dataset
stringclasses
6 values
model
stringclasses
13 values
seed
int64
42
46
n_topics
int64
10
50
topic_descriptions
sequencelengths
1
50
runtime_s
float64
0.93
52.3k
encoder
stringclasses
5 values
diversity
float64
0.09
1
c_npmi
float64
-0.38
0.21
wec_ex
float64
0.11
0.49
wec_in
float64
0.07
0.94
ArXiv ML Papers
BERTopic
43
10
[ [ "with", "of", "we", "for", "in", "and", "the", "that", "is", "to" ], [ "to", "the", "privacy", "federated", "learning", "of", "data", "and", "in", "we" ], [ "in", "of", "that", "the", "and", "to", "we", "on", "this", "for" ], [ "to", "time", "the", "and", "of", "series", "in", "for", "forecasting", "is" ], [ "of", "we", "the", "in", "graph", "graphs", "to", "and", "on", "node" ], [ "and", "the", "image", "in", "of", "we", "to", "for", "on", "with" ], [ "we", "of", "the", "to", "in", "and", "learning", "that", "policy", "is" ], [ "and", "eeg", "the", "to", "of", "signals", "in", "on", "data", "using" ], [ "to", "and", "is", "protein", "in", "the", "of", "side", "health", "for" ], [ "the", "to", "fairness", "and", "of", "fair", "we", "that", "in", "learning" ] ]
19.087417
all-MiniLM-L6-v2
0.32
-0.028712
0.232479
0.556292
ArXiv ML Papers
BERTopic
44
10
[ [ "we", "of", "in", "and", "the", "that", "is", "for", "to", "on" ], [ "to", "of", "the", "in", "and", "we", "that", "for", "is", "on" ], [ "and", "fairness", "of", "we", "in", "to", "the", "that", "fair", "learning" ], [ "the", "to", "in", "eeg", "on", "of", "signals", "and", "data", "as" ], [ "of", "graph", "the", "and", "to", "in", "graphs", "we", "on", "node" ], [ "we", "in", "the", "of", "and", "for", "image", "to", "with", "on" ], [ "attacks", "to", "the", "in", "that", "adversarial", "and", "we", "attack", "of" ], [ "of", "the", "to", "learning", "in", "and", "we", "that", "policy", "is" ], [ "treatment", "causal", "the", "of", "to", "in", "we", "is", "and", "outcome" ], [ "to", "learning", "federated", "the", "devices", "data", "and", "of", "in", "server" ] ]
14.536075
all-MiniLM-L6-v2
0.32
-0.025437
0.23001
0.543245
ArXiv ML Papers
BERTopic
45
10
[ [ "on", "in", "we", "the", "and", "of", "for", "to", "that", "is" ], [ "for", "of", "image", "to", "on", "the", "we", "and", "in", "with" ], [ "the", "privacy", "federated", "to", "and", "learning", "data", "we", "of", "in" ], [ "the", "in", "network", "we", "to", "neural", "of", "and", "with", "for" ], [ "for", "and", "to", "of", "the", "in", "is", "we", "data", "that" ], [ "to", "learning", "and", "fair", "the", "that", "fairness", "in", "we", "of" ], [ "of", "variational", "we", "to", "in", "the", "and", "bayesian", "is", "inference" ], [ "eeg", "and", "in", "to", "signals", "of", "data", "the", "on", "as" ], [ "to", "attacks", "and", "adversarial", "the", "of", "we", "attack", "in", "that" ], [ "of", "in", "the", "to", "and", "we", "learning", "that", "policy", "is" ] ]
14.676268
all-MiniLM-L6-v2
0.3
-0.026364
0.224854
0.539863
ArXiv ML Papers
BERTopic
46
10
[ [ "the", "of", "to", "time", "and", "in", "for", "data", "series", "is" ], [ "to", "the", "and", "of", "in", "for", "data", "learning", "we", "is" ], [ "attacks", "we", "in", "adversarial", "of", "and", "that", "the", "training", "to" ], [ "neural", "for", "is", "to", "the", "and", "in", "we", "of", "with" ], [ "image", "for", "and", "we", "to", "of", "in", "the", "on", "with" ], [ "to", "and", "the", "of", "graph", "we", "in", "on", "that", "graphs" ], [ "and", "of", "learning", "we", "that", "to", "in", "the", "policy", "is" ], [ "to", "and", "in", "of", "speech", "the", "we", "audio", "for", "on" ], [ "fair", "fairness", "the", "that", "and", "to", "of", "we", "in", "learning" ], [ "and", "of", "the", "to", "is", "in", "for", "we", "algorithm", "matrix" ] ]
14.212026
all-MiniLM-L6-v2
0.29
-0.030181
0.222051
0.532827
ArXiv ML Papers
BERTopic
43
20
[ [ "in", "with", "the", "image", "segmentation", "of", "for", "to", "we", "and" ], [ "and", "image", "to", "gans", "gan", "of", "images", "generative", "the", "we" ], [ "3d", "scene", "and", "point", "the", "semantic", "to", "depth", "we", "in" ], [ "to", "data", "of", "in", "the", "and", "is", "for", "we", "matrix" ], [ "gradient", "the", "stochastic", "of", "convex", "we", "and", "convergence", "is", "optimization" ], [ "the", "in", "pruning", "network", "to", "of", "on", "neural", "and", "hardware" ], [ "and", "in", "we", "equations", "to", "of", "neural", "the", "physics", "data" ], [ "of", "time", "the", "and", "forecasting", "series", "to", "for", "in", "model" ], [ "and", "the", "variational", "of", "to", "we", "in", "for", "is", "bayesian" ], [ "and", "graphs", "graph", "the", "of", "to", "in", "nodes", "node", "we" ], [ "in", "we", "and", "models", "to", "software", "of", "the", "explanation", "explanations" ], [ "to", "and", "the", "of", "in", "for", "on", "speech", "language", "we" ], [ "private", "data", "privacy", "differentially", "the", "to", "we", "that", "learning", "our" ], [ "federated", "to", "distributed", "learning", "the", "devices", "data", "and", "server", "communication" ], [ "the", "learning", "of", "in", "to", "and", "policy", "we", "that", "is" ], [ "the", "attacks", "to", "of", "adversarial", "attack", "we", "training", "in", "and" ], [ "to", "on", "eeg", "signals", "of", "the", "and", "in", "using", "data" ], [ "side", "protein", "the", "and", "of", "to", "health", "in", "is", "patient" ], [ "neural", "the", "network", "we", "of", "networks", "relu", "in", "to", "layer" ], [ "fairness", "the", "to", "we", "of", "and", "fair", "that", "learning", "machine" ] ]
15.018337
all-MiniLM-L6-v2
0.395
-0.00707
0.172834
0.582825
ArXiv ML Papers
BERTopic
44
20
[ [ "segmentation", "the", "and", "to", "in", "of", "with", "image", "for", "we" ], [ "for", "and", "the", "of", "we", "uncertainty", "to", "variational", "in", "is" ], [ "of", "to", "the", "pruning", "and", "neural", "network", "hardware", "on", "nas" ], [ "image", "gans", "gan", "the", "generative", "to", "images", "of", "generator", "and" ], [ "of", "for", "the", "in", "is", "and", "we", "to", "matrix", "data" ], [ "protein", "of", "to", "health", "the", "and", "in", "predict", "is", "this" ], [ "class", "classification", "the", "decision", "of", "classifier", "label", "to", "noise", "codes" ], [ "of", "eeg", "to", "and", "the", "signals", "in", "on", "data", "as" ], [ "the", "gradient", "of", "stochastic", "convex", "and", "is", "we", "optimization", "convergence" ], [ "of", "the", "graph", "and", "graphs", "we", "node", "to", "in", "on" ], [ "differential", "to", "that", "privacy", "private", "the", "differentially", "data", "we", "of" ], [ "model", "of", "time", "series", "in", "for", "forecasting", "to", "the", "and" ], [ "to", "fair", "of", "and", "fairness", "the", "we", "that", "machine", "learning" ], [ "and", "on", "of", "we", "for", "that", "to", "in", "the", "language" ], [ "adversarial", "the", "attacks", "to", "training", "and", "of", "we", "in", "attack" ], [ "we", "in", "learning", "to", "the", "policy", "that", "of", "and", "is" ], [ "the", "to", "of", "and", "neural", "we", "equations", "physics", "in", "data" ], [ "treatment", "in", "the", "outcome", "we", "causal", "of", "to", "observational", "is" ], [ "of", "relu", "in", "to", "network", "networks", "neural", "the", "we", "that" ], [ "data", "to", "federated", "training", "distributed", "server", "learning", "devices", "the", "and" ] ]
14.292851
all-MiniLM-L6-v2
0.395
-0.009121
0.174477
0.578752
ArXiv ML Papers
BERTopic
45
20
[ [ "we", "policy", "that", "learning", "in", "the", "and", "of", "to", "is" ], [ "and", "fairness", "fair", "to", "we", "that", "of", "the", "machine", "learning" ], [ "we", "the", "in", "of", "neural", "to", "and", "for", "data", "physics" ], [ "federated", "training", "learning", "and", "data", "devices", "to", "server", "the", "distributed" ], [ "to", "we", "the", "privacy", "private", "differentially", "data", "that", "differential", "learning" ], [ "attacks", "of", "the", "adversarial", "to", "attack", "we", "and", "in", "that" ], [ "the", "we", "that", "and", "in", "language", "on", "to", "for", "of" ], [ "signals", "eeg", "the", "to", "of", "and", "on", "in", "data", "as" ], [ "of", "graph", "and", "to", "the", "in", "we", "graphs", "on", "for" ], [ "the", "of", "neural", "network", "relu", "networks", "in", "that", "we", "to" ], [ "of", "time", "the", "series", "forecasting", "for", "to", "and", "in", "model" ], [ "is", "of", "for", "in", "to", "and", "the", "data", "we", "algorithm" ], [ "generative", "gans", "gan", "image", "to", "images", "and", "generator", "the", "of" ], [ "the", "of", "in", "we", "to", "image", "and", "medical", "segmentation", "for" ], [ "pruning", "to", "the", "on", "of", "neural", "and", "hardware", "in", "nas" ], [ "and", "in", "of", "detection", "the", "to", "we", "for", "object", "on" ], [ "convex", "optimization", "and", "stochastic", "gradient", "convergence", "of", "the", "we", "is" ], [ "and", "variational", "in", "we", "of", "latent", "to", "uncertainty", "the", "posterior" ], [ "the", "we", "inference", "on", "bayesian", "for", "and", "of", "to", "in" ] ]
37.529427
all-MiniLM-L6-v2
0.357895
-0.008285
0.179466
0.583949
ArXiv ML Papers
BERTopic
46
20
[ [ "the", "of", "and", "to", "in", "neural", "network", "we", "networks", "for" ], [ "of", "to", "metric", "the", "learning", "distance", "data", "in", "we", "is" ], [ "segmentation", "of", "and", "to", "the", "in", "for", "image", "with", "we" ], [ "we", "the", "point", "depth", "and", "3d", "of", "scene", "to", "in" ], [ "algorithm", "for", "is", "in", "and", "of", "the", "we", "to", "matrix" ], [ "series", "the", "time", "and", "forecasting", "of", "in", "to", "for", "model" ], [ "and", "to", "hardware", "of", "pruning", "neural", "the", "on", "network", "accuracy" ], [ "speech", "of", "in", "audio", "to", "and", "the", "speaker", "we", "for" ], [ "in", "the", "for", "is", "variational", "to", "we", "and", "of", "bayesian" ], [ "gan", "image", "generative", "the", "to", "and", "images", "of", "gans", "we" ], [ "the", "eeg", "and", "to", "of", "in", "signals", "on", "as", "using" ], [ "to", "and", "federated", "learning", "data", "devices", "server", "the", "distributed", "communication" ], [ "graph", "graphs", "of", "the", "and", "in", "on", "to", "node", "we" ], [ "in", "and", "to", "of", "the", "learning", "policy", "we", "is", "that" ], [ "the", "for", "in", "protein", "of", "to", "and", "are", "is", "health" ], [ "recommendation", "user", "to", "item", "items", "recommender", "of", "we", "the", "and" ], [ "privacy", "private", "sensitive", "that", "the", "attacks", "to", "we", "of", "differentially" ], [ "to", "and", "we", "of", "fairness", "that", "the", "fair", "machine", "learning" ], [ "the", "to", "adversarial", "attacks", "attack", "we", "of", "and", "in", "training" ], [ "language", "the", "to", "of", "models", "and", "we", "on", "in", "for" ] ]
13.924605
all-MiniLM-L6-v2
0.38
-0.006415
0.178178
0.564837
ArXiv ML Papers
BERTopic
43
30
[ [ "distributed", "the", "federated", "learning", "to", "devices", "data", "server", "communication", "and" ], [ "we", "learning", "of", "in", "the", "to", "policy", "and", "that", "is" ], [ "attacks", "the", "to", "attack", "adversarial", "of", "and", "we", "in", "training" ], [ "that", "data", "privacy", "differentially", "the", "private", "we", "to", "learning", "differential" ], [ "fairness", "and", "to", "fair", "the", "of", "machine", "we", "that", "learning" ], [ "to", "the", "and", "of", "speech", "language", "in", "we", "for", "on" ], [ "in", "we", "and", "models", "to", "software", "of", "the", "explanation", "explanations" ], [ "and", "of", "the", "to", "neural", "we", "physics", "equations", "in", "data" ], [ "network", "the", "neural", "networks", "relu", "of", "we", "in", "layer", "to" ], [ "to", "on", "eeg", "signals", "of", "the", "and", "in", "using", "data" ], [ "the", "graphs", "graph", "of", "and", "nodes", "node", "to", "we", "in" ], [ "side", "protein", "the", "and", "of", "to", "health", "in", "patient", "is" ], [ "time", "the", "of", "and", "forecasting", "to", "series", "in", "for", "model" ], [ "gradient", "convex", "and", "of", "the", "we", "stochastic", "convergence", "is", "optimization" ], [ "and", "the", "variational", "of", "to", "we", "in", "for", "is", "bayesian" ], [ "of", "for", "the", "is", "and", "in", "to", "data", "matrix", "we" ], [ "images", "image", "gans", "gan", "generative", "to", "the", "and", "of", "we" ], [ "of", "and", "the", "pruning", "to", "neural", "on", "hardware", "network", "in" ], [ "the", "segmentation", "of", "and", "to", "in", "image", "medical", "we", "for" ], [ "and", "with", "the", "object", "to", "in", "detection", "of", "on", "we" ], [ "we", "and", "depth", "scene", "the", "3d", "point", "to", "semantic", "in" ] ]
11.780045
all-MiniLM-L6-v2
0.390476
-0.006637
0.170969
0.576168
ArXiv ML Papers
BERTopic
44
30
[ [ "we", "in", "learning", "to", "the", "policy", "that", "of", "and", "is" ], [ "to", "fair", "of", "and", "fairness", "the", "we", "that", "machine", "learning" ], [ "data", "federated", "server", "to", "distributed", "devices", "learning", "and", "the", "training" ], [ "we", "to", "of", "attacks", "adversarial", "the", "attack", "and", "in", "training" ], [ "data", "private", "we", "the", "that", "privacy", "differential", "differentially", "to", "sensitive" ], [ "of", "the", "causal", "outcome", "treatment", "to", "we", "in", "is", "observational" ], [ "and", "on", "of", "we", "for", "that", "to", "in", "the", "language" ], [ "the", "of", "and", "to", "neural", "we", "physics", "equations", "in", "data" ], [ "the", "forecasting", "time", "series", "and", "in", "of", "to", "for", "model" ], [ "network", "relu", "in", "to", "the", "of", "neural", "networks", "we", "that" ], [ "the", "eeg", "to", "signals", "and", "of", "in", "on", "data", "as" ], [ "to", "health", "and", "protein", "the", "of", "in", "is", "predict", "machine" ], [ "of", "the", "graph", "and", "graphs", "we", "node", "to", "in", "nodes" ], [ "classification", "label", "the", "decision", "of", "classifier", "class", "to", "noise", "codes" ], [ "in", "and", "for", "of", "the", "is", "to", "matrix", "we", "data" ], [ "and", "of", "the", "gradient", "convex", "stochastic", "we", "is", "optimization", "convergence" ], [ "image", "gans", "gan", "generative", "to", "images", "of", "the", "generator", "and" ], [ "for", "and", "the", "of", "we", "uncertainty", "to", "variational", "in", "is" ], [ "to", "the", "pruning", "and", "of", "hardware", "network", "neural", "on", "nas" ], [ "in", "the", "with", "object", "of", "on", "and", "pooling", "detection", "to" ], [ "and", "segmentation", "to", "of", "the", "in", "image", "for", "we", "with" ] ]
11.743433
all-MiniLM-L6-v2
0.395238
-0.009103
0.171947
0.583365
ArXiv ML Papers
BERTopic
45
30
[ [ "we", "policy", "that", "learning", "in", "the", "and", "of", "to", "is" ], [ "and", "fairness", "fair", "to", "we", "that", "of", "the", "machine", "learning" ], [ "we", "the", "in", "of", "neural", "to", "and", "for", "data", "physics" ], [ "federated", "training", "learning", "and", "data", "devices", "to", "server", "the", "distributed" ], [ "to", "we", "the", "privacy", "private", "differentially", "data", "that", "differential", "learning" ], [ "attacks", "of", "the", "adversarial", "to", "attack", "we", "and", "in", "that" ], [ "the", "we", "that", "and", "in", "language", "on", "to", "for", "of" ], [ "signals", "eeg", "the", "to", "of", "and", "on", "in", "data", "as" ], [ "of", "graph", "and", "to", "the", "in", "we", "graphs", "on", "for" ], [ "the", "of", "neural", "network", "relu", "networks", "in", "that", "we", "to" ], [ "of", "time", "the", "series", "forecasting", "for", "to", "and", "in", "model" ], [ "is", "of", "for", "in", "to", "and", "the", "data", "we", "algorithm" ], [ "generative", "gans", "gan", "image", "to", "images", "and", "generator", "the", "of" ], [ "the", "of", "in", "we", "to", "image", "and", "medical", "segmentation", "for" ], [ "pruning", "to", "the", "on", "of", "neural", "and", "hardware", "in", "nas" ], [ "and", "in", "of", "detection", "the", "to", "we", "for", "object", "on" ], [ "convex", "optimization", "and", "stochastic", "gradient", "convergence", "of", "the", "we", "is" ], [ "and", "variational", "in", "we", "of", "latent", "to", "uncertainty", "the", "posterior" ], [ "the", "we", "inference", "on", "bayesian", "for", "and", "of", "to", "in" ] ]
11.415018
all-MiniLM-L6-v2
0.357895
-0.008285
0.179466
0.564226
ArXiv ML Papers
BERTopic
46
30
[ [ "learning", "policy", "in", "of", "the", "we", "and", "to", "that", "reinforcement" ], [ "learning", "federated", "the", "to", "data", "devices", "server", "and", "distributed", "communication" ], [ "that", "the", "and", "of", "we", "to", "fairness", "fair", "machine", "learning" ], [ "adversarial", "attacks", "attack", "the", "of", "we", "to", "and", "in", "training" ], [ "attacks", "differentially", "sensitive", "that", "private", "privacy", "the", "to", "differential", "of" ], [ "to", "of", "physics", "neural", "and", "the", "we", "equations", "in", "data" ], [ "to", "signals", "the", "as", "on", "and", "eeg", "in", "of", "using" ], [ "network", "neural", "relu", "the", "to", "activation", "in", "of", "networks", "layer" ], [ "protein", "of", "and", "to", "the", "in", "for", "are", "health", "is" ], [ "and", "graph", "of", "we", "the", "graphs", "to", "node", "network", "in" ], [ "the", "user", "of", "to", "item", "recommender", "we", "items", "recommendation", "users" ], [ "series", "of", "the", "time", "forecasting", "and", "to", "in", "for", "model" ], [ "speech", "of", "in", "audio", "to", "and", "the", "speaker", "we", "for" ], [ "of", "models", "language", "we", "to", "the", "and", "on", "in", "attention" ], [ "data", "the", "classification", "label", "is", "in", "of", "learning", "we", "labels" ], [ "image", "for", "of", "to", "medical", "segmentation", "the", "and", "in", "with" ], [ "is", "of", "the", "and", "for", "to", "in", "variational", "we", "bayesian" ], [ "convergence", "is", "optimization", "the", "stochastic", "of", "gradient", "convex", "and", "we" ], [ "generative", "image", "gan", "images", "of", "to", "gans", "and", "the", "generator" ], [ "metric", "distance", "the", "to", "we", "data", "of", "learning", "in", "and" ], [ "the", "matrix", "tensor", "algorithm", "rank", "for", "of", "is", "in", "and" ], [ "nas", "pruning", "to", "on", "of", "and", "neural", "the", "hardware", "accuracy" ], [ "the", "point", "scene", "depth", "and", "3d", "we", "of", "to", "semantic" ], [ "detection", "object", "the", "on", "and", "to", "of", "in", "cnn", "for" ] ]
12.662092
all-MiniLM-L6-v2
0.425
0.002061
0.164638
0.605135
ArXiv ML Papers
BERTopic
43
40
[ [ "distributed", "the", "federated", "learning", "to", "devices", "data", "server", "communication", "and" ], [ "we", "learning", "of", "in", "the", "to", "policy", "and", "that", "is" ], [ "attacks", "the", "to", "attack", "adversarial", "of", "and", "we", "in", "training" ], [ "that", "data", "privacy", "differentially", "the", "private", "we", "to", "learning", "differential" ], [ "fairness", "and", "to", "fair", "the", "of", "machine", "we", "that", "learning" ], [ "to", "the", "and", "of", "speech", "language", "in", "we", "for", "on" ], [ "in", "we", "and", "models", "to", "software", "of", "the", "explanation", "explanations" ], [ "and", "of", "the", "to", "neural", "we", "physics", "equations", "in", "data" ], [ "network", "the", "neural", "networks", "relu", "of", "we", "in", "layer", "to" ], [ "to", "on", "eeg", "signals", "of", "the", "and", "in", "using", "data" ], [ "the", "graphs", "graph", "of", "and", "nodes", "node", "to", "we", "in" ], [ "side", "protein", "the", "and", "of", "to", "health", "in", "patient", "is" ], [ "time", "the", "of", "and", "forecasting", "to", "series", "in", "for", "model" ], [ "gradient", "convex", "and", "of", "the", "we", "stochastic", "convergence", "is", "optimization" ], [ "and", "the", "variational", "of", "to", "we", "in", "for", "is", "bayesian" ], [ "of", "for", "the", "is", "and", "in", "to", "data", "matrix", "we" ], [ "images", "image", "gans", "gan", "generative", "to", "the", "and", "of", "we" ], [ "of", "and", "the", "pruning", "to", "neural", "on", "hardware", "network", "in" ], [ "the", "segmentation", "of", "and", "to", "in", "image", "medical", "we", "for" ], [ "and", "with", "the", "object", "to", "in", "detection", "of", "on", "we" ], [ "we", "and", "depth", "scene", "the", "3d", "point", "to", "semantic", "in" ] ]
15.499882
all-MiniLM-L6-v2
0.390476
-0.006637
0.170969
0.576788
ArXiv ML Papers
BERTopic
44
40
[ [ "we", "in", "learning", "to", "the", "policy", "that", "of", "and", "is" ], [ "to", "fair", "of", "and", "fairness", "the", "we", "that", "machine", "learning" ], [ "data", "federated", "server", "to", "distributed", "devices", "learning", "and", "the", "training" ], [ "we", "to", "of", "attacks", "adversarial", "the", "attack", "and", "in", "training" ], [ "data", "private", "we", "the", "that", "privacy", "differential", "differentially", "to", "sensitive" ], [ "of", "the", "causal", "outcome", "treatment", "to", "we", "in", "is", "observational" ], [ "and", "on", "of", "we", "for", "that", "to", "in", "the", "language" ], [ "the", "of", "and", "to", "neural", "we", "physics", "equations", "in", "data" ], [ "the", "forecasting", "time", "series", "and", "in", "of", "to", "for", "model" ], [ "network", "relu", "in", "to", "the", "of", "neural", "networks", "we", "that" ], [ "the", "eeg", "to", "signals", "and", "of", "in", "on", "data", "as" ], [ "to", "health", "and", "protein", "the", "of", "in", "is", "predict", "machine" ], [ "of", "the", "graph", "and", "graphs", "we", "node", "to", "in", "nodes" ], [ "classification", "label", "the", "decision", "of", "classifier", "class", "to", "noise", "codes" ], [ "in", "and", "for", "of", "the", "is", "to", "matrix", "we", "data" ], [ "and", "of", "the", "gradient", "convex", "stochastic", "we", "is", "optimization", "convergence" ], [ "image", "gans", "gan", "generative", "to", "images", "of", "the", "generator", "and" ], [ "for", "and", "the", "of", "we", "uncertainty", "to", "variational", "in", "is" ], [ "to", "the", "pruning", "and", "of", "hardware", "network", "neural", "on", "nas" ], [ "in", "the", "with", "object", "of", "on", "and", "pooling", "detection", "to" ], [ "and", "segmentation", "to", "of", "the", "in", "image", "for", "we", "with" ] ]
26.208948
all-MiniLM-L6-v2
0.395238
-0.009103
0.171947
0.589815
ArXiv ML Papers
BERTopic
45
40
[ [ "we", "policy", "that", "learning", "in", "the", "and", "of", "to", "is" ], [ "and", "fairness", "fair", "to", "we", "that", "of", "the", "machine", "learning" ], [ "we", "the", "in", "of", "neural", "to", "and", "for", "data", "physics" ], [ "federated", "training", "learning", "and", "data", "devices", "to", "server", "the", "distributed" ], [ "to", "we", "the", "privacy", "private", "differentially", "data", "that", "differential", "learning" ], [ "attacks", "of", "the", "adversarial", "to", "attack", "we", "and", "in", "that" ], [ "the", "we", "that", "and", "in", "language", "on", "to", "for", "of" ], [ "signals", "eeg", "the", "to", "of", "and", "on", "in", "data", "as" ], [ "of", "graph", "and", "to", "the", "in", "we", "graphs", "on", "for" ], [ "the", "of", "neural", "network", "relu", "networks", "in", "that", "we", "to" ], [ "of", "time", "the", "series", "forecasting", "for", "to", "and", "in", "model" ], [ "is", "of", "for", "in", "to", "and", "the", "data", "we", "algorithm" ], [ "generative", "gans", "gan", "image", "to", "images", "and", "generator", "the", "of" ], [ "the", "of", "in", "we", "to", "image", "and", "medical", "segmentation", "for" ], [ "pruning", "to", "the", "on", "of", "neural", "and", "hardware", "in", "nas" ], [ "and", "in", "of", "detection", "the", "to", "we", "for", "object", "on" ], [ "convex", "optimization", "and", "stochastic", "gradient", "convergence", "of", "the", "we", "is" ], [ "and", "variational", "in", "we", "of", "latent", "to", "uncertainty", "the", "posterior" ], [ "the", "we", "inference", "on", "bayesian", "for", "and", "of", "to", "in" ] ]
13.015473
all-MiniLM-L6-v2
0.357895
-0.008285
0.179466
0.561053
ArXiv ML Papers
BERTopic
46
40
[ [ "learning", "policy", "in", "of", "the", "we", "and", "to", "that", "reinforcement" ], [ "learning", "federated", "the", "to", "data", "devices", "server", "and", "distributed", "communication" ], [ "that", "the", "and", "of", "we", "to", "fairness", "fair", "machine", "learning" ], [ "adversarial", "attacks", "attack", "the", "of", "we", "to", "and", "in", "training" ], [ "attacks", "differentially", "sensitive", "that", "private", "privacy", "the", "to", "differential", "of" ], [ "to", "of", "physics", "neural", "and", "the", "we", "equations", "in", "data" ], [ "to", "signals", "the", "as", "on", "and", "eeg", "in", "of", "using" ], [ "network", "neural", "relu", "the", "to", "activation", "in", "of", "networks", "layer" ], [ "protein", "of", "and", "to", "the", "in", "for", "are", "health", "is" ], [ "and", "graph", "of", "we", "the", "graphs", "to", "node", "network", "in" ], [ "the", "user", "of", "to", "item", "recommender", "we", "items", "recommendation", "users" ], [ "series", "of", "the", "time", "forecasting", "and", "to", "in", "for", "model" ], [ "speech", "of", "in", "audio", "to", "and", "the", "speaker", "we", "for" ], [ "of", "models", "language", "we", "to", "the", "and", "on", "in", "attention" ], [ "data", "the", "classification", "label", "is", "in", "of", "learning", "we", "labels" ], [ "image", "for", "of", "to", "medical", "segmentation", "the", "and", "in", "with" ], [ "is", "of", "the", "and", "for", "to", "in", "variational", "we", "bayesian" ], [ "convergence", "is", "optimization", "the", "stochastic", "of", "gradient", "convex", "and", "we" ], [ "generative", "image", "gan", "images", "of", "to", "gans", "and", "the", "generator" ], [ "metric", "distance", "the", "to", "we", "data", "of", "learning", "in", "and" ], [ "the", "matrix", "tensor", "algorithm", "rank", "for", "of", "is", "in", "and" ], [ "nas", "pruning", "to", "on", "of", "and", "neural", "the", "hardware", "accuracy" ], [ "the", "point", "scene", "depth", "and", "3d", "we", "of", "to", "semantic" ], [ "detection", "object", "the", "on", "and", "to", "of", "in", "cnn", "for" ] ]
31.63049
all-MiniLM-L6-v2
0.425
0.002061
0.164638
0.597184
ArXiv ML Papers
BERTopic
43
50
[ [ "distributed", "the", "federated", "learning", "to", "devices", "data", "server", "communication", "and" ], [ "we", "learning", "of", "in", "the", "to", "policy", "and", "that", "is" ], [ "attacks", "the", "to", "attack", "adversarial", "of", "and", "we", "in", "training" ], [ "that", "data", "privacy", "differentially", "the", "private", "we", "to", "learning", "differential" ], [ "fairness", "and", "to", "fair", "the", "of", "machine", "we", "that", "learning" ], [ "to", "the", "and", "of", "speech", "language", "in", "we", "for", "on" ], [ "in", "we", "and", "models", "to", "software", "of", "the", "explanation", "explanations" ], [ "and", "of", "the", "to", "neural", "we", "physics", "equations", "in", "data" ], [ "network", "the", "neural", "networks", "relu", "of", "we", "in", "layer", "to" ], [ "to", "on", "eeg", "signals", "of", "the", "and", "in", "using", "data" ], [ "the", "graphs", "graph", "of", "and", "nodes", "node", "to", "we", "in" ], [ "side", "protein", "the", "and", "of", "to", "health", "in", "patient", "is" ], [ "time", "the", "of", "and", "forecasting", "to", "series", "in", "for", "model" ], [ "gradient", "convex", "and", "of", "the", "we", "stochastic", "convergence", "is", "optimization" ], [ "and", "the", "variational", "of", "to", "we", "in", "for", "is", "bayesian" ], [ "of", "for", "the", "is", "and", "in", "to", "data", "matrix", "we" ], [ "images", "image", "gans", "gan", "generative", "to", "the", "and", "of", "we" ], [ "of", "and", "the", "pruning", "to", "neural", "on", "hardware", "network", "in" ], [ "the", "segmentation", "of", "and", "to", "in", "image", "medical", "we", "for" ], [ "and", "with", "the", "object", "to", "in", "detection", "of", "on", "we" ], [ "we", "and", "depth", "scene", "the", "3d", "point", "to", "semantic", "in" ] ]
12.399162
all-MiniLM-L6-v2
0.390476
-0.006637
0.170969
0.587098
ArXiv ML Papers
BERTopic
45
50
[ [ "the", "to", "policy", "and", "learning", "that", "we", "of", "in", "is" ], [ "server", "to", "federated", "data", "the", "learning", "and", "distributed", "devices", "training" ], [ "we", "and", "to", "the", "that", "attack", "attacks", "in", "adversarial", "of" ], [ "fairness", "fair", "and", "to", "the", "of", "we", "that", "machine", "learning" ], [ "of", "for", "to", "in", "that", "language", "on", "we", "and", "the" ], [ "data", "of", "to", "the", "neural", "equations", "physics", "we", "and", "in" ], [ "the", "neural", "network", "in", "networks", "we", "that", "of", "relu", "to" ], [ "label", "labels", "is", "of", "the", "to", "classification", "classifier", "classifiers", "in" ], [ "time", "of", "forecasting", "series", "the", "and", "to", "in", "model", "for" ], [ "in", "signals", "and", "the", "eeg", "of", "using", "on", "to", "as" ], [ "are", "the", "of", "health", "and", "to", "protein", "in", "for", "disease" ], [ "graph", "graphs", "the", "of", "and", "to", "we", "in", "node", "on" ], [ "images", "of", "generative", "gans", "to", "the", "and", "image", "gan", "generator" ], [ "the", "stochastic", "convergence", "gradient", "convex", "and", "is", "of", "optimization", "we" ], [ "to", "of", "we", "and", "variational", "the", "in", "uncertainty", "is", "for" ], [ "of", "to", "and", "hardware", "the", "pruning", "neural", "on", "nas", "network" ], [ "in", "of", "to", "the", "image", "segmentation", "and", "medical", "for", "with" ], [ "the", "of", "in", "and", "to", "pooling", "with", "object", "on", "image" ], [ "3d", "point", "scene", "depth", "semantic", "and", "the", "we", "segmentation", "driving" ], [ "of", "data", "to", "the", "metric", "kernel", "we", "in", "distance", "learning" ], [ "the", "tensor", "in", "for", "algorithm", "rank", "of", "is", "matrix", "and" ] ]
498.173665
all-MiniLM-L6-v2
0.404762
-0.005565
0.17295
0.592389
ArXiv ML Papers
BERTopic
46
50
[ [ "in", "policy", "and", "of", "to", "learning", "the", "we", "that", "is" ], [ "distributed", "learning", "devices", "server", "federated", "the", "to", "data", "and", "training" ], [ "adversarial", "the", "attacks", "to", "attack", "of", "we", "training", "in", "and" ], [ "differentially", "that", "to", "our", "we", "private", "privacy", "the", "data", "as" ], [ "fairness", "fair", "and", "to", "the", "of", "machine", "we", "that", "learning" ], [ "to", "and", "the", "in", "of", "we", "on", "that", "language", "for" ], [ "data", "we", "and", "the", "to", "in", "physics", "neural", "of", "equations" ], [ "the", "of", "time", "forecasting", "and", "series", "to", "in", "model", "for" ], [ "relu", "neural", "networks", "network", "the", "of", "in", "to", "we", "and" ], [ "to", "eeg", "of", "signals", "the", "and", "on", "in", "data", "using" ], [ "the", "graph", "graphs", "of", "in", "to", "we", "and", "node", "network" ], [ "of", "protein", "the", "and", "health", "to", "in", "for", "from", "is" ], [ "of", "classification", "in", "label", "data", "the", "we", "to", "is", "and" ], [ "and", "with", "medical", "in", "to", "of", "images", "the", "segmentation", "image" ], [ "the", "distance", "metric", "of", "learning", "in", "we", "and", "to", "data" ], [ "the", "algorithm", "tensor", "matrix", "of", "and", "is", "rank", "for", "in" ], [ "generative", "gan", "image", "gans", "images", "to", "generator", "and", "of", "the" ], [ "nas", "neural", "on", "network", "of", "to", "pruning", "hardware", "and", "the" ], [ "and", "of", "detection", "to", "on", "object", "for", "the", "we", "in" ], [ "gradient", "stochastic", "is", "we", "the", "convergence", "of", "optimization", "and", "convex" ], [ "of", "the", "variational", "we", "and", "is", "in", "bayesian", "latent", "to" ] ]
529.846204
all-MiniLM-L6-v2
0.37619
-0.004715
0.17169
0.580724
ArXiv ML Papers
NMF
43
10
[ [ "the", "and", "of", "proposed", "to", "on", "by", "based", "method", "is" ], [ "to", "the", "be", "this", "neural", "model", "from", "can", "networks", "an" ], [ "to", "an", "of", "this", "the", "number", "these", "are", "our", "with" ], [ "models", "both", "are", "with", "based", "on", "and", "methods", "as", "performance" ], [ "show", "as", "the", "we", "on", "with", "our", "propose", "that", "can" ], [ "the", "in", "has", "this", "are", "we", "as", "or", "which", "been" ], [ "for", "with", "an", "we", "are", "be", "as", "algorithm", "can", "used" ], [ "training", "the", "for", "is", "data", "from", "as", "on", "model", "models" ], [ "machine", "to", "tasks", "on", "of", "algorithms", "and", "learning", "deep", "reinforcement" ], [ "algorithm", "and", "this", "is", "it", "that", "are", "be", "by", "which" ] ]
1.813817
all-MiniLM-L6-v2
0.49
-0.008151
0.195331
0.467242
ArXiv ML Papers
NMF
44
10
[ [ "method", "on", "of", "to", "is", "proposed", "and", "based", "the", "by" ], [ "the", "from", "this", "to", "model", "be", "neural", "can", "an", "networks" ], [ "to", "an", "of", "this", "the", "number", "these", "are", "our", "with" ], [ "models", "both", "are", "with", "based", "on", "and", "methods", "as", "performance" ], [ "show", "as", "the", "we", "on", "with", "our", "propose", "that", "can" ], [ "the", "in", "has", "this", "are", "we", "as", "or", "which", "been" ], [ "for", "with", "an", "we", "are", "be", "as", "algorithm", "can", "used" ], [ "training", "the", "for", "is", "data", "from", "as", "on", "model", "models" ], [ "machine", "to", "tasks", "on", "of", "algorithms", "and", "learning", "deep", "reinforcement" ], [ "algorithm", "and", "this", "is", "it", "that", "are", "be", "by", "which" ] ]
2.021709
all-MiniLM-L6-v2
0.49
-0.008151
0.195331
0.464778
ArXiv ML Papers
NMF
45
10
[ [ "the", "and", "of", "proposed", "to", "on", "by", "based", "method", "is" ], [ "the", "from", "this", "to", "model", "be", "neural", "can", "an", "networks" ], [ "to", "an", "of", "this", "the", "number", "these", "are", "our", "with" ], [ "models", "both", "are", "with", "based", "on", "and", "methods", "as", "performance" ], [ "show", "as", "the", "we", "on", "with", "our", "propose", "that", "can" ], [ "the", "in", "has", "this", "are", "we", "as", "or", "which", "been" ], [ "for", "with", "an", "we", "are", "be", "as", "algorithm", "can", "used" ], [ "training", "the", "for", "is", "data", "from", "as", "on", "model", "models" ], [ "machine", "to", "tasks", "on", "of", "algorithms", "and", "learning", "deep", "reinforcement" ], [ "algorithm", "and", "this", "is", "it", "that", "are", "be", "by", "which" ] ]
2.06287
all-MiniLM-L6-v2
0.49
-0.008151
0.195331
0.469553
ArXiv ML Papers
NMF
46
10
[ [ "the", "and", "of", "proposed", "to", "on", "by", "based", "method", "is" ], [ "to", "the", "be", "this", "neural", "model", "from", "can", "networks", "an" ], [ "to", "an", "of", "this", "the", "number", "these", "are", "our", "with" ], [ "models", "both", "are", "with", "based", "on", "and", "methods", "as", "performance" ], [ "show", "as", "the", "we", "on", "with", "our", "propose", "that", "can" ], [ "the", "in", "has", "this", "are", "we", "as", "or", "which", "been" ], [ "for", "with", "an", "we", "are", "be", "as", "algorithm", "can", "used" ], [ "training", "the", "for", "is", "data", "from", "as", "on", "model", "models" ], [ "machine", "to", "tasks", "on", "of", "algorithms", "and", "learning", "deep", "reinforcement" ], [ "algorithm", "and", "this", "is", "it", "that", "are", "be", "by", "which" ] ]
1.930904
all-MiniLM-L6-v2
0.49
-0.008151
0.195331
0.47231
ArXiv ML Papers
NMF
43
20
[ [ "proposed", "of", "the", "to", "by", "between", "information", "results", "using", "two" ], [ "from", "the", "it", "due", "an", "to", "their", "framework", "this", "approach" ], [ "these", "set", "the", "this", "number", "an", "of", "new", "state", "art" ], [ "both", "and", "two", "models", "are", "the", "research", "ml", "their", "including" ], [ "problem", "which", "our", "show", "we", "and", "the", "propose", "in", "algorithm" ], [ "in", "or", "are", "work", "paper", "the", "this", "been", "has", "many" ], [ "time", "are", "the", "each", "for", "test", "also", "used", "using", "analysis" ], [ "from", "to", "we", "data", "training", "real", "is", "classification", "of", "time" ], [ "deep", "of", "learning", "machine", "algorithms", "reinforcement", "tasks", "policy", "supervised", "has" ], [ "it", "problem", "which", "where", "algorithm", "is", "and", "this", "an", "at" ], [ "based", "an", "on", "performance", "by", "datasets", "this", "tasks", "classification", "task" ], [ "the", "representations", "are", "this", "show", "not", "more", "than", "also", "that" ], [ "training", "deep", "convolutional", "to", "is", "network", "networks", "neural", "we", "layer" ], [ "model", "which", "trained", "we", "performance", "models", "inference", "in", "our", "based" ], [ "to", "have", "an", "are", "as", "such", "or", "well", "these", "using" ], [ "with", "dataset", "high", "accuracy", "an", "by", "low", "performance", "results", "while" ], [ "and", "state", "methods", "gradient", "based", "method", "optimization", "proposed", "propose", "our" ], [ "we", "training", "models", "robustness", "attacks", "adversarial", "attack", "to", "against", "the" ], [ "by", "be", "this", "can", "from", "it", "or", "which", "used", "how" ], [ "the", "graphs", "graph", "gnns", "representation", "nodes", "node", "and", "our", "we" ] ]
2.422656
all-MiniLM-L6-v2
0.565
0.013258
0.170953
0.561412
ArXiv ML Papers
NMF
44
20
[ [ "proposed", "of", "the", "to", "by", "between", "information", "results", "using", "two" ], [ "from", "the", "it", "due", "an", "to", "their", "framework", "this", "approach" ], [ "these", "set", "the", "this", "number", "an", "of", "new", "state", "art" ], [ "ml", "the", "research", "both", "are", "and", "two", "models", "their", "including" ], [ "problem", "which", "our", "show", "we", "and", "the", "propose", "in", "algorithm" ], [ "in", "or", "are", "work", "paper", "the", "this", "been", "has", "many" ], [ "time", "are", "the", "each", "for", "test", "also", "used", "using", "analysis" ], [ "from", "to", "we", "data", "training", "real", "is", "classification", "of", "time" ], [ "deep", "of", "learning", "machine", "algorithms", "reinforcement", "tasks", "policy", "supervised", "has" ], [ "it", "problem", "which", "where", "algorithm", "is", "and", "this", "an", "at" ], [ "based", "an", "on", "performance", "by", "datasets", "this", "tasks", "classification", "task" ], [ "that", "show", "are", "this", "not", "representations", "also", "more", "than", "the" ], [ "training", "deep", "convolutional", "to", "is", "network", "networks", "neural", "we", "layer" ], [ "model", "which", "trained", "we", "performance", "models", "inference", "in", "our", "based" ], [ "to", "have", "an", "are", "as", "such", "or", "well", "these", "using" ], [ "with", "dataset", "high", "accuracy", "an", "by", "low", "performance", "results", "while" ], [ "and", "state", "methods", "gradient", "based", "method", "optimization", "proposed", "propose", "our" ], [ "models", "attack", "attacks", "training", "to", "we", "the", "robustness", "adversarial", "against" ], [ "by", "be", "this", "can", "from", "it", "or", "which", "used", "how" ], [ "the", "graphs", "graph", "gnns", "representation", "nodes", "node", "and", "our", "we" ] ]
2.538431
all-MiniLM-L6-v2
0.565
0.013258
0.170953
0.559352
ArXiv ML Papers
NMF
45
20
[ [ "by", "to", "of", "the", "proposed", "results", "using", "between", "information", "two" ], [ "from", "the", "it", "due", "an", "to", "their", "framework", "this", "approach" ], [ "these", "set", "the", "this", "number", "an", "of", "new", "state", "art" ], [ "ml", "the", "research", "both", "are", "and", "two", "models", "their", "including" ], [ "problem", "which", "our", "show", "we", "and", "the", "propose", "in", "algorithm" ], [ "in", "or", "are", "work", "paper", "the", "this", "been", "has", "many" ], [ "time", "are", "the", "each", "for", "test", "also", "used", "using", "analysis" ], [ "from", "to", "we", "data", "training", "real", "is", "classification", "of", "time" ], [ "deep", "of", "learning", "machine", "algorithms", "reinforcement", "tasks", "policy", "supervised", "has" ], [ "it", "problem", "which", "where", "algorithm", "is", "and", "this", "an", "at" ], [ "based", "an", "on", "performance", "by", "datasets", "this", "tasks", "classification", "task" ], [ "the", "representations", "are", "this", "show", "not", "more", "than", "also", "that" ], [ "training", "deep", "convolutional", "to", "is", "network", "networks", "neural", "we", "layer" ], [ "model", "which", "trained", "we", "performance", "models", "inference", "in", "our", "based" ], [ "to", "have", "an", "are", "as", "such", "or", "well", "these", "using" ], [ "with", "dataset", "high", "accuracy", "an", "by", "low", "performance", "results", "while" ], [ "and", "state", "methods", "gradient", "based", "method", "optimization", "proposed", "propose", "our" ], [ "we", "training", "models", "robustness", "attacks", "adversarial", "attack", "to", "against", "the" ], [ "by", "be", "this", "can", "from", "it", "or", "which", "used", "how" ], [ "the", "graphs", "graph", "gnns", "representation", "nodes", "node", "and", "our", "we" ] ]
2.535384
all-MiniLM-L6-v2
0.565
0.013258
0.170953
0.559035
ArXiv ML Papers
NMF
46
20
[ [ "proposed", "of", "the", "to", "by", "between", "information", "results", "using", "two" ], [ "the", "due", "to", "an", "from", "this", "their", "it", "approach", "framework" ], [ "these", "set", "the", "this", "number", "an", "of", "new", "state", "art" ], [ "both", "and", "two", "models", "are", "the", "research", "ml", "their", "including" ], [ "problem", "which", "our", "show", "we", "and", "the", "propose", "in", "algorithm" ], [ "in", "or", "are", "work", "paper", "the", "this", "been", "has", "many" ], [ "time", "are", "the", "each", "for", "test", "also", "used", "using", "analysis" ], [ "from", "to", "we", "data", "training", "real", "is", "classification", "of", "time" ], [ "deep", "of", "learning", "machine", "algorithms", "reinforcement", "tasks", "policy", "supervised", "has" ], [ "it", "problem", "which", "where", "algorithm", "is", "and", "this", "an", "at" ], [ "based", "an", "on", "performance", "by", "datasets", "this", "tasks", "classification", "task" ], [ "the", "representations", "are", "this", "show", "not", "more", "than", "also", "that" ], [ "training", "deep", "convolutional", "to", "is", "network", "networks", "neural", "we", "layer" ], [ "model", "which", "trained", "we", "performance", "models", "inference", "in", "our", "based" ], [ "to", "have", "an", "are", "as", "such", "or", "well", "these", "using" ], [ "with", "dataset", "high", "accuracy", "an", "by", "low", "performance", "results", "while" ], [ "and", "state", "methods", "gradient", "based", "method", "optimization", "proposed", "propose", "our" ], [ "models", "attack", "attacks", "training", "to", "we", "the", "robustness", "adversarial", "against" ], [ "by", "be", "this", "can", "from", "it", "or", "which", "used", "how" ], [ "the", "graphs", "graph", "gnns", "representation", "nodes", "node", "and", "our", "we" ] ]
2.505962
all-MiniLM-L6-v2
0.565
0.013258
0.170953
0.557773
ArXiv ML Papers
NMF
43
30
[ [ "of", "to", "problem", "results", "the", "between", "proposed", "information", "two", "space" ], [ "due", "the", "their", "to", "use", "have", "human", "it", "end", "search" ], [ "the", "of", "number", "set", "analysis", "large", "these", "terms", "art", "and" ], [ "research", "and", "two", "the", "both", "accuracy", "ml", "between", "including", "their" ], [ "the", "demonstrate", "we", "which", "show", "use", "present", "propose", "in", "where" ], [ "been", "in", "or", "applications", "has", "work", "many", "the", "recent", "have" ], [ "for", "the", "time", "test", "also", "used", "using", "each", "regression", "prediction" ], [ "data", "and", "classification", "real", "training", "time", "to", "synthetic", "large", "supervised" ], [ "supervised", "learning", "deep", "machine", "of", "has", "tasks", "reinforcement", "metric", "agent" ], [ "it", "is", "which", "where", "one", "than", "not", "at", "when", "information" ], [ "performance", "trained", "classification", "or", "art", "tasks", "work", "on", "task", "datasets" ], [ "that", "show", "results", "not", "representations", "the", "than", "find", "more", "also" ], [ "convolutional", "and", "networks", "network", "deep", "training", "neural", "to", "layer", "performance" ], [ "the", "model", "performance", "and", "which", "in", "trained", "proposed", "inference", "parameters" ], [ "or", "have", "well", "to", "such", "as", "the", "these", "been", "not" ], [ "dataset", "accuracy", "low", "high", "results", "while", "performance", "compared", "with", "achieve" ], [ "by", "feature", "high", "series", "images", "approach", "between", "time", "the", "into" ], [ "training", "to", "we", "adversarial", "attacks", "the", "robustness", "against", "examples", "attack" ], [ "this", "paper", "work", "process", "new", "distribution", "approach", "problem", "been", "system" ], [ "image", "which", "input", "an", "approach", "features", "classification", "using", "high", "its" ], [ "detection", "the", "based", "on", "system", "feature", "recognition", "propose", "proposed", "have" ], [ "image", "which", "used", "or", "quantum", "how", "it", "be", "they", "can" ], [ "more", "different", "they", "these", "are", "their", "algorithms", "some", "when", "where" ], [ "graph", "node", "nodes", "and", "the", "graphs", "representation", "we", "gnns", "which" ], [ "method", "methods", "proposed", "deep", "prediction", "datasets", "classification", "existing", "feature", "large" ], [ "of", "models", "generative", "language", "trained", "these", "fidelity", "performance", "inference", "have" ], [ "stochastic", "function", "policy", "convex", "and", "convergence", "gradient", "problems", "optimization", "optimal" ], [ "approach", "results", "our", "we", "method", "framework", "state", "to", "at", "art" ], [ "to", "algorithm", "problem", "algorithms", "matrix", "proposed", "linear", "clustering", "sample", "rank" ], [ "task", "language", "using", "tasks", "from", "different", "representations", "knowledge", "object", "information" ] ]
3.53165
all-MiniLM-L6-v2
0.566667
0.022401
0.162399
0.62859
ArXiv ML Papers
NMF
44
30
[ [ "to", "proposed", "between", "of", "the", "results", "problem", "two", "information", "space" ], [ "human", "to", "their", "due", "use", "it", "end", "have", "framework", "search" ], [ "these", "set", "and", "the", "of", "terms", "analysis", "number", "art", "large" ], [ "two", "both", "the", "and", "research", "accuracy", "ml", "including", "between", "their" ], [ "show", "use", "which", "we", "in", "present", "the", "propose", "and", "demonstrate" ], [ "or", "in", "work", "has", "been", "applications", "many", "recent", "have", "high" ], [ "using", "the", "prediction", "time", "test", "used", "each", "regression", "for", "also" ], [ "data", "and", "classification", "real", "training", "time", "to", "synthetic", "large", "supervised" ], [ "supervised", "of", "deep", "reinforcement", "learning", "machine", "tasks", "has", "metric", "agent" ], [ "it", "which", "where", "one", "is", "not", "at", "than", "when", "information" ], [ "on", "performance", "tasks", "datasets", "task", "or", "trained", "classification", "work", "art" ], [ "the", "find", "representations", "that", "not", "show", "results", "than", "more", "also" ], [ "training", "convolutional", "networks", "network", "neural", "and", "deep", "to", "layer", "performance" ], [ "the", "model", "in", "performance", "and", "which", "trained", "proposed", "inference", "parameters" ], [ "well", "such", "as", "to", "have", "or", "been", "these", "not", "using" ], [ "dataset", "high", "compared", "performance", "accuracy", "low", "results", "with", "while", "achieve" ], [ "which", "input", "image", "approach", "an", "features", "classification", "using", "high", "its" ], [ "attacks", "to", "adversarial", "training", "the", "we", "robustness", "attack", "against", "examples" ], [ "paper", "this", "problem", "approach", "the", "distribution", "process", "new", "work", "been" ], [ "generative", "fidelity", "models", "these", "of", "trained", "language", "performance", "inference", "have" ], [ "methods", "method", "datasets", "proposed", "deep", "feature", "existing", "classification", "prediction", "large" ], [ "be", "can", "used", "which", "how", "or", "it", "quantum", "they", "image" ], [ "the", "more", "are", "these", "algorithms", "when", "where", "they", "different", "their" ], [ "the", "graphs", "and", "node", "nodes", "graph", "representation", "gnns", "we", "which" ], [ "detection", "on", "proposed", "based", "the", "feature", "propose", "system", "have", "recognition" ], [ "by", "the", "time", "feature", "series", "approach", "images", "high", "between", "into" ], [ "gradient", "and", "stochastic", "optimization", "policy", "function", "convex", "convergence", "problems", "optimal" ], [ "framework", "results", "our", "approach", "method", "we", "state", "at", "art", "first" ], [ "linear", "problem", "matrix", "algorithm", "proposed", "to", "algorithms", "sample", "clustering", "rank" ], [ "from", "task", "using", "tasks", "language", "representations", "different", "knowledge", "information", "object" ] ]
3.492299
all-MiniLM-L6-v2
0.566667
0.022383
0.15836
0.63278
ArXiv ML Papers
NMF
45
30
[ [ "two", "of", "problem", "to", "proposed", "results", "between", "the", "information", "using" ], [ "to", "the", "due", "use", "their", "it", "human", "have", "search", "framework" ], [ "large", "art", "set", "terms", "state", "analysis", "of", "number", "new", "these" ], [ "the", "and", "both", "two", "ml", "research", "accuracy", "including", "their", "study" ], [ "and", "present", "which", "we", "the", "use", "show", "propose", "where", "consider" ], [ "or", "in", "been", "work", "has", "many", "the", "applications", "have", "recent" ], [ "also", "time", "for", "used", "each", "prediction", "test", "one", "using", "the" ], [ "training", "data", "real", "time", "classification", "synthetic", "large", "to", "supervised", "from" ], [ "machine", "reinforcement", "of", "learning", "tasks", "deep", "supervised", "has", "policy", "metric" ], [ "where", "which", "one", "it", "is", "not", "information", "at", "when", "than" ], [ "on", "performance", "datasets", "tasks", "or", "task", "work", "focus", "classification", "trained" ], [ "than", "not", "find", "representations", "show", "that", "the", "results", "also", "more" ], [ "to", "neural", "networks", "deep", "network", "convolutional", "training", "layer", "and", "architecture" ], [ "in", "performance", "parameters", "model", "which", "and", "inference", "the", "trained", "proposed" ], [ "well", "such", "to", "been", "as", "have", "these", "or", "the", "not" ], [ "performance", "with", "accuracy", "high", "compared", "results", "low", "dataset", "while", "achieve" ], [ "methods", "method", "proposed", "optimization", "existing", "gradient", "propose", "and", "datasets", "problems" ], [ "we", "attacks", "training", "to", "the", "adversarial", "robustness", "against", "attack", "examples" ], [ "process", "been", "new", "work", "problem", "paper", "this", "approach", "distribution", "system" ], [ "task", "information", "language", "from", "representations", "different", "features", "using", "knowledge", "tasks" ], [ "detection", "based", "proposed", "the", "on", "feature", "have", "propose", "system", "information" ], [ "it", "be", "or", "can", "how", "quantum", "which", "used", "they", "systems" ], [ "time", "the", "by", "approach", "series", "high", "between", "feature", "while", "into" ], [ "graph", "the", "which", "nodes", "and", "gnns", "graphs", "node", "we", "representation" ], [ "these", "more", "where", "when", "some", "algorithms", "their", "different", "are", "they" ], [ "inference", "trained", "fidelity", "models", "language", "machine", "these", "of", "generative", "performance" ], [ "an", "network", "input", "features", "which", "its", "and", "using", "approach", "method" ], [ "results", "state", "our", "approach", "framework", "method", "to", "we", "at", "art" ], [ "gradient", "problem", "algorithms", "algorithm", "optimization", "convergence", "to", "problems", "which", "stochastic" ], [ "text", "images", "image", "gan", "segmentation", "classification", "deep", "which", "we", "performance" ] ]
3.537155
all-MiniLM-L6-v2
0.543333
0.020288
0.16415
0.626309
ArXiv ML Papers
NMF
46
30
[ [ "proposed", "the", "to", "results", "of", "between", "by", "two", "problem", "information" ], [ "the", "it", "due", "search", "to", "human", "use", "this", "end", "their" ], [ "these", "art", "set", "analysis", "of", "terms", "the", "number", "large", "new" ], [ "and", "both", "accuracy", "the", "ml", "two", "between", "research", "including", "their" ], [ "problem", "which", "the", "show", "we", "propose", "demonstrate", "present", "use", "and" ], [ "or", "been", "in", "the", "has", "work", "this", "paper", "applications", "many" ], [ "time", "test", "also", "using", "regression", "used", "the", "for", "each", "one" ], [ "supervised", "and", "large", "training", "data", "real", "labeled", "synthetic", "classification", "time" ], [ "reinforcement", "deep", "machine", "tasks", "learning", "of", "supervised", "multi", "has", "metric" ], [ "this", "which", "it", "is", "where", "not", "one", "at", "than", "when" ], [ "datasets", "performance", "tasks", "on", "classification", "task", "work", "or", "trained", "art" ], [ "not", "this", "show", "than", "that", "representations", "the", "results", "find", "more" ], [ "neural", "deep", "network", "networks", "convolutional", "to", "and", "training", "performance", "layer" ], [ "performance", "the", "which", "and", "model", "trained", "in", "inference", "parameters", "speech" ], [ "such", "have", "or", "as", "to", "well", "these", "been", "the", "not" ], [ "compared", "low", "results", "accuracy", "high", "with", "dataset", "performance", "while", "achieve" ], [ "method", "prediction", "propose", "proposed", "deep", "feature", "and", "using", "classification", "optimization" ], [ "adversarial", "training", "attacks", "to", "the", "we", "robustness", "against", "attack", "examples" ], [ "time", "by", "images", "paper", "this", "system", "approach", "at", "while", "into" ], [ "using", "tasks", "knowledge", "language", "representations", "object", "task", "from", "different", "information" ], [ "on", "detection", "the", "have", "feature", "recognition", "system", "propose", "proposed", "based" ], [ "can", "which", "be", "used", "or", "quantum", "it", "how", "they", "image" ], [ "algorithms", "are", "these", "they", "when", "where", "different", "the", "more", "their" ], [ "graphs", "and", "node", "nodes", "the", "representation", "gnns", "graph", "we", "structure" ], [ "have", "methods", "datasets", "which", "performance", "and", "been", "both", "existing", "has" ], [ "fidelity", "of", "performance", "generative", "prediction", "language", "these", "models", "trained", "have" ], [ "policy", "and", "function", "convex", "optimization", "stochastic", "gradient", "problems", "convergence", "optimal" ], [ "results", "to", "framework", "in", "art", "our", "approach", "at", "we", "state" ], [ "problem", "algorithm", "algorithms", "to", "matrix", "proposed", "clustering", "linear", "sample", "rank" ], [ "which", "image", "input", "features", "an", "approach", "high", "its", "using", "classification" ] ]
3.519023
all-MiniLM-L6-v2
0.556667
0.019611
0.163618
0.617932
ArXiv ML Papers
NMF
43
40
[ [ "results", "of", "strategy", "between", "proposed", "information", "the", "test", "was", "error" ], [ "due", "their", "approach", "framework", "use", "to", "compared", "human", "able", "often" ], [ "analysis", "number", "of", "set", "and", "these", "new", "large", "art", "terms" ], [ "including", "and", "both", "two", "accuracy", "between", "ml", "research", "study", "respectively" ], [ "demonstrate", "we", "present", "new", "the", "propose", "use", "show", "study", "these" ], [ "or", "work", "in", "particular", "addition", "space", "order", "terms", "context", "many" ], [ "test", "analysis", "used", "one", "using", "for", "also", "the", "regression", "each" ], [ "driven", "data", "and", "synthetic", "classification", "large", "real", "high", "techniques", "labeled" ], [ "learning", "machine", "of", "agent", "metric", "reinforcement", "supervised", "policy", "learn", "the" ], [ "when", "where", "one", "the", "at", "than", "it", "is", "its", "information" ], [ "previous", "work", "or", "datasets", "on", "focus", "real", "rely", "all", "dataset" ], [ "show", "than", "results", "also", "find", "not", "more", "the", "that", "representations" ], [ "and", "networks", "neural", "convolutional", "over", "architectures", "have", "cnns", "physics", "weights" ], [ "uncertainty", "inference", "predictions", "parameters", "trained", "privacy", "the", "model", "proposed", "to" ], [ "such", "to", "the", "as", "well", "or", "known", "through", "using", "processing" ], [ "with", "accuracy", "results", "all", "dataset", "high", "while", "compared", "the", "respect" ], [ "between", "high", "by", "into", "approach", "the", "filter", "while", "function", "class" ], [ "of", "adversarial", "to", "robustness", "against", "attack", "attacks", "the", "detection", "and" ], [ "approach", "work", "new", "paper", "this", "process", "system", "challenge", "presents", "video" ], [ "language", "source", "domain", "tasks", "target", "multi", "knowledge", "task", "different", "representations" ], [ "and", "samples", "distribution", "problems", "problem", "optimal", "the", "matrix", "sample", "rank" ], [ "be", "used", "how", "it", "can", "or", "systems", "quantum", "they", "at" ], [ "these", "they", "algorithms", "where", "when", "are", "the", "more", "some", "different" ], [ "node", "graph", "graphs", "and", "the", "nodes", "representation", "gnns", "we", "clustering" ], [ "datasets", "proposed", "prediction", "methods", "method", "propose", "large", "existing", "noise", "optimization" ], [ "prediction", "generative", "inference", "trained", "fidelity", "models", "language", "of", "machine", "these" ], [ "gradient", "stochastic", "optimization", "policy", "convergence", "and", "convex", "function", "problems", "descent" ], [ "results", "framework", "approach", "method", "our", "we", "first", "state", "work", "best" ], [ "algorithm", "proposed", "clustering", "to", "algorithms", "the", "label", "regret", "efficient", "multi" ], [ "features", "using", "from", "set", "information", "representations", "language", "between", "object", "different" ], [ "detection", "based", "proposed", "information", "the", "propose", "system", "systems", "user", "state" ], [ "time", "of", "real", "series", "at", "forecasting", "temporal", "world", "state", "hierarchical" ], [ "which", "of", "not", "propose", "two", "feature", "information", "only", "to", "state" ], [ "state", "performance", "classification", "of", "art", "accuracy", "better", "achieve", "high", "using" ], [ "images", "image", "gan", "segmentation", "text", "generative", "resolution", "object", "we", "visual" ], [ "neural", "layer", "network", "to", "structure", "prediction", "show", "nodes", "architecture", "convolutional" ], [ "deep", "feature", "speech", "learning", "features", "to", "detection", "accuracy", "results", "ensemble" ], [ "an", "features", "input", "approach", "using", "its", "system", "technique", "high", "it" ], [ "train", "during", "set", "loss", "training", "pre", "trained", "to", "we", "federated" ], [ "has", "have", "it", "been", "not", "these", "machine", "the", "methods", "research" ] ]
5.106921
all-MiniLM-L6-v2
0.5825
0.008913
0.149585
0.670219
ArXiv ML Papers
NMF
44
40
[ [ "the", "error", "of", "was", "proposed", "between", "results", "information", "two", "test" ], [ "due", "it", "to", "human", "search", "use", "their", "have", "approach", "able" ], [ "large", "these", "number", "set", "of", "terms", "art", "analysis", "and", "new" ], [ "both", "and", "research", "between", "ml", "study", "accuracy", "their", "including", "two" ], [ "provide", "use", "show", "we", "demonstrate", "propose", "present", "new", "consider", "study" ], [ "many", "high", "recent", "have", "has", "applications", "in", "work", "or", "been" ], [ "used", "for", "test", "regression", "the", "analysis", "also", "one", "using", "large" ], [ "real", "the", "techniques", "large", "high", "driven", "synthetic", "and", "classification", "data" ], [ "learning", "machine", "active", "framework", "has", "supervised", "metric", "learn", "of", "algorithms" ], [ "is", "it", "when", "where", "the", "at", "one", "than", "its", "information" ], [ "the", "work", "datasets", "or", "real", "on", "performance", "previous", "focus", "rely" ], [ "show", "that", "not", "results", "than", "also", "the", "more", "representations", "find" ], [ "neural", "network", "convolutional", "networks", "and", "layer", "architecture", "performance", "architectures", "layers" ], [ "model", "the", "performance", "inference", "and", "trained", "uncertainty", "predictions", "parameters", "in" ], [ "to", "well", "the", "such", "as", "have", "using", "been", "these", "or" ], [ "with", "accuracy", "performance", "compared", "the", "results", "dataset", "high", "low", "while" ], [ "the", "attacks", "attack", "to", "adversarial", "of", "robustness", "detection", "against", "and" ], [ "training", "performance", "trained", "in", "to", "during", "and", "we", "loss", "pre" ], [ "paper", "approach", "problem", "been", "this", "new", "process", "work", "the", "challenge" ], [ "algorithms", "these", "where", "are", "more", "the", "when", "they", "their", "some" ], [ "system", "based", "the", "propose", "proposed", "systems", "detection", "have", "recognition", "user" ], [ "how", "be", "can", "used", "it", "or", "quantum", "they", "to", "systems" ], [ "between", "filter", "by", "while", "the", "dimensional", "approach", "high", "into", "images" ], [ "graph", "node", "nodes", "and", "gnns", "graphs", "the", "representation", "we", "structure" ], [ "performance", "datasets", "methods", "have", "the", "both", "existing", "and", "many", "other" ], [ "fidelity", "models", "generative", "machine", "these", "trained", "inference", "language", "prediction", "high" ], [ "stochastic", "gradient", "and", "optimization", "problems", "function", "convergence", "order", "descent", "convex" ], [ "framework", "our", "approach", "results", "we", "best", "state", "art", "first", "experiments" ], [ "an", "the", "it", "input", "approach", "high", "technique", "system", "its", "using" ], [ "from", "using", "different", "object", "set", "language", "information", "representations", "samples", "between" ], [ "language", "knowledge", "multi", "tasks", "task", "prediction", "performance", "different", "each", "meta" ], [ "series", "time", "of", "in", "at", "real", "forecasting", "temporal", "world", "state" ], [ "two", "which", "to", "propose", "of", "not", "only", "has", "distribution", "local" ], [ "to", "problem", "algorithms", "matrix", "the", "algorithm", "proposed", "sample", "clustering", "linear" ], [ "segmentation", "images", "image", "text", "gan", "generative", "we", "to", "resolution", "object" ], [ "domains", "domain", "knowledge", "transfer", "source", "the", "target", "to", "adaptation", "proposed" ], [ "deep", "have", "been", "learning", "has", "ensemble", "accuracy", "recognition", "paper", "results" ], [ "prediction", "method", "proposed", "propose", "and", "noise", "optimization", "large", "using", "low" ], [ "agents", "policy", "rl", "reward", "reinforcement", "state", "the", "and", "agent", "learning" ], [ "and", "features", "feature", "to", "classification", "classifier", "attention", "in", "the", "selection" ] ]
5.076314
all-MiniLM-L6-v2
0.56
0.017912
0.157493
0.644569
ArXiv ML Papers
NMF
45
40
[ [ "the", "of", "proposed", "results", "between", "error", "two", "information", "test", "problem" ], [ "use", "their", "to", "due", "it", "search", "framework", "human", "have", "however" ], [ "and", "of", "number", "these", "set", "art", "terms", "large", "analysis", "study" ], [ "both", "and", "two", "ml", "research", "accuracy", "study", "including", "their", "between" ], [ "we", "propose", "demonstrate", "present", "use", "show", "new", "provide", "consider", "study" ], [ "has", "many", "applications", "in", "or", "been", "work", "recent", "high", "particular" ], [ "test", "analysis", "for", "also", "used", "regression", "large", "one", "using", "set" ], [ "synthetic", "large", "classification", "driven", "data", "training", "real", "high", "labeled", "techniques" ], [ "the", "metric", "learning", "of", "supervised", "machine", "framework", "active", "learn", "to" ], [ "it", "the", "is", "one", "when", "where", "than", "at", "and", "its" ], [ "performance", "datasets", "on", "work", "or", "focus", "previous", "real", "rely", "all" ], [ "show", "that", "not", "the", "results", "more", "than", "also", "find", "representations" ], [ "neural", "architectures", "convolutional", "more", "have", "training", "over", "cnns", "and", "networks" ], [ "trained", "parameters", "model", "inference", "performance", "the", "and", "to", "uncertainty", "predictions" ], [ "well", "or", "not", "as", "have", "the", "to", "these", "such", "been" ], [ "compared", "accuracy", "results", "with", "performance", "high", "dataset", "while", "low", "achieve" ], [ "method", "proposed", "prediction", "the", "and", "large", "propose", "noise", "using", "optimization" ], [ "to", "robustness", "the", "training", "adversarial", "attacks", "we", "attack", "against", "examples" ], [ "paper", "the", "problem", "distribution", "this", "process", "new", "approach", "been", "work" ], [ "which", "not", "has", "of", "propose", "two", "the", "only", "distribution", "information" ], [ "its", "approach", "using", "input", "an", "the", "technique", "high", "system", "it" ], [ "or", "how", "it", "used", "they", "quantum", "be", "can", "systems", "at" ], [ "are", "these", "where", "the", "more", "they", "when", "different", "their", "some" ], [ "nodes", "we", "graph", "and", "representation", "graphs", "gnns", "the", "node", "structure" ], [ "training", "different", "knowledge", "source", "language", "domain", "performance", "target", "task", "tasks" ], [ "these", "models", "fidelity", "machine", "trained", "generative", "inference", "performance", "language", "prediction" ], [ "gradient", "convergence", "function", "and", "convex", "stochastic", "optimization", "problems", "descent", "order" ], [ "our", "approach", "results", "first", "framework", "we", "state", "to", "art", "best" ], [ "problem", "the", "to", "matrix", "sample", "algorithm", "regret", "rank", "clustering", "linear" ], [ "set", "object", "samples", "language", "using", "information", "different", "from", "between", "representations" ], [ "system", "based", "have", "detection", "proposed", "propose", "systems", "the", "recognition", "user" ], [ "series", "in", "time", "real", "at", "of", "forecasting", "temporal", "state", "world" ], [ "other", "datasets", "the", "existing", "have", "both", "performance", "methods", "many", "been" ], [ "algorithms", "to", "two", "the", "problems", "other", "search", "these", "variance", "more" ], [ "between", "while", "the", "high", "dimensional", "images", "by", "approach", "into", "filter" ], [ "neural", "to", "performance", "layer", "network", "training", "structure", "architecture", "convolutional", "prediction" ], [ "been", "deep", "learning", "have", "ensemble", "accuracy", "has", "recognition", "paper", "speech" ], [ "segmentation", "resolution", "images", "image", "object", "gan", "generative", "we", "text", "trained" ], [ "the", "learning", "policy", "agent", "and", "reward", "state", "agents", "reinforcement", "rl" ], [ "selection", "in", "features", "attention", "to", "the", "feature", "classification", "classifier", "and" ] ]
5.337978
all-MiniLM-L6-v2
0.5525
0.016879
0.153486
0.64701
ArXiv ML Papers
NMF
46
40
[ [ "the", "problem", "error", "results", "of", "between", "proposed", "two", "space", "information" ], [ "use", "due", "to", "it", "human", "their", "have", "framework", "search", "approach" ], [ "terms", "analysis", "set", "number", "of", "these", "new", "and", "large", "art" ], [ "research", "two", "both", "and", "ml", "including", "between", "accuracy", "their", "study" ], [ "new", "propose", "we", "demonstrate", "study", "provide", "consider", "show", "present", "use" ], [ "has", "or", "the", "been", "recent", "many", "work", "applications", "have", "in" ], [ "one", "using", "used", "also", "the", "test", "for", "analysis", "regression", "large" ], [ "large", "driven", "real", "data", "classification", "synthetic", "high", "and", "techniques", "labeled" ], [ "the", "of", "supervised", "reinforcement", "machine", "learning", "metric", "policy", "learn", "agent" ], [ "when", "is", "one", "where", "it", "than", "the", "at", "its", "and" ], [ "on", "datasets", "the", "focus", "real", "work", "or", "rely", "previous", "all" ], [ "not", "show", "more", "the", "results", "also", "that", "than", "find", "representations" ], [ "network", "structure", "layer", "architecture", "nodes", "neural", "to", "and", "prediction", "convolutional" ], [ "the", "model", "trained", "and", "in", "inference", "uncertainty", "parameters", "to", "predictions" ], [ "well", "as", "such", "to", "or", "the", "have", "these", "using", "been" ], [ "with", "low", "dataset", "accuracy", "results", "high", "compared", "the", "respect", "all" ], [ "an", "approach", "the", "input", "its", "using", "system", "it", "high", "technique" ], [ "to", "attacks", "the", "against", "of", "attack", "adversarial", "robustness", "detection", "and" ], [ "approach", "paper", "distribution", "new", "the", "problem", "work", "process", "this", "been" ], [ "state", "the", "distribution", "which", "not", "of", "only", "propose", "two", "has" ], [ "the", "based", "detection", "proposed", "system", "propose", "approaches", "user", "systems", "have" ], [ "they", "can", "quantum", "or", "be", "used", "it", "how", "to", "systems" ], [ "these", "more", "are", "some", "algorithms", "where", "the", "their", "when", "they" ], [ "representation", "graphs", "graph", "and", "nodes", "node", "the", "gnns", "we", "clustering" ], [ "by", "the", "approach", "between", "high", "while", "filter", "dimensional", "function", "into" ], [ "models", "machine", "generative", "trained", "these", "fidelity", "language", "of", "prediction", "inference" ], [ "policy", "stochastic", "function", "optimization", "gradient", "and", "convergence", "problems", "convex", "optimal" ], [ "approach", "results", "framework", "our", "we", "experiments", "art", "state", "to", "first" ], [ "algorithms", "algorithm", "proposed", "to", "clustering", "matrix", "the", "problem", "linear", "sample" ], [ "object", "different", "using", "set", "from", "information", "between", "samples", "language", "representations" ], [ "domain", "tasks", "task", "knowledge", "language", "multi", "different", "to", "target", "source" ], [ "series", "time", "real", "in", "world", "of", "at", "temporal", "forecasting", "state" ], [ "have", "the", "methods", "datasets", "problems", "both", "other", "existing", "been", "many" ], [ "performance", "of", "state", "art", "to", "accuracy", "high", "proposed", "classification", "achieve" ], [ "image", "generative", "segmentation", "images", "we", "text", "gan", "resolution", "trained", "object" ], [ "architectures", "cnns", "over", "neural", "and", "networks", "convolutional", "physics", "have", "recurrent" ], [ "deep", "learning", "to", "been", "has", "accuracy", "ensemble", "have", "paper", "recognition" ], [ "the", "features", "in", "classification", "feature", "selection", "to", "attention", "classifier", "information" ], [ "set", "trained", "training", "in", "pre", "during", "federated", "we", "to", "loss" ], [ "proposed", "optimization", "low", "prediction", "large", "using", "method", "noise", "propose", "existing" ] ]
5.098675
all-MiniLM-L6-v2
0.5475
0.01484
0.159135
0.638216
ArXiv ML Papers
NMF
43
50
[ [ "of", "test", "proposed", "error", "the", "was", "between", "results", "strategy", "one" ], [ "to", "due", "use", "human", "able", "approach", "their", "often", "compared", "framework" ], [ "set", "number", "these", "of", "analysis", "large", "terms", "study", "parameters", "new" ], [ "accuracy", "including", "research", "and", "ml", "two", "both", "study", "between", "respectively" ], [ "we", "use", "propose", "show", "present", "new", "study", "demonstrate", "provide", "these" ], [ "context", "in", "addition", "computational", "terms", "work", "may", "particular", "many", "or" ], [ "test", "one", "used", "regression", "for", "also", "analysis", "each", "large", "fast" ], [ "data", "real", "framework", "labeled", "driven", "high", "large", "classification", "synthetic", "techniques" ], [ "learning", "of", "metric", "active", "machine", "supervised", "learn", "framework", "reinforcement", "self" ], [ "one", "when", "it", "is", "and", "the", "than", "its", "at", "where" ], [ "on", "work", "datasets", "focus", "or", "rely", "all", "user", "real", "previous" ], [ "these", "where", "more", "they", "when", "are", "their", "some", "new", "or" ], [ "network", "structure", "layer", "neural", "architecture", "nodes", "convolutional", "show", "prediction", "learned" ], [ "parameters", "trained", "uncertainty", "inference", "while", "the", "predictions", "model", "free", "global" ], [ "through", "to", "as", "well", "such", "or", "using", "processing", "known", "continuous" ], [ "with", "accuracy", "results", "dataset", "all", "respect", "compared", "high", "while", "multiple" ], [ "art", "of", "state", "the", "propose", "datasets", "classification", "results", "over", "approaches" ], [ "trained", "loss", "we", "training", "during", "train", "pre", "set", "federated", "large" ], [ "approach", "paper", "process", "presents", "new", "challenge", "work", "this", "system", "class" ], [ "user", "the", "between", "information", "semantic", "channel", "two", "order", "variables", "label" ], [ "has", "have", "it", "been", "not", "these", "machine", "the", "their", "research" ], [ "can", "be", "used", "to", "the", "it", "they", "how", "quantum", "or" ], [ "filter", "between", "into", "approach", "high", "by", "the", "while", "however", "images" ], [ "the", "graph", "graphs", "nodes", "gnns", "node", "representation", "and", "we", "clustering" ], [ "than", "more", "results", "show", "also", "that", "not", "find", "prior", "representations" ], [ "models", "trained", "fidelity", "generative", "inference", "prediction", "these", "at", "language", "machine" ], [ "system", "framework", "high", "approach", "input", "technique", "3d", "the", "its", "an" ], [ "our", "results", "first", "framework", "approach", "best", "work", "at", "experiments", "only" ], [ "to", "the", "algorithm", "proposed", "selection", "clustering", "sample", "exchange", "is", "convergence" ], [ "from", "using", "different", "the", "set", "between", "object", "language", "knowledge", "representations" ], [ "based", "the", "proposed", "systems", "user", "on", "flow", "propose", "approaches", "is" ], [ "forecasting", "real", "series", "of", "hierarchical", "time", "world", "temporal", "forecast", "at" ], [ "methods", "datasets", "both", "many", "existing", "other", "and", "be", "linear", "hierarchical" ], [ "task", "knowledge", "different", "multi", "prediction", "tasks", "solving", "each", "meta", "language" ], [ "images", "image", "gan", "resolution", "medical", "generative", "text", "trained", "segmentation", "we" ], [ "networks", "neural", "convolutional", "over", "architectures", "physics", "and", "have", "cnns", "more" ], [ "adversarial", "attacks", "to", "attack", "the", "against", "robustness", "and", "of", "robust" ], [ "proposed", "the", "method", "prediction", "has", "propose", "large", "noise", "using", "new" ], [ "control", "policy", "the", "policies", "learning", "and", "reinforcement", "rl", "reward", "optimal" ], [ "results", "recognition", "deep", "paper", "ensemble", "learning", "accuracy", "in", "to", "dnn" ], [ "optimization", "stochastic", "convergence", "gradient", "function", "problems", "convex", "descent", "and", "order" ], [ "better", "achieve", "their", "computing", "high", "cost", "to", "performance", "accuracy", "design" ], [ "domain", "target", "the", "source", "adaptation", "knowledge", "transfer", "domains", "to", "datasets" ], [ "dataset", "using", "object", "detection", "system", "classification", "the", "rate", "detect", "attacks" ], [ "of", "two", "not", "which", "to", "the", "propose", "distribution", "only", "has" ], [ "classification", "feature", "features", "and", "of", "the", "selection", "to", "results", "attention" ], [ "two", "more", "algorithms", "problems", "variance", "search", "other", "than", "the", "robust" ], [ "matrix", "rank", "problems", "problem", "low", "optimal", "samples", "distribution", "linear", "given" ], [ "agents", "communication", "agent", "to", "each", "where", "regret", "the", "at", "multi" ], [ "visual", "and", "language", "speech", "to", "audio", "the", "text", "end", "attention" ] ]
6.24702
all-MiniLM-L6-v2
0.56
0.007192
0.147988
0.67742
ArXiv ML Papers
NMF
44
50
[ [ "between", "information", "the", "error", "proposed", "results", "was", "of", "strategy", "conditions" ], [ "compared", "due", "to", "use", "able", "their", "approach", "framework", "often", "end" ], [ "of", "and", "number", "set", "analysis", "these", "terms", "art", "large", "study" ], [ "various", "respectively", "research", "including", "two", "between", "study", "both", "and", "covid" ], [ "use", "show", "demonstrate", "provide", "present", "study", "new", "we", "propose", "introduce" ], [ "context", "particular", "work", "or", "in", "addition", "terms", "order", "many", "computational" ], [ "test", "used", "one", "for", "analysis", "also", "large", "regression", "fast", "useful" ], [ "data", "real", "large", "driven", "synthetic", "labeled", "techniques", "process", "sets", "framework" ], [ "learning", "supervised", "metric", "of", "machine", "framework", "active", "learn", "self", "has" ], [ "show", "that", "not", "than", "also", "more", "find", "results", "representations", "such" ], [ "datasets", "on", "work", "or", "focus", "user", "all", "previous", "rely", "trained" ], [ "and", "prediction", "accuracy", "is", "of", "dataset", "traffic", "two", "speed", "market" ], [ "structure", "show", "layer", "architecture", "neural", "network", "convolutional", "learned", "nodes", "teacher" ], [ "trained", "privacy", "parameters", "inference", "model", "while", "free", "the", "predictions", "global" ], [ "well", "to", "such", "as", "or", "through", "known", "using", "processing", "these" ], [ "compared", "accuracy", "dataset", "results", "with", "high", "respect", "while", "low", "all" ], [ "noise", "large", "propose", "proposed", "method", "optimization", "space", "has", "the", "low" ], [ "trained", "training", "loss", "set", "we", "during", "pre", "federated", "train", "large" ], [ "this", "paper", "work", "new", "process", "approach", "challenge", "presents", "strategies", "approaches" ], [ "ml", "such", "how", "human", "quantum", "machine", "system", "systems", "techniques", "user" ], [ "have", "the", "it", "recent", "these", "but", "been", "not", "has", "however" ], [ "it", "be", "can", "used", "how", "or", "they", "at", "to", "the" ], [ "where", "are", "the", "these", "they", "when", "more", "their", "some", "different" ], [ "node", "and", "graph", "graphs", "representation", "the", "nodes", "gnns", "we", "structure" ], [ "accuracy", "classifier", "classification", "class", "high", "label", "state", "the", "classifiers", "art" ], [ "models", "trained", "language", "fidelity", "generative", "these", "inference", "at", "modeling", "high" ], [ "approach", "the", "framework", "3d", "technique", "an", "its", "input", "using", "high" ], [ "our", "approach", "framework", "results", "state", "first", "art", "best", "work", "experiments" ], [ "algorithm", "the", "to", "proposed", "sample", "matrix", "regret", "exchange", "clustering", "selection" ], [ "between", "set", "from", "language", "information", "using", "different", "the", "object", "representations" ], [ "flow", "the", "propose", "on", "proposed", "information", "based", "approaches", "user", "state" ], [ "series", "real", "forecasting", "at", "time", "world", "temporal", "of", "state", "hierarchical" ], [ "datasets", "many", "other", "both", "existing", "methods", "linear", "large", "several", "problems" ], [ "it", "is", "one", "the", "where", "than", "at", "when", "its", "information" ], [ "segmentation", "gan", "generative", "image", "text", "images", "resolution", "we", "visual", "trained" ], [ "architectures", "and", "convolutional", "networks", "neural", "over", "have", "more", "cnns", "physics" ], [ "against", "to", "and", "adversarial", "of", "the", "robust", "attack", "robustness", "attacks" ], [ "using", "in", "detect", "attacks", "the", "detection", "object", "rate", "dataset", "system" ], [ "policy", "the", "rl", "agent", "reinforcement", "state", "action", "agents", "policies", "reward" ], [ "learning", "ensemble", "results", "recognition", "deep", "accuracy", "speech", "in", "paper", "design" ], [ "convex", "optimization", "descent", "function", "problems", "order", "non", "stochastic", "gradient", "convergence" ], [ "achieve", "performance", "state", "of", "better", "cost", "proposed", "art", "design", "different" ], [ "than", "other", "more", "two", "algorithms", "these", "problems", "variance", "search", "the" ], [ "and", "features", "feature", "the", "selection", "information", "to", "attention", "interaction", "speech" ], [ "not", "of", "novel", "only", "two", "propose", "which", "to", "local", "framework" ], [ "distributions", "distribution", "estimation", "over", "samples", "bayesian", "sample", "bounds", "variational", "inference" ], [ "rank", "problem", "matrix", "optimal", "the", "problems", "solution", "optimization", "low", "given" ], [ "the", "between", "by", "approach", "into", "filter", "function", "while", "images", "dimensional" ], [ "tasks", "task", "to", "language", "meta", "knowledge", "different", "multi", "each", "representations" ], [ "the", "knowledge", "domains", "domain", "source", "target", "transfer", "to", "adaptation", "proposed" ] ]
6.245105
all-MiniLM-L6-v2
0.59
0.007457
0.150291
0.68471
ArXiv ML Papers
NMF
45
50
[ [ "of", "between", "proposed", "results", "error", "the", "was", "strategy", "conditions", "one" ], [ "often", "able", "use", "approach", "to", "their", "compared", "due", "framework", "ability" ], [ "analysis", "terms", "of", "set", "large", "art", "these", "number", "and", "state" ], [ "research", "and", "study", "including", "both", "two", "between", "respectively", "their", "various" ], [ "introduce", "new", "study", "provide", "present", "propose", "we", "show", "use", "demonstrate" ], [ "addition", "terms", "work", "or", "many", "particular", "in", "context", "computational", "applications" ], [ "test", "used", "for", "one", "also", "analysis", "large", "fast", "each", "useful" ], [ "driven", "synthetic", "real", "high", "classification", "large", "data", "labeled", "techniques", "framework" ], [ "of", "learning", "supervised", "metric", "machine", "to", "learn", "framework", "active", "self" ], [ "where", "than", "one", "its", "it", "is", "the", "when", "and", "at" ], [ "datasets", "on", "work", "or", "rely", "focus", "previous", "all", "trained", "real" ], [ "where", "when", "more", "some", "their", "are", "different", "they", "these", "new" ], [ "neural", "network", "structure", "show", "layer", "architecture", "convolutional", "nodes", "learned", "to" ], [ "trained", "model", "the", "parameters", "inference", "predictions", "while", "free", "train", "global" ], [ "or", "using", "well", "as", "to", "through", "such", "known", "processing", "these" ], [ "with", "accuracy", "compared", "high", "results", "dataset", "respect", "all", "while", "low" ], [ "has", "the", "which", "not", "propose", "novel", "two", "of", "only", "most" ], [ "trained", "we", "loss", "training", "during", "set", "pre", "datasets", "train", "federated" ], [ "how", "they", "used", "or", "be", "can", "it", "at", "any", "show" ], [ "these", "but", "it", "the", "been", "not", "have", "has", "using", "recent" ], [ "process", "work", "approach", "paper", "the", "this", "new", "presents", "challenge", "strategies" ], [ "ml", "how", "system", "machine", "human", "systems", "quantum", "such", "techniques", "research" ], [ "by", "approach", "between", "the", "into", "high", "filter", "while", "function", "images" ], [ "node", "and", "the", "graph", "graphs", "nodes", "representation", "we", "gnns", "structure" ], [ "show", "than", "results", "not", "that", "more", "representations", "find", "also", "such" ], [ "inference", "generative", "language", "models", "surrogate", "modeling", "at", "fidelity", "trained", "these" ], [ "input", "an", "its", "technique", "the", "approach", "framework", "high", "3d", "shape" ], [ "approach", "experiments", "state", "our", "results", "first", "best", "art", "work", "framework" ], [ "algorithm", "to", "the", "proposed", "sample", "matrix", "regret", "exchange", "clustering", "selection" ], [ "using", "between", "language", "set", "from", "different", "object", "features", "representations", "derived" ], [ "flow", "proposed", "is", "approaches", "based", "on", "user", "the", "propose", "state" ], [ "time", "at", "real", "series", "forecasting", "of", "temporal", "world", "state", "each" ], [ "existing", "datasets", "methods", "other", "the", "both", "many", "linear", "art", "large" ], [ "task", "different", "meta", "knowledge", "language", "tasks", "representations", "multi", "the", "trained" ], [ "image", "resolution", "images", "segmentation", "gan", "text", "generative", "we", "visual", "trained" ], [ "neural", "networks", "convolutional", "and", "over", "architectures", "have", "cnns", "more", "physics" ], [ "robustness", "the", "attack", "attacks", "to", "of", "against", "adversarial", "and", "robust" ], [ "performance", "art", "state", "of", "the", "achieve", "better", "proposed", "cost", "high" ], [ "rl", "agent", "the", "policy", "reinforcement", "reward", "state", "action", "policies", "agents" ], [ "classification", "the", "feature", "to", "features", "and", "selection", "of", "speech", "attention" ], [ "stochastic", "function", "gradient", "convex", "convergence", "optimization", "problems", "descent", "order", "non" ], [ "to", "learning", "deep", "ensemble", "results", "recognition", "accuracy", "speech", "paper", "design" ], [ "the", "algorithms", "search", "two", "other", "problems", "variance", "to", "more", "than" ], [ "prediction", "is", "speed", "accuracy", "two", "of", "traffic", "dataset", "market", "and" ], [ "optimal", "problem", "problems", "matrix", "rank", "solution", "optimization", "given", "linear", "low" ], [ "over", "samples", "bayesian", "estimation", "distributions", "distribution", "sample", "bounds", "variational", "inference" ], [ "propose", "the", "proposed", "noise", "method", "large", "space", "has", "optimization", "low" ], [ "detection", "using", "classification", "object", "classifier", "accuracy", "dataset", "detect", "in", "rate" ], [ "each", "the", "between", "information", "channel", "two", "multi", "user", "local", "semantic" ], [ "target", "domain", "the", "source", "transfer", "knowledge", "adaptation", "domains", "to", "datasets" ] ]
6.301346
all-MiniLM-L6-v2
0.58
0.009219
0.151183
0.683603
ArXiv ML Papers
NMF
46
50
[ [ "two", "results", "between", "proposed", "error", "the", "of", "information", "test", "was" ], [ "framework", "their", "use", "due", "to", "able", "compared", "approach", "often", "ability" ], [ "set", "analysis", "and", "these", "terms", "number", "of", "study", "parameters", "different" ], [ "two", "and", "both", "including", "between", "study", "research", "respectively", "various", "present" ], [ "these", "we", "study", "new", "show", "demonstrate", "propose", "provide", "present", "use" ], [ "work", "particular", "or", "terms", "in", "context", "space", "addition", "many", "order" ], [ "one", "each", "also", "analysis", "test", "regression", "for", "used", "day", "specific" ], [ "synthetic", "data", "driven", "process", "techniques", "real", "sets", "framework", "privacy", "labeled" ], [ "learning", "of", "framework", "supervised", "machine", "metric", "learn", "active", "reinforcement", "has" ], [ "show", "more", "that", "than", "not", "find", "also", "results", "prior", "representations" ], [ "work", "previous", "focus", "rely", "or", "on", "all", "user", "trained", "provide" ], [ "when", "where", "more", "are", "these", "they", "algorithms", "their", "some", "new" ], [ "nodes", "convolutional", "layer", "prediction", "show", "network", "structure", "architecture", "learned", "neural" ], [ "free", "trained", "parameters", "privacy", "uncertainty", "model", "predictions", "the", "while", "inference" ], [ "as", "such", "well", "using", "processing", "the", "through", "or", "to", "known" ], [ "compared", "with", "accuracy", "results", "respect", "all", "while", "level", "multiple", "parameters" ], [ "propose", "noise", "the", "has", "method", "prediction", "proposed", "optimization", "new", "effectiveness" ], [ "training", "loss", "trained", "we", "during", "pre", "set", "federated", "train", "generalization" ], [ "presents", "work", "approach", "challenge", "paper", "process", "new", "this", "distribution", "strategies" ], [ "machine", "system", "systems", "quantum", "such", "ml", "human", "the", "techniques", "how" ], [ "these", "not", "but", "has", "the", "it", "have", "been", "recent", "however" ], [ "the", "be", "can", "to", "used", "how", "it", "they", "or", "any" ], [ "into", "between", "the", "approach", "by", "filter", "while", "however", "function", "inspired" ], [ "graph", "the", "and", "we", "node", "gnns", "graphs", "nodes", "representation", "structure" ], [ "samples", "datasets", "scale", "over", "real", "accuracy", "world", "dataset", "large", "the" ], [ "models", "fidelity", "generative", "inference", "prediction", "high", "language", "trained", "these", "modeling" ], [ "approach", "input", "an", "its", "framework", "shape", "technique", "3d", "using", "the" ], [ "results", "our", "approach", "best", "framework", "first", "we", "work", "experiments", "only" ], [ "the", "to", "algorithm", "algorithms", "clustering", "proposed", "sample", "label", "two", "linear" ], [ "language", "using", "information", "object", "from", "between", "different", "set", "representations", "the" ], [ "information", "proposed", "is", "user", "flow", "based", "the", "propose", "approaches", "better" ], [ "time", "series", "real", "at", "of", "temporal", "forecasting", "world", "each", "hierarchical" ], [ "both", "other", "existing", "methods", "problems", "many", "different", "be", "linear", "several" ], [ "it", "the", "is", "one", "where", "when", "at", "than", "its", "information" ], [ "segmentation", "images", "gan", "image", "text", "generative", "resolution", "we", "trained", "visual" ], [ "networks", "have", "weights", "physics", "convolutional", "over", "and", "cnns", "architectures", "neural" ], [ "against", "the", "perturbations", "robust", "robustness", "and", "attack", "attacks", "adversarial", "to" ], [ "approaches", "state", "we", "previous", "propose", "art", "action", "results", "the", "of" ], [ "reward", "agent", "reinforcement", "the", "policy", "policies", "rl", "agents", "and", "environment" ], [ "deep", "learning", "recognition", "accuracy", "ensemble", "results", "paper", "speech", "in", "design" ], [ "convergence", "problems", "optimization", "descent", "function", "gradient", "convex", "stochastic", "order", "non" ], [ "achieve", "cost", "performance", "different", "to", "better", "design", "their", "computing", "high" ], [ "domain", "to", "transfer", "domains", "adaptation", "target", "knowledge", "the", "source", "proposed" ], [ "information", "to", "features", "interaction", "feature", "selection", "and", "speech", "the", "attention" ], [ "accuracy", "classification", "of", "classifier", "classes", "classifiers", "high", "using", "class", "label" ], [ "language", "task", "multi", "to", "tasks", "different", "knowledge", "prediction", "each", "meta" ], [ "detection", "using", "object", "the", "in", "detect", "rate", "attacks", "system", "dataset" ], [ "linear", "samples", "given", "where", "problems", "the", "optimal", "solution", "to", "problem" ], [ "to", "which", "not", "only", "propose", "of", "two", "distribution", "local", "has" ], [ "matrix", "and", "rank", "low", "completion", "tensor", "alternating", "approximation", "measurements", "high" ] ]
6.587935
all-MiniLM-L6-v2
0.586
0.010488
0.153716
0.689687
ArXiv ML Papers
LDA
43
10
[ [ "of", "to", "the", "in", "and", "we", "is", "for", "learning", "that" ], [ "and", "the", "that", "of", "for", "in", "we", "is", "to", "on" ], [ "image", "generative", "gan", "gans", "and", "the", "generator", "of", "generation", "adversarial" ], [ "to", "in", "and", "of", "the", "is", "this", "data", "we", "learning" ], [ "the", "to", "time", "series", "forecasting", "we", "and", "that", "of", "methods" ], [ "in", "on", "is", "and", "the", "we", "for", "to", "of", "this" ], [ "to", "of", "the", "and", "is", "in", "we", "for", "that", "on" ], [ "the", "we", "of", "and", "in", "to", "learning", "on", "with", "that" ], [ "and", "we", "for", "the", "of", "is", "to", "in", "that", "with" ], [ "the", "we", "and", "to", "of", "in", "on", "is", "that", "network" ] ]
10.928315
all-MiniLM-L6-v2
0.26
-0.028659
0.262092
0.52677
ArXiv ML Papers
LDA
44
10
[ [ "of", "the", "series", "and", "time", "to", "in", "we", "for", "models" ], [ "on", "and", "is", "we", "to", "in", "the", "of", "with", "for" ], [ "to", "for", "graph", "in", "of", "we", "that", "the", "and", "on" ], [ "the", "of", "and", "algorithm", "gradient", "is", "for", "convergence", "in", "we" ], [ "we", "attacks", "of", "training", "to", "adversarial", "in", "that", "the", "and" ], [ "hardware", "kernel", "distillation", "kernels", "precision", "quantization", "student", "bit", "teacher", "and" ], [ "in", "we", "is", "the", "of", "to", "and", "for", "that", "on" ], [ "of", "the", "is", "to", "learning", "for", "in", "this", "and", "we" ], [ "of", "to", "the", "that", "in", "is", "learning", "and", "we", "this" ], [ "the", "on", "of", "we", "and", "in", "to", "for", "model", "with" ] ]
11.216451
all-MiniLM-L6-v2
0.33
-0.048647
0.253008
0.557253
ArXiv ML Papers
LDA
45
10
[ [ "we", "on", "that", "the", "and", "of", "graph", "to", "in", "is" ], [ "the", "we", "of", "to", "and", "in", "learning", "that", "this", "on" ], [ "on", "for", "to", "of", "and", "in", "the", "we", "is", "learning" ], [ "of", "in", "the", "to", "is", "and", "we", "this", "on", "for" ], [ "we", "the", "and", "in", "to", "of", "that", "for", "neural", "as" ], [ "user", "on", "the", "to", "recommendation", "of", "and", "for", "users", "we" ], [ "generator", "resolution", "gan", "generative", "gans", "and", "images", "image", "generation", "we" ], [ "of", "the", "and", "for", "is", "we", "to", "in", "on", "data" ], [ "flow", "encoder", "flows", "traffic", "translation", "decoder", "level", "speed", "normalizing", "character" ], [ "the", "and", "of", "to", "is", "in", "we", "for", "that", "with" ] ]
19.839619
all-MiniLM-L6-v2
0.38
-0.04485
0.241503
0.567928
ArXiv ML Papers
LDA
46
10
[ [ "to", "in", "and", "of", "the", "we", "is", "neural", "for", "on" ], [ "we", "adversarial", "of", "the", "and", "to", "attacks", "training", "that", "on" ], [ "the", "of", "to", "and", "in", "on", "with", "for", "feature", "we" ], [ "that", "in", "we", "the", "of", "to", "and", "for", "learning", "this" ], [ "algorithm", "power", "coupled", "quantum", "derived", "shift", "stability", "hilbert", "mode", "under" ], [ "in", "and", "the", "of", "is", "to", "this", "we", "for", "as" ], [ "in", "of", "the", "to", "and", "we", "for", "is", "that", "data" ], [ "the", "of", "we", "and", "in", "to", "is", "for", "that", "with" ], [ "and", "we", "that", "the", "of", "on", "in", "to", "our", "this" ], [ "traffic", "driving", "of", "and", "the", "demand", "is", "with", "for", "are" ] ]
10.933623
all-MiniLM-L6-v2
0.35
-0.055283
0.261051
0.565739
ArXiv ML Papers
LDA
43
20
[ [ "to", "in", "learning", "the", "of", "data", "is", "and", "communication", "model" ], [ "and", "network", "neural", "of", "to", "the", "networks", "we", "in", "that" ], [ "image", "generator", "gans", "likelihood", "generative", "gan", "adversarial", "and", "the", "images" ], [ "quantum", "traffic", "the", "to", "and", "is", "driving", "vehicle", "in", "of" ], [ "sensitivity", "of", "subspace", "construction", "the", "subspaces", "in", "is", "to", "mathematical" ], [ "to", "and", "the", "of", "in", "we", "that", "adversarial", "training", "attacks" ], [ "and", "for", "this", "we", "of", "to", "in", "is", "that", "the" ], [ "learning", "the", "methods", "data", "labels", "supervised", "semi", "and", "labeled", "label" ], [ "to", "of", "in", "the", "that", "and", "ml", "can", "we", "for" ], [ "and", "the", "we", "of", "on", "to", "is", "network", "in", "graph" ], [ "to", "in", "we", "for", "that", "the", "is", "policy", "of", "and" ], [ "of", "and", "in", "the", "is", "we", "on", "that", "for", "to" ], [ "of", "in", "and", "the", "that", "to", "we", "learning", "for", "is" ], [ "we", "to", "on", "the", "of", "and", "in", "for", "model", "with" ], [ "in", "of", "the", "and", "with", "we", "to", "for", "this", "is" ], [ "and", "classification", "of", "to", "data", "the", "classifier", "for", "in", "is" ], [ "this", "we", "the", "of", "and", "for", "in", "to", "is", "on" ], [ "trees", "stochastic", "varepsilon", "left", "right", "frac", "convergence", "rate", "optimal", "oracle" ], [ "to", "we", "that", "the", "on", "for", "learning", "of", "and", "data" ], [ "data", "tensor", "and", "the", "video", "metric", "to", "for", "of", "learning" ] ]
12.112303
all-MiniLM-L6-v2
0.315
-0.023955
0.215658
0.56059
ArXiv ML Papers
LDA
44
20
[ [ "to", "series", "time", "of", "and", "the", "in", "video", "we", "for" ], [ "the", "to", "and", "of", "in", "for", "we", "is", "on", "with" ], [ "and", "the", "to", "of", "in", "we", "data", "is", "that", "this" ], [ "search", "and", "the", "variables", "variational", "of", "autoencoders", "vae", "generative", "is" ], [ "adversarial", "to", "of", "we", "attacks", "the", "that", "attack", "training", "robustness" ], [ "kernel", "bayesian", "gaussian", "posterior", "processes", "the", "generalization", "kernels", "compositional", "and" ], [ "in", "the", "is", "we", "and", "of", "to", "that", "model", "this" ], [ "of", "the", "and", "to", "is", "in", "for", "this", "be", "machine" ], [ "in", "policy", "of", "and", "learning", "to", "the", "that", "we", "reinforcement" ], [ "and", "speech", "was", "to", "model", "the", "for", "in", "of", "signals" ], [ "of", "the", "and", "to", "for", "graph", "in", "on", "that", "we" ], [ "and", "interactive", "users", "of", "sensing", "user", "to", "for", "data", "that" ], [ "answer", "black", "query", "box", "question", "questions", "system", "rate", "answering", "based" ], [ "of", "the", "and", "we", "in", "data", "to", "learning", "is", "on" ], [ "of", "the", "we", "to", "in", "and", "that", "is", "for", "with" ], [ "xgboost", "bandits", "and", "approaches", "for", "in", "are", "using", "text", "proposal" ], [ "in", "learning", "on", "this", "for", "the", "of", "we", "to", "and" ], [ "to", "of", "the", "by", "for", "cnn", "resolution", "and", "neural", "with" ], [ "of", "we", "in", "to", "for", "and", "the", "is", "that", "algorithm" ], [ "to", "and", "we", "the", "in", "that", "of", "for", "is", "learning" ] ]
21.244746
all-MiniLM-L6-v2
0.355
-0.032252
0.226536
0.583911
ArXiv ML Papers
LDA
45
20
[ [ "on", "and", "the", "in", "to", "we", "is", "of", "that", "for" ], [ "the", "to", "of", "we", "and", "in", "learning", "that", "on", "for" ], [ "and", "to", "target", "we", "the", "of", "in", "learning", "domain", "for" ], [ "of", "the", "to", "and", "we", "in", "learning", "is", "this", "as" ], [ "of", "to", "the", "in", "and", "we", "as", "for", "learning", "that" ], [ "of", "user", "users", "the", "and", "recommendation", "to", "for", "we", "items" ], [ "generative", "to", "resolution", "images", "gan", "and", "image", "we", "gans", "generator" ], [ "is", "the", "and", "for", "in", "of", "to", "data", "we", "on" ], [ "mild", "actively", "suggested", "integrating", "poisoning", "regardless", "growth", "devise", "existence", "conclusion" ], [ "the", "of", "is", "and", "in", "we", "to", "for", "that", "with" ], [ "the", "text", "and", "of", "in", "language", "attention", "with", "recurrent", "neural" ], [ "and", "we", "to", "fairness", "that", "fair", "of", "functional", "are", "this" ], [ "of", "to", "the", "graph", "and", "we", "in", "that", "on", "graphs" ], [ "explanations", "protein", "speaker", "to", "and", "the", "of", "hardware", "feature", "that" ], [ "activation", "layer", "scaling", "network", "networks", "neural", "functions", "the", "of", "depth" ], [ "to", "is", "in", "on", "the", "and", "of", "for", "we", "training" ], [ "of", "and", "we", "the", "to", "in", "an", "object", "is", "for" ], [ "of", "to", "the", "learning", "we", "and", "supervised", "data", "that", "on" ], [ "and", "is", "of", "in", "to", "the", "for", "we", "this", "on" ], [ "the", "of", "to", "and", "training", "that", "we", "adversarial", "in", "attacks" ] ]
20.694685
all-MiniLM-L6-v2
0.335
-0.048407
0.212405
0.557579
ArXiv ML Papers
LDA
46
20
[ [ "network", "the", "to", "of", "we", "in", "neural", "networks", "and", "on" ], [ "the", "of", "and", "to", "for", "with", "on", "we", "is", "that" ], [ "on", "of", "and", "the", "with", "in", "to", "for", "we", "is" ], [ "in", "of", "and", "we", "learning", "to", "the", "that", "graphs", "for" ], [ "in", "this", "for", "quantum", "noise", "and", "is", "stability", "the", "power" ], [ "for", "is", "to", "of", "this", "the", "and", "as", "in", "we" ], [ "in", "of", "to", "we", "is", "the", "and", "for", "that", "data" ], [ "for", "to", "is", "that", "of", "we", "and", "the", "in", "with" ], [ "and", "on", "of", "to", "in", "we", "the", "that", "our", "this" ], [ "are", "driving", "and", "the", "service", "expert", "pooling", "of", "with", "experts" ], [ "diffusion", "equations", "traffic", "equation", "neural", "speed", "network", "solution", "mathematical", "the" ], [ "policy", "for", "and", "of", "gradient", "to", "the", "we", "in", "learning" ], [ "the", "and", "for", "in", "of", "to", "we", "domain", "models", "speech" ], [ "data", "learning", "trees", "metric", "distance", "the", "conditional", "from", "distances", "we" ], [ "of", "and", "to", "supervised", "in", "learning", "the", "data", "we", "with" ], [ "graph", "we", "embedding", "node", "nodes", "the", "to", "representations", "embeddings", "representation" ], [ "dramatically", "greatly", "newton", "concrete", "exhibits", "proof", "reconstructing", "decreases", "caused", "asymptotically" ], [ "concepts", "words", "word", "vectors", "to", "promising", "vector", "we", "similarity", "this" ], [ "to", "attacks", "of", "adversarial", "we", "and", "that", "the", "in", "model" ], [ "to", "of", "the", "in", "and", "for", "we", "detection", "samples", "with" ] ]
11.831478
all-MiniLM-L6-v2
0.39
-0.050428
0.208954
0.586658
ArXiv ML Papers
LDA
43
30
[ [ "to", "is", "learning", "the", "and", "of", "in", "transfer", "with", "supervised" ], [ "neural", "network", "the", "of", "networks", "to", "and", "speech", "layer", "this" ], [ "generalization", "as", "functions", "tasks", "synthesis", "log", "which", "negative", "loss", "generalized" ], [ "and", "to", "of", "data", "traffic", "learning", "policy", "reinforcement", "the", "stream" ], [ "of", "robust", "pooling", "consumption", "in", "algorithms", "case", "the", "energy", "and" ], [ "the", "of", "attacks", "in", "training", "to", "we", "adversarial", "and", "that" ], [ "we", "to", "the", "in", "of", "that", "and", "is", "this", "for" ], [ "face", "in", "treatment", "data", "the", "of", "for", "estimator", "causal", "bias" ], [ "the", "to", "of", "in", "and", "ml", "that", "can", "we", "be" ], [ "and", "to", "the", "of", "we", "images", "image", "with", "on", "in" ], [ "the", "of", "quantum", "in", "and", "as", "for", "to", "we", "kernel" ], [ "of", "the", "and", "in", "for", "to", "is", "we", "that", "this" ], [ "of", "the", "to", "in", "this", "and", "we", "for", "learning", "that" ], [ "in", "the", "we", "to", "and", "learning", "for", "of", "task", "on" ], [ "of", "and", "method", "to", "for", "is", "in", "we", "learning", "the" ], [ "of", "feature", "for", "and", "active", "the", "that", "fairness", "we", "in" ], [ "we", "of", "and", "for", "in", "to", "the", "on", "is", "this" ], [ "candidate", "edge", "frac", "partitioning", "nearest", "partition", "right", "neighbor", "neighbors", "left" ], [ "that", "the", "we", "to", "and", "of", "queries", "on", "clustering", "learning" ], [ "supervised", "the", "detection", "data", "to", "and", "for", "from", "we", "of" ], [ "of", "to", "the", "and", "is", "we", "in", "on", "with", "that" ], [ "source", "domain", "the", "target", "audio", "video", "cross", "adaptation", "net", "videos" ], [ "our", "questions", "we", "shift", "question", "answer", "framework", "and", "consistency", "distributional" ], [ "communication", "data", "of", "the", "privacy", "users", "federated", "user", "and", "to" ], [ "the", "of", "we", "and", "in", "for", "with", "to", "that", "is" ], [ "of", "to", "distillation", "teacher", "and", "in", "the", "student", "knowledge", "network" ], [ "and", "to", "that", "of", "for", "the", "in", "networks", "we", "on" ], [ "generation", "text", "image", "generative", "to", "sequence", "generate", "models", "the", "and" ], [ "music", "recommendation", "expert", "contextual", "context", "knowledge", "regret", "scenarios", "real", "bandit" ], [ "of", "the", "graphs", "to", "and", "graph", "we", "node", "for", "nodes" ] ]
13.241964
all-MiniLM-L6-v2
0.4
-0.039303
0.200602
0.60234
ArXiv ML Papers
LDA
44
30
[ [ "in", "we", "deep", "design", "to", "the", "and", "of", "models", "translation" ], [ "in", "and", "the", "we", "of", "is", "on", "to", "for", "network" ], [ "to", "of", "and", "in", "graph", "we", "the", "that", "on", "this" ], [ "and", "devices", "the", "communication", "federated", "model", "channel", "proposed", "distributed", "to" ], [ "the", "attacks", "adversarial", "attack", "robustness", "training", "of", "to", "quantum", "against" ], [ "kernels", "are", "recursive", "functions", "gaussian", "activation", "width", "compositional", "mixed", "kernel" ], [ "in", "of", "to", "we", "the", "and", "that", "this", "on", "is" ], [ "in", "the", "of", "and", "to", "for", "can", "this", "systems", "be" ], [ "to", "and", "agent", "tasks", "of", "in", "that", "the", "learning", "we" ], [ "to", "net", "of", "and", "the", "in", "this", "hardware", "for", "demand" ], [ "is", "that", "and", "the", "in", "to", "of", "we", "on", "network" ], [ "sensing", "of", "free", "user", "motion", "for", "text", "and", "study", "head" ], [ "tree", "on", "the", "learning", "this", "meaningful", "trees", "valued", "is", "averaged" ], [ "of", "we", "to", "data", "and", "is", "learning", "supervised", "the", "on" ], [ "of", "to", "the", "and", "in", "that", "we", "for", "on", "with" ], [ "privacy", "private", "xgboost", "differentially", "data", "learning", "and", "for", "model", "inference" ], [ "in", "for", "and", "to", "of", "the", "we", "data", "learning", "this" ], [ "to", "resolution", "for", "and", "images", "scale", "of", "the", "by", "games" ], [ "the", "of", "we", "to", "for", "and", "that", "in", "algorithm", "is" ], [ "of", "and", "in", "learning", "we", "that", "the", "to", "for", "is" ], [ "to", "the", "of", "and", "data", "for", "is", "that", "in", "we" ], [ "to", "in", "metric", "series", "of", "time", "the", "on", "is", "similarity" ], [ "of", "to", "in", "the", "and", "model", "we", "learning", "policy", "speech" ], [ "optimization", "method", "the", "black", "and", "box", "we", "gradient", "problems", "gans" ], [ "bit", "massive", "diffusion", "the", "constrained", "we", "in", "to", "and", "that" ], [ "the", "in", "of", "and", "to", "this", "is", "we", "for", "with" ], [ "sampling", "carlo", "dl", "monte", "relevance", "markov", "multi", "label", "mcmc", "chain" ], [ "to", "data", "region", "detection", "proposal", "boxes", "object", "interpolation", "mapping", "points" ], [ "imbalance", "arises", "sampled", "al", "actively", "detail", "idea", "corpus", "imbalanced", "overlapping" ], [ "search", "and", "to", "the", "of", "we", "forecasting", "for", "neural", "power" ] ]
13.062546
all-MiniLM-L6-v2
0.416667
-0.046981
0.187724
0.604609
ArXiv ML Papers
LDA
45
30
[ [ "in", "to", "model", "of", "the", "we", "and", "on", "is", "data" ], [ "to", "of", "learning", "that", "we", "the", "and", "policy", "in", "reinforcement" ], [ "in", "domain", "and", "transfer", "to", "source", "target", "the", "of", "for" ], [ "is", "the", "to", "of", "method", "this", "and", "in", "fairness", "from" ], [ "time", "for", "in", "of", "and", "series", "the", "to", "forecasting", "term" ], [ "user", "of", "users", "and", "for", "model", "items", "the", "algorithms", "to" ], [ "personalized", "perspective", "performed", "performing", "performs", "period", "periodic", "periods", "person", "personal" ], [ "of", "and", "the", "to", "in", "for", "is", "on", "data", "with" ], [ "personalized", "perspective", "performed", "performing", "performs", "period", "periodic", "periods", "person", "personal" ], [ "the", "for", "algorithm", "is", "of", "we", "in", "to", "and", "that" ], [ "the", "language", "recurrent", "translation", "rnn", "and", "label", "languages", "multi", "to" ], [ "nearest", "social", "media", "neighbor", "diffusion", "classifier", "communities", "topological", "information", "in" ], [ "the", "of", "graph", "clustering", "and", "we", "nodes", "to", "node", "for" ], [ "speaker", "to", "end", "hardware", "differentiable", "accelerators", "forgetting", "and", "neural", "catastrophic" ], [ "structured", "belief", "partition", "mean", "approximations", "field", "counterpart", "positive", "trees", "naive" ], [ "the", "of", "and", "in", "we", "training", "to", "is", "on", "with" ], [ "the", "channel", "to", "in", "object", "and", "is", "detection", "of", "attacks" ], [ "the", "learning", "data", "to", "latent", "supervised", "of", "we", "metric", "distance" ], [ "the", "of", "and", "to", "is", "for", "in", "queries", "we", "ensemble" ], [ "in", "to", "of", "the", "with", "we", "and", "is", "for", "neural" ], [ "the", "code", "and", "to", "training", "github", "https", "pre", "com", "networks" ], [ "of", "sequential", "methods", "for", "solving", "coupled", "stochastic", "framework", "games", "to" ], [ "networks", "random", "of", "that", "layer", "and", "error", "the", "we", "generalization" ], [ "learning", "in", "data", "deep", "and", "of", "to", "the", "we", "for" ], [ "of", "in", "the", "local", "points", "point", "and", "kernels", "kernel", "that" ], [ "as", "and", "of", "flow", "the", "to", "learning", "in", "we", "tree" ], [ "the", "and", "of", "to", "graph", "in", "for", "tasks", "we", "on" ], [ "the", "of", "to", "and", "in", "we", "on", "for", "that", "this" ], [ "and", "image", "generative", "the", "images", "we", "to", "text", "gan", "of" ], [ "changing", "processes", "inefficient", "on", "conditions", "process", "rapidly", "act", "value", "wide" ] ]
12.018622
all-MiniLM-L6-v2
0.443333
-0.071942
0.169726
0.634942
ArXiv ML Papers
LDA
46
30
[ [ "the", "and", "to", "of", "in", "neural", "we", "network", "on", "networks" ], [ "is", "for", "the", "in", "and", "we", "our", "to", "of", "with" ], [ "the", "and", "in", "of", "on", "to", "data", "image", "we", "is" ], [ "learning", "of", "to", "in", "the", "and", "we", "that", "for", "agents" ], [ "personalized", "perspective", "performed", "performing", "performs", "period", "periodic", "periods", "person", "personal" ], [ "the", "of", "and", "in", "to", "is", "for", "flow", "we", "on" ], [ "of", "the", "and", "to", "in", "for", "is", "we", "that", "data" ], [ "of", "in", "and", "the", "we", "is", "to", "for", "that", "algorithm" ], [ "to", "the", "image", "of", "we", "on", "images", "and", "that", "in" ], [ "the", "and", "segmentation", "semantic", "traffic", "driving", "of", "in", "expert", "pooling" ], [ "personalized", "perspective", "performed", "performing", "performs", "period", "periodic", "periods", "person", "personal" ], [ "of", "the", "and", "gradient", "we", "for", "to", "policy", "optimization", "in" ], [ "we", "of", "to", "the", "and", "models", "that", "for", "our", "object" ], [ "label", "multi", "tensor", "the", "labels", "query", "and", "relevance", "algorithm", "given" ], [ "the", "of", "model", "to", "methods", "and", "ensemble", "generative", "data", "we" ], [ "nodes", "graph", "node", "graphs", "the", "embedding", "to", "of", "network", "structure" ], [ "personalized", "perspective", "performed", "performing", "performs", "period", "periodic", "periods", "person", "personal" ], [ "personalized", "perspective", "performed", "performing", "performs", "period", "periodic", "periods", "person", "personal" ], [ "of", "to", "the", "that", "we", "attacks", "adversarial", "training", "in", "attack" ], [ "of", "for", "in", "energy", "to", "the", "we", "monte", "and", "carlo" ], [ "fusion", "video", "feature", "frame", "frames", "the", "of", "this", "in", "we" ], [ "the", "of", "to", "and", "that", "we", "on", "in", "for", "is" ], [ "of", "the", "and", "this", "for", "to", "in", "is", "we", "that" ], [ "language", "representations", "words", "in", "of", "speech", "the", "languages", "we", "embeddings" ], [ "shift", "clustering", "and", "distributional", "algorithm", "consistency", "mode", "we", "consistent", "under" ], [ "of", "and", "the", "we", "inference", "data", "to", "variational", "latent", "bayesian" ], [ "is", "in", "to", "we", "learning", "the", "and", "domain", "of", "target" ], [ "feature", "classifier", "metric", "features", "classification", "forecasting", "the", "filter", "classifiers", "of" ], [ "the", "to", "that", "of", "robot", "we", "in", "and", "for", "privacy" ], [ "personalized", "perspective", "performed", "performing", "performs", "period", "periodic", "periods", "person", "personal" ] ]
12.875089
all-MiniLM-L6-v2
0.336667
-0.098042
0.191397
0.648427
ArXiv ML Papers
LDA
43
40
[ [ "the", "for", "is", "in", "of", "distributed", "and", "to", "communication", "flow" ], [ "to", "the", "speech", "of", "signals", "and", "text", "in", "we", "word" ], [ "loss", "open", "source", "goals", "and", "gans", "an", "as", "library", "implementations" ], [ "traffic", "quantum", "in", "and", "noise", "the", "of", "to", "is", "data" ], [ "reference", "vs", "2019", "the", "best", "of", "on", "in", "and", "be" ], [ "to", "adversarial", "in", "and", "of", "the", "attacks", "we", "training", "models" ], [ "for", "the", "we", "is", "in", "that", "and", "of", "to", "this" ], [ "system", "systems", "energy", "control", "and", "physics", "in", "the", "of", "to" ], [ "to", "the", "of", "and", "that", "in", "ml", "is", "we", "reward" ], [ "we", "the", "and", "images", "image", "to", "segmentation", "of", "in", "with" ], [ "driving", "and", "of", "dnn", "in", "the", "to", "for", "on", "as" ], [ "and", "clustering", "in", "of", "the", "based", "for", "to", "is", "we" ], [ "in", "to", "the", "of", "and", "we", "for", "neural", "on", "networks" ], [ "the", "to", "of", "and", "learning", "we", "in", "tasks", "that", "task" ], [ "is", "for", "and", "in", "of", "to", "we", "the", "this", "method" ], [ "of", "for", "short", "protein", "term", "the", "and", "in", "lstm", "is" ], [ "and", "in", "of", "to", "we", "for", "the", "is", "on", "this" ], [ "candidate", "diversity", "candidates", "retrieval", "user", "mixture", "set", "recall", "clusters", "generation" ], [ "the", "to", "that", "and", "private", "privacy", "of", "clustering", "in", "we" ], [ "is", "and", "we", "to", "of", "the", "data", "causal", "in", "local" ], [ "and", "is", "to", "on", "in", "the", "of", "we", "with", "for" ], [ "the", "adaptation", "source", "domain", "video", "target", "cross", "fine", "grained", "videos" ], [ "personalized", "perspective", "performed", "performing", "performs", "period", "periodic", "periods", "person", "personal" ], [ "mixed", "in", "the", "and", "languages", "of", "to", "compositional", "for", "kernels" ], [ "in", "the", "of", "is", "to", "for", "algorithm", "and", "we", "that" ], [ "and", "the", "to", "of", "left", "sparse", "right", "we", "in", "frac" ], [ "to", "in", "with", "and", "that", "of", "performance", "we", "for", "the" ], [ "image", "generation", "the", "and", "generate", "of", "to", "forecasting", "for", "text" ], [ "activation", "relu", "functions", "layers", "neurons", "neuron", "complex", "function", "sign", "unit" ], [ "graph", "the", "and", "to", "for", "graphs", "of", "we", "embedding", "that" ], [ "and", "to", "in", "the", "of", "we", "that", "as", "is", "demand" ], [ "bounds", "the", "we", "of", "to", "and", "bayesian", "with", "for", "optimization" ], [ "and", "for", "we", "to", "in", "classifier", "of", "the", "learning", "label" ], [ "the", "in", "of", "to", "net", "and", "geometric", "graph", "convolutional", "style" ], [ "the", "to", "for", "data", "we", "of", "and", "this", "in", "that" ], [ "the", "to", "we", "of", "and", "on", "for", "models", "data", "in" ], [ "of", "the", "to", "and", "in", "network", "we", "networks", "object", "on" ], [ "gan", "graph", "the", "generative", "gnns", "node", "and", "is", "of", "to" ], [ "to", "and", "in", "is", "we", "of", "that", "the", "on", "learning" ], [ "and", "of", "policy", "we", "to", "the", "in", "learning", "for", "that" ] ]
12.933721
all-MiniLM-L6-v2
0.3475
-0.043755
0.195528
0.61362
ArXiv ML Papers
LDA
44
40
[ [ "of", "to", "the", "we", "manifold", "in", "for", "neural", "on", "and" ], [ "the", "and", "to", "with", "neural", "in", "of", "on", "we", "for" ], [ "data", "we", "on", "of", "the", "that", "and", "this", "in", "to" ], [ "the", "dynamics", "and", "to", "of", "that", "model", "data", "vae", "groups" ], [ "attack", "attacks", "adversarial", "to", "the", "we", "of", "training", "robustness", "against" ], [ "personalized", "perspective", "performed", "performing", "performs", "period", "periodic", "periods", "person", "personal" ], [ "to", "the", "of", "we", "that", "in", "and", "is", "this", "on" ], [ "be", "for", "in", "the", "and", "to", "of", "can", "is", "this" ], [ "to", "learning", "the", "of", "in", "that", "policy", "and", "reinforcement", "we" ], [ "of", "for", "was", "we", "to", "and", "the", "in", "long", "term" ], [ "and", "in", "the", "of", "to", "for", "we", "that", "on", "learning" ], [ "word", "of", "head", "smart", "motion", "text", "natural", "selection", "for", "accuracy" ], [ "dynamical", "processes", "system", "library", "online", "systems", "the", "production", "an", "to" ], [ "the", "to", "of", "and", "learning", "data", "in", "is", "we", "methods" ], [ "for", "and", "the", "on", "of", "that", "we", "to", "in", "is" ], [ "for", "deep", "models", "in", "net", "are", "diffusion", "and", "paper", "of" ], [ "and", "of", "the", "to", "in", "for", "we", "learning", "this", "on" ], [ "is", "the", "and", "cnn", "to", "dnn", "resolution", "by", "for", "of" ], [ "of", "we", "to", "that", "for", "the", "gradient", "in", "and", "algorithm" ], [ "we", "the", "in", "and", "of", "that", "for", "learning", "to", "is" ], [ "for", "we", "and", "to", "in", "of", "the", "is", "data", "that" ], [ "time", "series", "and", "of", "in", "the", "for", "to", "we", "is" ], [ "speech", "the", "to", "on", "in", "of", "based", "and", "we", "model" ], [ "epsilon", "method", "we", "optimization", "black", "box", "and", "gradient", "query", "propose" ], [ "the", "feature", "classifier", "class", "classification", "classes", "data", "and", "we", "of" ], [ "and", "for", "we", "of", "to", "is", "the", "with", "in", "on" ], [ "carlo", "recommendation", "monte", "the", "and", "chain", "of", "item", "mcmc", "parallel" ], [ "self", "contrastive", "clustering", "supervised", "representations", "in", "transformations", "data", "to", "augmentation" ], [ "imbalanced", "for", "catastrophic", "to", "forgetting", "the", "event", "imbalance", "we", "that" ], [ "to", "architecture", "and", "of", "search", "channel", "the", "with", "is", "for" ], [ "the", "in", "and", "of", "to", "is", "we", "for", "that", "problem" ], [ "to", "tensors", "and", "in", "tensor", "on", "rank", "networks", "multiple", "of" ], [ "of", "we", "data", "the", "to", "in", "and", "for", "that", "as" ], [ "of", "type", "and", "types", "task", "any", "associated", "answer", "to", "the" ], [ "forecasting", "of", "and", "forecast", "the", "probabilistic", "we", "demand", "to", "that" ], [ "on", "data", "to", "systems", "for", "is", "of", "and", "the", "this" ], [ "the", "of", "graph", "graphs", "nodes", "node", "and", "to", "on", "networks" ], [ "to", "in", "and", "we", "of", "the", "are", "for", "is", "that" ], [ "of", "the", "and", "to", "approximately", "sound", "in", "reliable", "on", "is" ], [ "bias", "of", "in", "to", "the", "and", "nn", "that", "error", "is" ] ]
13.056187
all-MiniLM-L6-v2
0.3425
-0.037738
0.207822
0.592155
ArXiv ML Papers
LDA
45
40
[ [ "rank", "the", "and", "matrix", "of", "tensor", "we", "is", "for", "in" ], [ "with", "of", "learning", "in", "we", "rl", "to", "the", "that", "and" ], [ "domain", "the", "of", "and", "to", "target", "for", "source", "label", "in" ], [ "quantum", "to", "learning", "is", "and", "the", "of", "this", "in", "for" ], [ "for", "of", "and", "the", "time", "forecasting", "series", "as", "in", "ensemble" ], [ "the", "items", "recommendation", "to", "users", "user", "of", "based", "on", "recommender" ], [ "profiles", "scale", "details", "variable", "spatial", "resolution", "patches", "images", "at", "conditioning" ], [ "the", "of", "and", "to", "data", "in", "is", "for", "on", "with" ], [ "personalized", "perspective", "performed", "performing", "performs", "period", "periodic", "periods", "person", "personal" ], [ "is", "for", "data", "the", "and", "to", "distributed", "with", "in", "of" ], [ "time", "recurrent", "series", "dynamics", "the", "rnn", "for", "of", "temporal", "measurement" ], [ "net", "diffusion", "recovery", "communities", "community", "information", "in", "to", "this", "of" ], [ "the", "nodes", "node", "clustering", "to", "of", "graph", "graphs", "and", "we" ], [ "search", "vae", "mobile", "architecture", "end", "hardware", "to", "and", "architectures", "neural" ], [ "implementations", "extensions", "updated", "millions", "base", "minimal", "entities", "down", "transforms", "having" ], [ "adversarial", "training", "to", "in", "the", "attacks", "of", "and", "we", "is" ], [ "to", "and", "the", "of", "is", "channel", "an", "semantic", "3d", "for" ], [ "semi", "learning", "of", "supervised", "the", "we", "and", "random", "features", "ct" ], [ "to", "of", "the", "and", "in", "for", "is", "this", "sharing", "eeg" ], [ "to", "the", "and", "of", "we", "for", "in", "on", "with", "that" ], [ "neural", "to", "and", "network", "the", "networks", "energy", "convolutional", "of", "training" ], [ "on", "protein", "generative", "sequence", "sequences", "to", "learning", "methods", "of", "knowledge" ], [ "learning", "meta", "reinforcement", "games", "the", "policy", "and", "that", "agent", "action" ], [ "and", "learning", "the", "we", "to", "in", "of", "for", "tasks", "methods" ], [ "10", "confidence", "100", "cifar", "to", "unlabeled", "that", "on", "we", "only" ], [ "of", "fairness", "we", "learning", "and", "to", "the", "in", "distance", "metric" ], [ "for", "text", "in", "we", "on", "of", "language", "and", "the", "to" ], [ "of", "in", "we", "for", "that", "the", "this", "to", "and", "on" ], [ "the", "object", "objects", "and", "uncertainty", "estimation", "driving", "to", "of", "we" ], [ "personalized", "perspective", "performed", "performing", "performs", "period", "periodic", "periods", "person", "personal" ], [ "precision", "on", "of", "and", "that", "the", "in", "to", "models", "learner" ], [ "with", "to", "and", "of", "in", "the", "is", "for", "we", "this" ], [ "in", "the", "to", "of", "and", "we", "is", "that", "for", "this" ], [ "bandit", "the", "and", "to", "agents", "agent", "of", "by", "in", "is" ], [ "and", "we", "the", "of", "for", "in", "to", "is", "with", "that" ], [ "super", "global", "minimax", "to", "points", "point", "local", "model", "score", "the" ], [ "to", "contrastive", "our", "image", "the", "method", "data", "fine", "tasks", "imagenet" ], [ "of", "in", "to", "the", "and", "is", "that", "we", "model", "by" ], [ "learning", "of", "on", "model", "to", "we", "the", "training", "and", "in" ], [ "memory", "the", "of", "term", "and", "network", "to", "flow", "in", "attention" ] ]
13.227212
all-MiniLM-L6-v2
0.4
-0.058412
0.186957
0.626419
ArXiv ML Papers
LDA
46
40
[ [ "in", "the", "networks", "network", "of", "neural", "we", "and", "to", "that" ], [ "fairness", "that", "in", "and", "to", "is", "we", "the", "for", "of" ], [ "the", "in", "on", "we", "and", "of", "to", "graph", "our", "with" ], [ "learning", "of", "the", "and", "to", "in", "we", "that", "communication", "agents" ], [ "svm", "tensor", "machine", "vector", "overfitting", "the", "support", "rank", "an", "one" ], [ "to", "learning", "in", "the", "policy", "and", "of", "reinforcement", "we", "is" ], [ "of", "and", "the", "to", "in", "is", "we", "for", "that", "on" ], [ "the", "in", "of", "we", "is", "to", "and", "that", "with", "for" ], [ "images", "on", "of", "to", "the", "and", "we", "that", "in", "with" ], [ "the", "points", "local", "traffic", "with", "driving", "to", "sharing", "and", "of" ], [ "is", "pooling", "25", "rnn", "expression", "the", "in", "recurrent", "average", "of" ], [ "the", "and", "of", "we", "gradient", "optimization", "for", "to", "stochastic", "convergence" ], [ "to", "objects", "and", "of", "the", "for", "that", "we", "signals", "sequences" ], [ "text", "vae", "the", "trees", "generation", "for", "and", "of", "joint", "that" ], [ "the", "and", "data", "model", "gan", "we", "of", "generative", "methods", "in" ], [ "embeddings", "3d", "representations", "segmentation", "the", "shape", "for", "learn", "to", "pipeline" ], [ "personalized", "perspective", "performed", "performing", "performs", "period", "periodic", "periods", "person", "personal" ], [ "personalized", "perspective", "performed", "performing", "performs", "period", "periodic", "periods", "person", "personal" ], [ "to", "attacks", "adversarial", "the", "of", "that", "and", "training", "attack", "we" ], [ "and", "the", "to", "in", "we", "carlo", "of", "monte", "that", "samples" ], [ "video", "feature", "features", "the", "frame", "in", "fusion", "of", "module", "this" ], [ "of", "to", "the", "and", "we", "on", "in", "that", "for", "our" ], [ "and", "of", "to", "for", "the", "in", "this", "we", "that", "is" ], [ "processes", "process", "conditional", "density", "of", "we", "gaussian", "the", "on", "in" ], [ "languages", "probabilistic", "inference", "algorithm", "improve", "impossible", "substantial", "program", "make", "programming" ], [ "of", "to", "the", "metric", "data", "in", "learning", "we", "ensemble", "with" ], [ "the", "and", "learning", "to", "in", "we", "of", "domain", "on", "is" ], [ "extracted", "extraction", "co", "classifier", "regions", "filter", "image", "filters", "feature", "by" ], [ "to", "the", "deep", "in", "learning", "we", "of", "for", "and", "speech" ], [ "personalized", "perspective", "performed", "performing", "performs", "period", "periodic", "periods", "person", "personal" ], [ "word", "embedding", "similarity", "concepts", "between", "words", "vectors", "vector", "measure", "to" ], [ "personalized", "perspective", "performed", "performing", "performs", "period", "periodic", "periods", "person", "personal" ], [ "the", "latent", "to", "manifold", "variational", "of", "dimensional", "and", "we", "variables" ], [ "for", "of", "and", "to", "time", "we", "the", "that", "in", "series" ], [ "to", "the", "of", "in", "and", "image", "is", "model", "method", "this" ], [ "the", "in", "of", "and", "face", "to", "flow", "we", "for", "on" ], [ "with", "and", "of", "on", "in", "the", "for", "to", "classification", "is" ], [ "to", "the", "of", "is", "in", "and", "for", "by", "are", "this" ], [ "term", "temporal", "short", "prediction", "long", "spatial", "dynamics", "and", "view", "demand" ], [ "the", "is", "of", "net", "quantum", "and", "for", "to", "estimator", "treatment" ] ]
13.765348
all-MiniLM-L6-v2
0.3825
-0.072976
0.192218
0.644522
ArXiv ML Papers
LDA
43
50
[ [ "is", "of", "for", "distributed", "the", "communication", "to", "and", "data", "as" ], [ "cifar", "transformation", "the", "10", "transformations", "relu", "activation", "neurons", "100", "of" ], [ "exploration", "bandit", "strategies", "depth", "regret", "scaling", "combinatorial", "and", "strategy", "in" ], [ "quantum", "classification", "traffic", "the", "data", "in", "to", "of", "and", "is" ], [ "personalized", "perspective", "performed", "performing", "performs", "period", "periodic", "periods", "person", "personal" ], [ "adversarial", "we", "the", "in", "of", "to", "and", "attacks", "that", "training" ], [ "of", "we", "the", "to", "and", "in", "that", "is", "for", "this" ], [ "semi", "unsupervised", "supervised", "learning", "the", "self", "data", "and", "in", "labeled" ], [ "of", "ml", "and", "that", "the", "to", "in", "can", "be", "is" ], [ "the", "and", "to", "of", "with", "is", "network", "training", "neural", "architecture" ], [ "to", "tensor", "of", "agent", "agents", "the", "we", "online", "and", "regret" ], [ "to", "in", "the", "is", "of", "and", "for", "based", "are", "algorithms" ], [ "the", "of", "to", "and", "in", "we", "learning", "that", "for", "tasks" ], [ "and", "of", "gradient", "the", "we", "in", "to", "for", "optimization", "on" ], [ "is", "to", "for", "and", "the", "method", "in", "of", "model", "an" ], [ "and", "to", "for", "in", "federated", "the", "of", "is", "that", "on" ], [ "of", "the", "and", "for", "to", "is", "this", "we", "in", "on" ], [ "items", "recommendation", "user", "right", "left", "frac", "recommender", "item", "systems", "varepsilon" ], [ "fairness", "in", "we", "and", "to", "the", "bias", "that", "of", "on" ], [ "and", "privacy", "private", "learning", "data", "to", "the", "sensitive", "of", "we" ], [ "the", "of", "and", "to", "is", "in", "we", "on", "this", "that" ], [ "moving", "selection", "head", "of", "smart", "motion", "sensing", "for", "and", "free" ], [ "questions", "question", "and", "our", "that", "answer", "given", "materials", "framework", "science" ], [ "kernels", "imitation", "of", "wireless", "in", "mixed", "imputation", "compositional", "conditions", "kernel" ], [ "for", "we", "and", "the", "of", "in", "to", "with", "algorithm", "is" ], [ "the", "to", "event", "and", "of", "in", "cnn", "events", "location", "based" ], [ "of", "the", "to", "for", "that", "and", "in", "with", "performance", "models" ], [ "and", "the", "text", "of", "language", "in", "for", "natural", "translation", "generation" ], [ "personalized", "perspective", "performed", "performing", "performs", "period", "periodic", "periods", "person", "personal" ], [ "for", "to", "search", "of", "we", "from", "sparse", "the", "that", "and" ], [ "student", "in", "the", "and", "teacher", "based", "flow", "to", "we", "sequences" ], [ "bayesian", "of", "the", "we", "processes", "and", "posterior", "model", "to", "with" ], [ "in", "and", "we", "of", "for", "the", "classifier", "classification", "label", "that" ], [ "to", "for", "of", "resolution", "graphs", "in", "the", "video", "graph", "and" ], [ "we", "data", "that", "in", "the", "for", "of", "this", "to", "and" ], [ "the", "and", "we", "of", "for", "models", "to", "in", "on", "inference" ], [ "the", "and", "object", "in", "for", "by", "we", "to", "of", "objects" ], [ "recommendations", "fair", "and", "of", "group", "algorithms", "our", "groups", "the", "to" ], [ "of", "that", "to", "in", "and", "the", "we", "is", "learning", "on" ], [ "of", "to", "and", "distance", "in", "the", "for", "we", "is", "this" ], [ "we", "to", "and", "in", "of", "the", "is", "data", "for", "that" ], [ "document", "word", "topic", "semantic", "representations", "level", "words", "documents", "sentences", "topics" ], [ "the", "and", "of", "in", "to", "we", "machine", "speech", "channel", "on" ], [ "energy", "systems", "of", "to", "sharing", "demand", "devices", "the", "and", "heterogeneous" ], [ "policies", "rl", "reinforcement", "epsilon", "in", "learning", "the", "of", "policy", "optimal" ], [ "deep", "on", "we", "and", "of", "to", "the", "in", "learning", "with" ], [ "of", "learning", "that", "the", "to", "training", "and", "in", "from", "we" ], [ "and", "for", "to", "of", "the", "in", "on", "is", "model", "performance" ], [ "surrogate", "models", "fidelity", "off", "trade", "to", "high", "and", "programs", "sampling" ], [ "and", "the", "graph", "of", "signals", "on", "speech", "in", "to", "network" ] ]
23.874444
all-MiniLM-L6-v2
0.364
-0.051004
0.189744
0.61902
ArXiv ML Papers
LDA
44
50
[ [ "image", "generative", "gan", "the", "to", "of", "generator", "images", "gans", "and" ], [ "of", "and", "in", "the", "to", "we", "for", "domain", "model", "on" ], [ "of", "in", "to", "and", "that", "the", "we", "this", "are", "on" ], [ "search", "the", "architecture", "and", "of", "shift", "to", "nas", "parameter", "ranking" ], [ "adversarial", "attacks", "training", "the", "robustness", "attack", "to", "we", "that", "robust" ], [ "processes", "kernels", "mixed", "gaussian", "are", "compositional", "recursive", "gps", "stochastic", "the" ], [ "to", "the", "and", "of", "we", "is", "in", "that", "model", "on" ], [ "to", "of", "and", "the", "can", "for", "in", "machine", "be", "we" ], [ "to", "learning", "the", "and", "of", "in", "we", "reinforcement", "agent", "that" ], [ "for", "term", "time", "net", "we", "the", "and", "in", "to", "of" ], [ "the", "of", "on", "in", "and", "we", "to", "metric", "learning", "that" ], [ "of", "people", "and", "for", "sensing", "user", "text", "users", "head", "study" ], [ "systems", "system", "to", "learning", "continuous", "production", "rules", "asynchronous", "logic", "of" ], [ "data", "the", "we", "to", "of", "learning", "and", "methods", "in", "is" ], [ "and", "to", "in", "of", "we", "the", "that", "on", "for", "is" ], [ "xgboost", "in", "the", "dense", "using", "agents", "for", "single", "different", "of" ], [ "for", "to", "the", "and", "learning", "in", "of", "we", "machine", "this" ], [ "for", "with", "layer", "the", "of", "to", "cnn", "and", "based", "in" ], [ "the", "to", "and", "of", "we", "for", "in", "algorithm", "is", "convergence" ], [ "and", "we", "of", "the", "that", "learning", "in", "to", "for", "is" ], [ "that", "the", "for", "in", "and", "is", "of", "data", "to", "we" ], [ "ensemble", "to", "and", "of", "on", "the", "in", "memory", "is", "we" ], [ "the", "of", "and", "in", "model", "policy", "we", "to", "speech", "for" ], [ "optimization", "gradient", "method", "and", "we", "the", "stochastic", "epsilon", "order", "of" ], [ "in", "selection", "and", "features", "classification", "feature", "the", "dataset", "we", "data" ], [ "with", "this", "in", "we", "of", "the", "and", "for", "to", "is" ], [ "markov", "carlo", "monte", "chain", "mcmc", "sampling", "distribution", "the", "parallel", "we" ], [ "contrastive", "self", "augmentation", "distributed", "data", "across", "10", "supervised", "locations", "to" ], [ "runtime", "generalize", "al", "2018", "strategy", "samples", "nearly", "runs", "et", "action" ], [ "network", "of", "to", "the", "we", "for", "convolutional", "layers", "and", "with" ], [ "in", "we", "is", "for", "of", "to", "that", "and", "the", "algorithm" ], [ "low", "tensor", "in", "rank", "on", "pruning", "to", "and", "tensors", "of" ], [ "and", "in", "of", "data", "the", "we", "to", "that", "on", "training" ], [ "personalized", "perspective", "performed", "performing", "performs", "period", "periodic", "periods", "person", "personal" ], [ "of", "the", "for", "sensitivity", "hardware", "that", "to", "neural", "network", "in" ], [ "and", "of", "data", "the", "to", "system", "in", "is", "systems", "for" ], [ "node", "of", "and", "on", "graph", "graphs", "nodes", "the", "we", "to" ], [ "for", "is", "in", "to", "that", "of", "and", "the", "we", "as" ], [ "confidence", "delta", "the", "and", "to", "costs", "lower", "left", "of", "right" ], [ "and", "the", "to", "sequence", "of", "mean", "is", "are", "in", "for" ], [ "and", "of", "the", "map", "to", "profiles", "network", "interpretation", "forgetting", "generalized" ], [ "that", "in", "is", "the", "to", "of", "and", "deep", "network", "segmentation" ], [ "games", "of", "classical", "that", "game", "learning", "cloud", "we", "for", "the" ], [ "the", "of", "to", "and", "in", "detection", "on", "is", "for", "we" ], [ "in", "series", "data", "to", "for", "signals", "time", "of", "the", "and" ], [ "data", "the", "for", "and", "is", "energy", "sensor", "of", "dl", "in" ], [ "to", "network", "neural", "networks", "the", "of", "in", "and", "that", "we" ], [ "quantum", "the", "of", "big", "and", "data", "to", "for", "in", "learning" ], [ "model", "and", "bert", "the", "event", "we", "to", "of", "based", "in" ], [ "in", "and", "that", "we", "the", "noise", "flow", "to", "dimensional", "high" ] ]
14.476796
all-MiniLM-L6-v2
0.356
-0.035463
0.186418
0.596312
ArXiv ML Papers
LDA
45
50
[ [ "and", "of", "the", "matrix", "is", "we", "that", "for", "in", "on" ], [ "the", "to", "learning", "in", "and", "policy", "we", "reinforcement", "that", "of" ], [ "classification", "in", "the", "is", "learning", "and", "classifier", "of", "label", "domain" ], [ "and", "the", "spatial", "sharing", "to", "grained", "demand", "fine", "reconstruction", "of" ], [ "of", "the", "segmentation", "and", "to", "for", "neural", "as", "image", "in" ], [ "inference", "based", "model", "speech", "recognition", "free", "performance", "ratio", "algorithm", "using" ], [ "personalized", "perspective", "performed", "performing", "performs", "period", "periodic", "periods", "person", "personal" ], [ "data", "with", "for", "of", "the", "in", "and", "to", "is", "we" ], [ "personalized", "perspective", "performed", "performing", "performs", "period", "periodic", "periods", "person", "personal" ], [ "the", "for", "of", "and", "is", "regression", "to", "with", "data", "in" ], [ "the", "rnn", "recurrent", "dynamics", "of", "sequence", "temporal", "time", "lstm", "system" ], [ "groups", "outcomes", "interventions", "machine", "that", "across", "fair", "ml", "fairness", "we" ], [ "the", "to", "clustering", "we", "and", "of", "flow", "for", "with", "network" ], [ "post", "start", "means", "outcome", "lack", "technique", "knowledge", "effects", "hoc", "interpretability" ], [ "explanations", "decision", "schemes", "explain", "aggregation", "aggregated", "aggregate", "methods", "explaining", "explanation" ], [ "the", "to", "of", "and", "is", "we", "on", "in", "search", "our" ], [ "user", "to", "the", "and", "items", "users", "we", "of", "recommendation", "in" ], [ "of", "active", "similarity", "the", "distance", "learning", "metric", "to", "data", "variables" ], [ "the", "computing", "for", "of", "to", "and", "in", "we", "is", "data" ], [ "question", "the", "and", "research", "in", "questions", "of", "to", "for", "open" ], [ "networks", "energy", "to", "for", "and", "we", "of", "privacy", "the", "traffic" ], [ "devices", "sequence", "learning", "protein", "sequences", "device", "power", "edge", "on", "to" ], [ "bias", "the", "games", "error", "that", "we", "variance", "in", "and", "of" ], [ "of", "the", "and", "quantum", "in", "for", "that", "we", "noise", "to" ], [ "the", "points", "point", "local", "global", "to", "signals", "of", "minimax", "equations" ], [ "the", "to", "and", "of", "learning", "in", "geometric", "translation", "we", "view" ], [ "of", "graph", "to", "node", "we", "and", "graphs", "in", "the", "on" ], [ "the", "we", "and", "of", "to", "in", "that", "for", "on", "this" ], [ "image", "of", "and", "images", "gan", "the", "we", "to", "generative", "generation" ], [ "transmission", "wireless", "anomaly", "coding", "channel", "codes", "detection", "anomalies", "that", "the" ], [ "the", "information", "to", "in", "latent", "variational", "of", "likelihood", "autoencoders", "and" ], [ "the", "of", "clustering", "for", "to", "we", "in", "and", "that", "with" ], [ "is", "to", "the", "that", "we", "in", "and", "of", "for", "this" ], [ "of", "is", "objects", "to", "feedback", "online", "the", "algorithms", "by", "that" ], [ "for", "of", "to", "the", "we", "and", "in", "is", "that", "optimization" ], [ "covid", "19", "amenable", "sample", "spectrum", "our", "evidence", "provable", "analysis", "we" ], [ "scene", "we", "tensor", "of", "the", "low", "rank", "approximation", "for", "and" ], [ "the", "in", "we", "to", "model", "and", "of", "is", "this", "that" ], [ "learning", "the", "and", "of", "to", "for", "in", "model", "we", "with" ], [ "of", "the", "and", "prediction", "in", "time", "long", "short", "traffic", "term" ], [ "to", "for", "we", "and", "in", "on", "data", "is", "of", "the" ], [ "off", "fidelity", "trade", "high", "accuracy", "models", "to", "cost", "surrogate", "of" ], [ "robust", "concepts", "to", "and", "that", "the", "cloud", "learning", "meta", "we" ], [ "to", "the", "we", "in", "and", "learning", "on", "of", "training", "data" ], [ "personalized", "perspective", "performed", "performing", "performs", "period", "periodic", "periods", "person", "personal" ], [ "the", "in", "and", "to", "is", "on", "of", "for", "this", "with" ], [ "the", "depth", "manifold", "and", "we", "of", "to", "that", "subspace", "shift" ], [ "of", "and", "networks", "network", "to", "neural", "the", "with", "we", "in" ], [ "and", "adversarial", "the", "to", "attacks", "we", "of", "training", "detection", "in" ], [ "the", "to", "and", "data", "is", "of", "in", "for", "that", "on" ] ]
14.255966
all-MiniLM-L6-v2
0.39
-0.061498
0.18372
0.637337
ArXiv ML Papers
LDA
46
50
[ [ "the", "of", "to", "in", "and", "on", "is", "we", "network", "model" ], [ "and", "the", "of", "we", "is", "in", "to", "for", "with", "that" ], [ "the", "and", "in", "to", "tensor", "traffic", "of", "net", "on", "for" ], [ "inference", "we", "and", "to", "learning", "the", "of", "for", "in", "variational" ], [ "to", "regression", "protein", "knowledge", "learning", "in", "this", "sequences", "for", "sequence" ], [ "of", "in", "and", "the", "to", "channel", "as", "is", "this", "for" ], [ "the", "to", "and", "of", "in", "is", "we", "for", "that", "data" ], [ "the", "of", "to", "we", "that", "in", "is", "and", "with", "for" ], [ "of", "to", "we", "the", "and", "on", "our", "that", "in", "object" ], [ "actions", "in", "local", "minimax", "and", "of", "to", "driving", "action", "the" ], [ "uncertainty", "in", "not", "to", "the", "predictions", "decision", "confidence", "want", "which" ], [ "for", "gradient", "optimization", "policy", "of", "the", "we", "and", "stochastic", "with" ], [ "to", "detection", "the", "of", "objects", "in", "for", "transformation", "we", "cnn" ], [ "metric", "distance", "the", "we", "learning", "to", "of", "from", "trees", "and" ], [ "of", "and", "data", "series", "we", "model", "time", "the", "to", "forecasting" ], [ "representation", "node", "of", "graphs", "embedding", "the", "graph", "nodes", "to", "network" ], [ "personalized", "perspective", "performed", "performing", "performs", "period", "periodic", "periods", "person", "personal" ], [ "personalized", "perspective", "performed", "performing", "performs", "period", "periodic", "periods", "person", "personal" ], [ "in", "is", "that", "to", "the", "and", "of", "with", "we", "problem" ], [ "for", "in", "the", "we", "of", "to", "and", "energy", "cost", "monte" ], [ "features", "video", "each", "fusion", "as", "frame", "frames", "this", "feature", "uses" ], [ "and", "to", "we", "of", "the", "in", "that", "for", "on", "language" ], [ "for", "to", "and", "in", "the", "of", "we", "this", "that", "is" ], [ "of", "are", "in", "region", "regions", "the", "reports", "this", "with", "to" ], [ "personalized", "perspective", "performed", "performing", "performs", "period", "periodic", "periods", "person", "personal" ], [ "data", "of", "the", "kernel", "in", "and", "we", "to", "on", "is" ], [ "in", "domain", "and", "the", "to", "of", "transfer", "target", "learning", "knowledge" ], [ "and", "demand", "bottleneck", "distributed", "each", "in", "communication", "this", "popular", "data" ], [ "the", "and", "privacy", "to", "of", "user", "users", "we", "on", "private" ], [ "personalized", "perspective", "performed", "performing", "performs", "period", "periodic", "periods", "person", "personal" ], [ "in", "measure", "words", "between", "similarity", "vectors", "word", "to", "concepts", "this" ], [ "processes", "our", "complexity", "gaussian", "sample", "approximation", "full", "we", "latent", "with" ], [ "and", "rl", "agent", "the", "learning", "agents", "to", "reinforcement", "environment", "we" ], [ "adversarial", "the", "to", "we", "training", "and", "that", "of", "attacks", "on" ], [ "image", "of", "to", "the", "and", "in", "we", "for", "images", "on" ], [ "on", "to", "the", "in", "method", "ensemble", "prediction", "of", "and", "with" ], [ "of", "the", "and", "to", "in", "for", "from", "with", "that", "is" ], [ "we", "this", "of", "to", "in", "the", "for", "and", "model", "is" ], [ "term", "short", "prediction", "long", "temporal", "dynamics", "spatial", "view", "and", "system" ], [ "for", "to", "the", "learning", "risk", "and", "of", "is", "in", "prediction" ], [ "resolution", "super", "types", "type", "identity", "of", "the", "and", "to", "task" ], [ "algorithm", "of", "to", "we", "the", "that", "is", "in", "for", "and" ], [ "personalized", "perspective", "performed", "performing", "performs", "period", "periodic", "periods", "person", "personal" ], [ "personalized", "perspective", "performed", "performing", "performs", "period", "periodic", "periods", "person", "personal" ], [ "entropy", "function", "loss", "shift", "functions", "margin", "cross", "networks", "and", "neural" ], [ "of", "the", "to", "is", "in", "and", "that", "this", "we", "by" ], [ "in", "the", "and", "to", "of", "memory", "with", "on", "networks", "neural" ], [ "the", "of", "and", "to", "is", "we", "by", "data", "this", "in" ], [ "in", "to", "supervised", "representations", "unsupervised", "of", "the", "we", "training", "be" ], [ "features", "feature", "learning", "deep", "games", "methods", "to", "semi", "for", "and" ] ]
14.030783
all-MiniLM-L6-v2
0.33
-0.072721
0.191163
0.638094
ArXiv ML Papers
Top2Vec
43
10
[ [ "cnns", "cnn", "adversary", "classifiers", "supervised", "adversarial", "adversarially", "softmax", "classifier", "imagenet" ], [ "learnt", "learns", "ai", "bandit", "bandits", "learnable", "reinforcement", "learning", "planning", "softmax" ], [ "rnns", "lstm", "softmax", "corpus", "learns", "autoencoder", "autoencoders", "embeddings", "supervised", "cnn" ], [ "softmax", "forecasts", "predicts", "predict", "prediction", "forecasting", "predicting", "forecast", "classifiers", "rnns" ], [ "regularization", "optimizer", "supervised", "sparse", "lasso", "softmax", "classifiers", "regularized", "classification", "boosting" ], [ "nodes", "graphs", "graph", "networks", "cnn", "cnns", "vertex", "embeddings", "softmax", "supervised" ], [ "supervised", "imagenet", "softmax", "classifying", "cnn", "classifiers", "cnns", "classification", "autoencoders", "neural" ], [ "backpropagation", "imagenet", "softmax", "cnn", "rnns", "cnns", "autoencoders", "neural", "networks", "learns" ], [ "gans", "autoencoders", "imagenet", "cnn", "cnns", "adversarial", "convolutions", "autoencoder", "adversarially", "recognition" ], [ "supervised", "softmax", "generative", "autoencoders", "rnns", "probabilistic", "learns", "backpropagation", "models", "priors" ] ]
656.29102
all-MiniLM-L6-v2
0.54
-0.278403
0.14962
0.846283
ArXiv ML Papers
Top2Vec
44
10
[ [ "supervised", "adversarial", "adversarially", "classifiers", "classifier", "softmax", "classification", "classifying", "adversary", "cnns" ], [ "ai", "reinforcement", "bandit", "planning", "learning", "learns", "learnt", "bandits", "learnable", "softmax" ], [ "softmax", "autoencoders", "learns", "cnns", "cnn", "neural", "networks", "imagenet", "rnns", "backpropagation" ], [ "supervised", "softmax", "embeddings", "graphs", "networks", "cnn", "cnns", "rnns", "nodes", "embedding" ], [ "predicting", "forecasts", "forecasting", "forecast", "prediction", "softmax", "rnns", "classifiers", "predicts", "predict" ], [ "autoencoder", "softmax", "corpus", "learns", "rnns", "cnn", "neural", "autoencoders", "lstm", "cnns" ], [ "lasso", "sparse", "regularization", "minimization", "optimizer", "softmax", "optimizing", "minimize", "optimization", "regularized" ], [ "models", "supervised", "autoencoders", "softmax", "generative", "priors", "learning", "rnns", "probabilistic", "learns" ], [ "softmax", "cnns", "cnn", "autoencoders", "supervised", "recognition", "imagenet", "neural", "classifiers", "classifying" ], [ "autoencoders", "supervised", "cnn", "adversarial", "cnns", "imagenet", "gans", "adversarially", "autoencoder", "convolutions" ] ]
673.252285
all-MiniLM-L6-v2
0.56
-0.26459
0.160135
0.845661
ArXiv ML Papers
Top2Vec
45
10
[ [ "learning", "learns", "learnt", "reinforcement", "bandit", "bandits", "ai", "planning", "learnable", "softmax" ], [ "adversarial", "softmax", "adversarially", "classifier", "classifiers", "supervised", "adversary", "cnns", "classification", "classifying" ], [ "rnns", "lstm", "learns", "autoencoders", "softmax", "corpus", "autoencoder", "supervised", "classifiers", "cnn" ], [ "backpropagation", "neural", "cnns", "cnn", "imagenet", "softmax", "rnns", "autoencoders", "networks", "learns" ], [ "predict", "forecasts", "forecasting", "forecast", "softmax", "predicts", "rnns", "prediction", "predicting", "lstm" ], [ "networks", "nodes", "graphs", "graph", "cnn", "cnns", "embeddings", "softmax", "vertex", "supervised" ], [ "softmax", "autoencoders", "supervised", "models", "rnns", "generative", "backpropagation", "probabilistic", "priors", "learns" ], [ "neural", "classifying", "autoencoders", "softmax", "classifiers", "imagenet", "supervised", "cnns", "classification", "cnn" ], [ "cnn", "autoencoders", "generative", "adversarially", "adversarial", "gans", "cnns", "imagenet", "autoencoder", "convolutions" ], [ "softmax", "benchmarks", "minimize", "minimization", "regularized", "regularization", "optimizer", "lasso", "sparse", "optimizing" ] ]
584.320705
all-MiniLM-L6-v2
0.56
-0.273268
0.174848
0.846271
ArXiv ML Papers
Top2Vec
46
10
[ [ "learning", "ai", "planning", "reinforcement", "bandit", "learnt", "learns", "bandits", "learnable", "softmax" ], [ "adversary", "softmax", "adversarial", "adversarially", "classifiers", "cnns", "supervised", "cnn", "classifier", "imagenet" ], [ "softmax", "autoencoder", "lstm", "supervised", "rnns", "embeddings", "learns", "corpus", "autoencoders", "cnn" ], [ "forecasts", "forecasting", "forecast", "prediction", "predict", "predicting", "predicts", "softmax", "classifiers", "rnns" ], [ "networks", "graphs", "nodes", "graph", "cnn", "softmax", "cnns", "supervised", "vertex", "embeddings" ], [ "classification", "classifying", "imagenet", "autoencoders", "cnn", "classifiers", "cnns", "supervised", "softmax", "classify" ], [ "regularization", "lasso", "softmax", "supervised", "regularized", "optimizer", "classifiers", "sparse", "boosting", "minimization" ], [ "rnns", "softmax", "neural", "imagenet", "backpropagation", "autoencoders", "cnn", "cnns", "networks", "learns" ], [ "adversarially", "cnn", "imagenet", "autoencoders", "cnns", "gans", "adversarial", "generative", "supervised", "autoencoder" ], [ "autoencoders", "learns", "rnns", "generative", "softmax", "models", "supervised", "backpropagation", "priors", "autoencoder" ] ]
634.524863
all-MiniLM-L6-v2
0.53
-0.269991
0.153078
0.853081
ArXiv ML Papers
Top2Vec
43
20
[ [ "adversarially", "classifiers", "discrimination", "bias", "biases", "fairness", "supervised", "adversarial", "discriminate", "classifier" ], [ "softmax", "distributed", "learns", "federated", "adversarially", "learnt", "learning", "supervised", "rnns", "rnn" ], [ "adversary", "security", "private", "adversarial", "softmax", "regularization", "classifiers", "randomized", "adversarially", "privacy" ], [ "adversarially", "adversarial", "cnn", "cnns", "classifiers", "gans", "softmax", "autoencoders", "adversary", "imagenet" ], [ "learnt", "learns", "ai", "bandit", "bandits", "learnable", "reinforcement", "learning", "planning", "softmax" ], [ "backpropagation", "rnns", "learns", "softmax", "neural", "modeling", "models", "autoencoders", "rnn", "learning" ], [ "personalized", "factorization", "recommender", "recommendation", "softmax", "recommendations", "embeddings", "retrieval", "supervised", "ranking" ], [ "lstm", "corpus", "learns", "cnn", "autoencoders", "autoencoder", "rnns", "softmax", "neural", "cnns" ], [ "neural", "cnns", "softmax", "backpropagation", "networks", "regularization", "rnns", "imagenet", "autoencoders", "cnn" ], [ "softmax", "forecasts", "predicts", "predict", "prediction", "forecasting", "predicting", "forecast", "classifiers", "rnns" ], [ "classify", "classifying", "neural", "classification", "classifier", "softmax", "lstm", "classifiers", "autoencoders", "recognition" ], [ "classifiers", "classification", "supervised", "lasso", "sparse", "softmax", "classifying", "regularization", "svm", "regularized" ], [ "nodes", "graphs", "graph", "networks", "cnn", "cnns", "vertex", "embeddings", "softmax", "supervised" ], [ "supervised", "predicting", "dataset", "datasets", "classifying", "classify", "classifiers", "classifier", "classification", "softmax" ], [ "gan", "gans", "adversarial", "generative", "autoencoders", "adversarially", "cnn", "imagenet", "autoencoder", "cnns" ], [ "imagenet", "cnn", "cnns", "supervised", "softmax", "recognition", "autoencoders", "neural", "segmentation", "recognizing" ], [ "bottleneck", "benchmarks", "autoencoder", "networks", "rnns", "imagenet", "autoencoders", "cnns", "cnn", "optimized" ], [ "convolutions", "autoencoders", "recognizing", "supervised", "imagenet", "cnns", "recognition", "cnn", "convolutional", "softmax" ], [ "supervised", "softmax", "generative", "autoencoders", "rnns", "probabilistic", "learns", "backpropagation", "models", "priors" ], [ "optimization", "optimizer", "regularization", "lasso", "minimize", "minimization", "optimizing", "optimize", "optimized", "optimal" ] ]
608.984458
all-MiniLM-L6-v2
0.45
-0.266672
0.173274
0.85077
ArXiv ML Papers
Top2Vec
44
20
[ [ "distributed", "softmax", "federated", "learnt", "learns", "learning", "supervised", "adversarially", "rnns", "rnn" ], [ "classifiers", "adversary", "cnns", "cnn", "attacks", "adversarial", "adversarially", "softmax", "exploiting", "adversaries" ], [ "fairness", "biases", "classifiers", "classifier", "adversarial", "discrimination", "discriminate", "adversarially", "supervised", "bias" ], [ "ai", "reinforcement", "bandit", "planning", "learning", "learns", "learnt", "bandits", "learnable", "softmax" ], [ "backpropagation", "rnns", "softmax", "neural", "autoencoders", "learns", "models", "learning", "modeling", "rnn" ], [ "softmax", "neural", "classifying", "classify", "autoencoders", "classifiers", "classifier", "classification", "supervised", "lstm" ], [ "graphs", "softmax", "supervised", "networks", "cnn", "nodes", "graph", "cnns", "embeddings", "rnns" ], [ "cnns", "cnn", "imagenet", "softmax", "regularization", "neural", "networks", "backpropagation", "autoencoders", "rnns" ], [ "predicting", "forecasts", "forecasting", "forecast", "prediction", "softmax", "rnns", "classifiers", "predicts", "predict" ], [ "embeddings", "recommender", "recommendations", "recommendation", "supervised", "softmax", "factorization", "personalized", "ranking", "retrieval" ], [ "autoencoder", "softmax", "corpus", "learns", "rnns", "cnn", "neural", "autoencoders", "lstm", "cnns" ], [ "classifier", "classifying", "boosting", "supervised", "classify", "ensemble", "classification", "classifiers", "ensembles", "softmax" ], [ "classification", "metrics", "classifiers", "supervised", "softmax", "metric", "similarity", "classifying", "classifier", "regularization" ], [ "optimizer", "regularization", "sparse", "lasso", "regularized", "clustering", "tensors", "softmax", "minimization", "algorithms" ], [ "autoencoders", "imagenet", "benchmarks", "bottleneck", "cnn", "rnns", "cnns", "networks", "neural", "optimized" ], [ "imagenet", "generative", "cnn", "autoencoder", "autoencoders", "gans", "adversarially", "adversarial", "gan", "cnns" ], [ "lasso", "optimizer", "optimize", "optimization", "minimization", "minimize", "regularization", "optimizing", "optimized", "optimizes" ], [ "models", "supervised", "autoencoders", "softmax", "generative", "priors", "learning", "rnns", "probabilistic", "learns" ], [ "imagenet", "cnn", "cnns", "recognition", "softmax", "supervised", "autoencoders", "neural", "segmentation", "recognizing" ], [ "recognizing", "supervised", "convolutions", "imagenet", "cnns", "recognition", "cnn", "autoencoders", "convolutional", "classifiers" ] ]
535.264348
all-MiniLM-L6-v2
0.47
-0.253657
0.174505
0.844253
ArXiv ML Papers
Top2Vec
45
20
[ [ "learning", "learns", "learnt", "reinforcement", "bandit", "bandits", "ai", "planning", "learnable", "softmax" ], [ "softmax", "distributed", "learns", "federated", "adversarially", "learnt", "learning", "supervised", "rnns", "rnn" ], [ "adversary", "adversarial", "adversarially", "cnns", "classifiers", "softmax", "cnn", "exploiting", "attacks", "adversaries" ], [ "biases", "bias", "fairness", "adversarially", "classifiers", "adversarial", "supervised", "discrimination", "discriminate", "classifier" ], [ "rnns", "lstm", "learns", "autoencoders", "softmax", "corpus", "autoencoder", "supervised", "classifiers", "cnn" ], [ "softmax", "learning", "backpropagation", "autoencoders", "modeling", "rnns", "learns", "neural", "models", "rnn" ], [ "imagenet", "networks", "regularization", "neural", "neuron", "backpropagation", "cnns", "cnn", "softmax", "adversarial" ], [ "classifier", "classifiers", "classifying", "classification", "classify", "supervised", "softmax", "ensemble", "boosting", "ensembles" ], [ "predict", "forecasts", "forecasting", "forecast", "softmax", "predicts", "rnns", "prediction", "predicting", "lstm" ], [ "classifier", "classification", "lstm", "classifiers", "classifying", "softmax", "classify", "supervised", "recognition", "neural" ], [ "classifiers", "predicting", "softmax", "supervised", "datasets", "classifying", "classification", "classifier", "dataset", "classify" ], [ "networks", "nodes", "graphs", "graph", "cnn", "cnns", "embeddings", "softmax", "vertex", "supervised" ], [ "gan", "gans", "generative", "adversarial", "autoencoders", "adversarially", "cnn", "imagenet", "autoencoder", "cnns" ], [ "optimization", "minimization", "optimizer", "optimizing", "lasso", "minimize", "regularization", "optimal", "softmax", "optimize" ], [ "softmax", "autoencoders", "supervised", "models", "rnns", "generative", "backpropagation", "probabilistic", "priors", "learns" ], [ "autoencoders", "imagenet", "benchmarks", "bottleneck", "cnn", "rnns", "cnns", "networks", "neural", "optimized" ], [ "cnn", "imagenet", "autoencoders", "neural", "supervised", "softmax", "cnns", "recognition", "segmentation", "recognizing" ], [ "cnn", "imagenet", "cnns", "recognition", "convolutions", "autoencoders", "recognizing", "supervised", "softmax", "convolutional" ], [ "supervised", "metrics", "classifiers", "classifying", "softmax", "classification", "classifier", "regularization", "metric", "similarity" ], [ "regularized", "lasso", "sparse", "regularization", "minimization", "softmax", "tensors", "optimizer", "clustering", "tensor" ] ]
588.672674
all-MiniLM-L6-v2
0.455
-0.250703
0.178586
0.840558
ArXiv ML Papers
Top2Vec
46
20
[ [ "learning", "ai", "planning", "reinforcement", "bandit", "learnt", "learns", "bandits", "learnable", "softmax" ], [ "adversarially", "distributed", "softmax", "federated", "learning", "learns", "learnt", "supervised", "rnns", "rnn" ], [ "adversarial", "adversarially", "cnns", "adversary", "cnn", "adversaries", "attacks", "softmax", "classifiers", "autoencoders" ], [ "fairness", "adversarially", "bias", "biases", "classifiers", "discrimination", "adversarial", "discriminate", "supervised", "classifier" ], [ "softmax", "autoencoder", "lstm", "supervised", "rnns", "embeddings", "learns", "corpus", "autoencoders", "cnn" ], [ "learns", "rnns", "neural", "backpropagation", "learning", "softmax", "autoencoders", "models", "rnn", "modeling" ], [ "forecasts", "forecasting", "forecast", "prediction", "predict", "predicting", "predicts", "softmax", "classifiers", "rnns" ], [ "softmax", "networks", "neural", "backpropagation", "cnns", "imagenet", "regularization", "cnn", "neuron", "rnns" ], [ "classify", "classifying", "eeg", "supervised", "classifiers", "classifier", "classification", "recognition", "lstm", "svm" ], [ "networks", "graphs", "nodes", "graph", "cnn", "softmax", "cnns", "supervised", "vertex", "embeddings" ], [ "dataset", "classifying", "classifiers", "predicting", "datasets", "softmax", "supervised", "classification", "classify", "rnns" ], [ "classifiers", "ensemble", "softmax", "boosting", "supervised", "classification", "classifying", "classify", "classifier", "ensembles" ], [ "imagenet", "cnns", "cnn", "supervised", "softmax", "neural", "recognition", "autoencoders", "segmentation", "recognizing" ], [ "regularization", "classifiers", "metrics", "classifier", "classifying", "classification", "softmax", "supervised", "similarity", "metric" ], [ "sparse", "softmax", "clustering", "minimization", "optimizer", "lasso", "algorithms", "regularized", "regularization", "supervised" ], [ "gan", "autoencoders", "adversarially", "adversarial", "autoencoder", "generative", "imagenet", "gans", "cnns", "cnn" ], [ "cnns", "rnns", "cnn", "imagenet", "autoencoders", "bottleneck", "benchmarks", "networks", "softmax", "neural" ], [ "imagenet", "recognizing", "recognition", "convolutions", "cnns", "cnn", "supervised", "autoencoders", "convolutional", "softmax" ], [ "optimization", "minimization", "optimizer", "optimizing", "optimize", "minimize", "regularization", "lasso", "optimized", "optimizes" ], [ "autoencoders", "learns", "rnns", "generative", "softmax", "models", "supervised", "backpropagation", "priors", "autoencoder" ] ]
631.574545
all-MiniLM-L6-v2
0.45
-0.259184
0.173666
0.84807
ArXiv ML Papers
Top2Vec
43
30
[ [ "adversarially", "classifiers", "discrimination", "bias", "biases", "fairness", "supervised", "adversarial", "discriminate", "classifier" ], [ "softmax", "distributed", "learns", "federated", "adversarially", "learnt", "learning", "supervised", "rnns", "rnn" ], [ "adversary", "security", "private", "adversarial", "softmax", "regularization", "classifiers", "randomized", "adversarially", "privacy" ], [ "adversary", "cnns", "attacks", "adversarial", "adversarially", "cnn", "classifiers", "adversaries", "softmax", "attacker" ], [ "learnt", "learns", "ai", "bandit", "bandits", "learnable", "reinforcement", "learning", "planning", "softmax" ], [ "backpropagation", "rnns", "learns", "softmax", "neural", "modeling", "models", "autoencoders", "rnn", "learning" ], [ "interpretability", "classifiers", "adversarially", "explanations", "softmax", "adversarial", "supervised", "neural", "cnns", "cnn" ], [ "personalized", "factorization", "recommender", "recommendation", "softmax", "recommendations", "embeddings", "retrieval", "supervised", "ranking" ], [ "lstm", "corpus", "learns", "cnn", "autoencoders", "autoencoder", "rnns", "softmax", "neural", "cnns" ], [ "neural", "cnns", "softmax", "backpropagation", "networks", "regularization", "rnns", "imagenet", "autoencoders", "cnn" ], [ "softmax", "forecasts", "predicts", "predict", "prediction", "forecasting", "predicting", "forecast", "classifiers", "rnns" ], [ "classify", "classifying", "neural", "classification", "classifier", "softmax", "lstm", "classifiers", "autoencoders", "recognition" ], [ "classifiers", "classification", "supervised", "lasso", "sparse", "softmax", "classifying", "regularization", "svm", "regularized" ], [ "nodes", "graphs", "graph", "networks", "cnn", "cnns", "vertex", "embeddings", "softmax", "supervised" ], [ "supervised", "predicting", "dataset", "datasets", "classifying", "classify", "classifiers", "classifier", "classification", "softmax" ], [ "gan", "gans", "adversarial", "generative", "autoencoders", "adversarially", "cnn", "imagenet", "autoencoder", "cnns" ], [ "imagenet", "cnn", "cnns", "supervised", "softmax", "recognition", "autoencoders", "neural", "segmentation", "recognizing" ], [ "bottleneck", "benchmarks", "autoencoder", "networks", "rnns", "imagenet", "autoencoders", "cnns", "cnn", "optimized" ], [ "convolutions", "autoencoders", "recognizing", "supervised", "imagenet", "cnns", "recognition", "cnn", "convolutional", "softmax" ], [ "supervised", "softmax", "generative", "autoencoders", "rnns", "probabilistic", "learns", "backpropagation", "models", "priors" ], [ "optimization", "optimizer", "regularization", "lasso", "minimize", "minimization", "optimizing", "optimize", "optimized", "optimal" ] ]
577.164902
all-MiniLM-L6-v2
0.452381
-0.264192
0.17378
0.84458
ArXiv ML Papers
Top2Vec
44
30
[ [ "distributed", "softmax", "federated", "learnt", "learns", "learning", "supervised", "adversarially", "rnns", "rnn" ], [ "classifiers", "adversary", "cnns", "cnn", "attacks", "adversarial", "adversarially", "softmax", "exploiting", "adversaries" ], [ "fairness", "biases", "classifiers", "classifier", "adversarial", "discrimination", "discriminate", "adversarially", "supervised", "bias" ], [ "supervised", "leveraging", "inference", "estimating", "interventions", "causal", "models", "priors", "generalization", "observational" ], [ "ai", "reinforcement", "bandit", "planning", "learning", "learns", "learnt", "bandits", "learnable", "softmax" ], [ "backpropagation", "rnns", "softmax", "neural", "autoencoders", "learns", "models", "learning", "modeling", "rnn" ], [ "softmax", "neural", "classifying", "classify", "autoencoders", "classifiers", "classifier", "classification", "supervised", "lstm" ], [ "graphs", "softmax", "supervised", "networks", "cnn", "nodes", "graph", "cnns", "embeddings", "rnns" ], [ "cnns", "cnn", "imagenet", "softmax", "regularization", "neural", "networks", "backpropagation", "autoencoders", "rnns" ], [ "predicting", "forecasts", "forecasting", "forecast", "prediction", "softmax", "rnns", "classifiers", "predicts", "predict" ], [ "embeddings", "recommender", "recommendations", "recommendation", "supervised", "softmax", "factorization", "personalized", "ranking", "retrieval" ], [ "autoencoder", "softmax", "corpus", "learns", "rnns", "cnn", "neural", "autoencoders", "lstm", "cnns" ], [ "classifier", "classifying", "boosting", "supervised", "classify", "ensemble", "classification", "classifiers", "ensembles", "softmax" ], [ "classification", "metrics", "classifiers", "supervised", "softmax", "metric", "similarity", "classifying", "classifier", "regularization" ], [ "optimizer", "regularization", "sparse", "lasso", "regularized", "clustering", "tensors", "softmax", "minimization", "algorithms" ], [ "autoencoders", "imagenet", "benchmarks", "bottleneck", "cnn", "rnns", "cnns", "networks", "neural", "optimized" ], [ "imagenet", "generative", "cnn", "autoencoder", "autoencoders", "gans", "adversarially", "adversarial", "gan", "cnns" ], [ "lasso", "optimizer", "optimize", "optimization", "minimization", "minimize", "regularization", "optimizing", "optimized", "optimizes" ], [ "supervised", "learns", "priors", "autoencoders", "backpropagation", "probabilistic", "rnns", "softmax", "generative", "models" ], [ "imagenet", "cnn", "cnns", "recognition", "softmax", "supervised", "autoencoders", "neural", "segmentation", "recognizing" ], [ "recognizing", "supervised", "convolutions", "imagenet", "cnns", "recognition", "cnn", "autoencoders", "convolutional", "classifiers" ] ]
601.105897
all-MiniLM-L6-v2
0.480952
-0.2564
0.172465
0.844825
ArXiv ML Papers
Top2Vec
45
30
[ [ "learning", "learns", "learnt", "reinforcement", "bandit", "bandits", "ai", "planning", "learnable", "softmax" ], [ "softmax", "distributed", "learns", "federated", "adversarially", "learnt", "learning", "supervised", "rnns", "rnn" ], [ "adversary", "adversarial", "adversarially", "cnns", "classifiers", "softmax", "cnn", "exploiting", "attacks", "adversaries" ], [ "biases", "bias", "fairness", "adversarially", "classifiers", "adversarial", "supervised", "discrimination", "discriminate", "classifier" ], [ "rnns", "lstm", "learns", "autoencoders", "softmax", "corpus", "autoencoder", "supervised", "classifiers", "cnn" ], [ "softmax", "learning", "backpropagation", "autoencoders", "modeling", "rnns", "learns", "neural", "models", "rnn" ], [ "imagenet", "networks", "regularization", "neural", "neuron", "backpropagation", "cnns", "cnn", "softmax", "adversarial" ], [ "classifier", "classifiers", "classifying", "classification", "classify", "supervised", "softmax", "ensemble", "boosting", "ensembles" ], [ "predict", "forecasts", "forecasting", "forecast", "softmax", "predicts", "rnns", "prediction", "predicting", "lstm" ], [ "classifier", "classification", "lstm", "classifiers", "classifying", "softmax", "classify", "supervised", "recognition", "neural" ], [ "classifiers", "predicting", "softmax", "supervised", "datasets", "classifying", "classification", "classifier", "dataset", "classify" ], [ "networks", "nodes", "graphs", "graph", "cnn", "cnns", "embeddings", "softmax", "vertex", "supervised" ], [ "gan", "gans", "generative", "adversarial", "autoencoders", "adversarially", "cnn", "imagenet", "autoencoder", "cnns" ], [ "optimization", "minimization", "optimizer", "optimizing", "lasso", "minimize", "regularization", "optimal", "softmax", "optimize" ], [ "softmax", "autoencoders", "supervised", "models", "rnns", "generative", "backpropagation", "probabilistic", "priors", "learns" ], [ "autoencoders", "imagenet", "benchmarks", "bottleneck", "cnn", "rnns", "cnns", "networks", "neural", "optimized" ], [ "cnn", "imagenet", "autoencoders", "neural", "supervised", "softmax", "cnns", "recognition", "segmentation", "recognizing" ], [ "cnns", "recognition", "imagenet", "cnn", "convolutions", "recognizing", "supervised", "autoencoders", "softmax", "classifiers" ], [ "maps", "cnns", "cnn", "imagenet", "3d", "recognition", "recognizing", "convolutions", "supervised", "vision" ], [ "supervised", "metrics", "classifiers", "classifying", "softmax", "classification", "classifier", "regularization", "metric", "similarity" ], [ "regularized", "lasso", "sparse", "regularization", "minimization", "softmax", "tensors", "optimizer", "clustering", "tensor" ] ]
597.81598
all-MiniLM-L6-v2
0.442857
-0.252986
0.175005
0.843511
ArXiv ML Papers
Top2Vec
46
30
[ [ "learning", "ai", "planning", "reinforcement", "bandit", "learnt", "learns", "bandits", "learnable", "softmax" ], [ "adversarially", "distributed", "softmax", "federated", "learning", "learns", "learnt", "supervised", "rnns", "rnn" ], [ "adversary", "adversarial", "adversarially", "cnns", "adversaries", "cnn", "classifiers", "attacks", "softmax", "attacker" ], [ "regularization", "adversarially", "adversarial", "privacy", "private", "randomized", "adversary", "normalization", "regularized", "softmax" ], [ "fairness", "adversarially", "bias", "biases", "classifiers", "discrimination", "adversarial", "discriminate", "supervised", "classifier" ], [ "softmax", "autoencoder", "lstm", "supervised", "rnns", "embeddings", "learns", "corpus", "autoencoders", "cnn" ], [ "learns", "rnns", "neural", "backpropagation", "learning", "softmax", "autoencoders", "models", "rnn", "modeling" ], [ "forecasts", "forecasting", "forecast", "prediction", "predict", "predicting", "predicts", "softmax", "classifiers", "rnns" ], [ "softmax", "networks", "neural", "backpropagation", "cnns", "imagenet", "regularization", "cnn", "neuron", "rnns" ], [ "classify", "classifying", "eeg", "supervised", "classifiers", "classifier", "classification", "recognition", "lstm", "svm" ], [ "networks", "graphs", "nodes", "graph", "cnn", "softmax", "cnns", "supervised", "vertex", "embeddings" ], [ "dataset", "classifying", "classifiers", "predicting", "datasets", "softmax", "supervised", "classification", "classify", "rnns" ], [ "classifiers", "ensemble", "softmax", "boosting", "supervised", "classification", "classifying", "classify", "classifier", "ensembles" ], [ "imagenet", "cnns", "cnn", "supervised", "softmax", "neural", "recognition", "autoencoders", "segmentation", "recognizing" ], [ "regularization", "classifiers", "metrics", "classifier", "classifying", "classification", "softmax", "supervised", "similarity", "metric" ], [ "sparse", "softmax", "clustering", "minimization", "optimizer", "lasso", "algorithms", "regularized", "regularization", "supervised" ], [ "gan", "autoencoders", "adversarially", "adversarial", "autoencoder", "generative", "imagenet", "gans", "cnns", "cnn" ], [ "cnns", "rnns", "cnn", "imagenet", "autoencoders", "bottleneck", "benchmarks", "networks", "softmax", "neural" ], [ "imagenet", "recognizing", "recognition", "convolutions", "cnns", "cnn", "supervised", "autoencoders", "convolutional", "softmax" ], [ "optimization", "minimization", "optimizer", "optimizing", "optimize", "minimize", "regularization", "lasso", "optimized", "optimizes" ], [ "autoencoders", "learns", "rnns", "generative", "softmax", "models", "supervised", "backpropagation", "priors", "autoencoder" ] ]
498.065006
all-MiniLM-L6-v2
0.452381
-0.262968
0.171184
0.850088
ArXiv ML Papers
Top2Vec
43
40
[ [ "adversarially", "classifiers", "discrimination", "bias", "biases", "fairness", "supervised", "adversarial", "discriminate", "classifier" ], [ "softmax", "distributed", "learns", "federated", "adversarially", "learnt", "learning", "supervised", "rnns", "rnn" ], [ "adversary", "security", "private", "adversarial", "softmax", "regularization", "classifiers", "randomized", "adversarially", "privacy" ], [ "adversary", "cnns", "attacks", "adversarial", "adversarially", "cnn", "classifiers", "adversaries", "softmax", "attacker" ], [ "learnt", "learns", "ai", "bandit", "bandits", "learnable", "reinforcement", "learning", "planning", "softmax" ], [ "backpropagation", "rnns", "learns", "softmax", "neural", "modeling", "models", "autoencoders", "rnn", "learning" ], [ "interpretability", "classifiers", "adversarially", "explanations", "softmax", "adversarial", "supervised", "neural", "cnns", "cnn" ], [ "personalized", "factorization", "recommender", "recommendation", "softmax", "recommendations", "embeddings", "retrieval", "supervised", "ranking" ], [ "lstm", "corpus", "learns", "cnn", "autoencoders", "autoencoder", "rnns", "softmax", "neural", "cnns" ], [ "neural", "cnns", "softmax", "backpropagation", "networks", "regularization", "rnns", "imagenet", "autoencoders", "cnn" ], [ "softmax", "forecasts", "predicts", "predict", "prediction", "forecasting", "predicting", "forecast", "classifiers", "rnns" ], [ "classify", "classifying", "neural", "classification", "classifier", "softmax", "lstm", "classifiers", "autoencoders", "recognition" ], [ "classifiers", "classification", "supervised", "lasso", "sparse", "softmax", "classifying", "regularization", "svm", "regularized" ], [ "nodes", "graphs", "graph", "networks", "cnn", "cnns", "vertex", "embeddings", "softmax", "supervised" ], [ "supervised", "predicting", "dataset", "datasets", "classifying", "classify", "classifiers", "classifier", "classification", "softmax" ], [ "gan", "gans", "adversarial", "generative", "autoencoders", "adversarially", "cnn", "imagenet", "autoencoder", "cnns" ], [ "imagenet", "cnn", "cnns", "supervised", "softmax", "recognition", "autoencoders", "neural", "segmentation", "recognizing" ], [ "bottleneck", "benchmarks", "autoencoder", "networks", "rnns", "imagenet", "autoencoders", "cnns", "cnn", "optimized" ], [ "convolutions", "autoencoders", "recognizing", "supervised", "imagenet", "cnns", "recognition", "cnn", "convolutional", "softmax" ], [ "supervised", "softmax", "generative", "autoencoders", "rnns", "probabilistic", "learns", "backpropagation", "models", "priors" ], [ "optimization", "optimizer", "regularization", "lasso", "minimize", "minimization", "optimizing", "optimize", "optimized", "optimal" ] ]
640.464483
all-MiniLM-L6-v2
0.452381
-0.264192
0.17378
0.845774
ArXiv ML Papers
Top2Vec
44
40
[ [ "distributed", "softmax", "federated", "learnt", "learns", "learning", "supervised", "adversarially", "rnns", "rnn" ], [ "classifiers", "adversary", "cnns", "cnn", "attacks", "adversarial", "adversarially", "softmax", "exploiting", "adversaries" ], [ "fairness", "biases", "classifiers", "classifier", "adversarial", "discrimination", "discriminate", "adversarially", "supervised", "bias" ], [ "supervised", "leveraging", "inference", "estimating", "interventions", "causal", "models", "priors", "generalization", "observational" ], [ "ai", "reinforcement", "bandit", "planning", "learning", "learns", "learnt", "bandits", "learnable", "softmax" ], [ "backpropagation", "rnns", "softmax", "neural", "autoencoders", "learns", "models", "learning", "modeling", "rnn" ], [ "softmax", "neural", "classifying", "classify", "autoencoders", "classifiers", "classifier", "classification", "supervised", "lstm" ], [ "graphs", "softmax", "supervised", "networks", "cnn", "nodes", "graph", "cnns", "embeddings", "rnns" ], [ "cnns", "cnn", "imagenet", "softmax", "regularization", "neural", "networks", "backpropagation", "autoencoders", "rnns" ], [ "predicting", "forecasts", "forecasting", "forecast", "prediction", "softmax", "rnns", "classifiers", "predicts", "predict" ], [ "embeddings", "recommender", "recommendations", "recommendation", "supervised", "softmax", "factorization", "personalized", "ranking", "retrieval" ], [ "autoencoder", "softmax", "corpus", "learns", "rnns", "cnn", "neural", "autoencoders", "lstm", "cnns" ], [ "classifier", "classifying", "boosting", "supervised", "classify", "ensemble", "classification", "classifiers", "ensembles", "softmax" ], [ "classification", "metrics", "classifiers", "supervised", "softmax", "metric", "similarity", "classifying", "classifier", "regularization" ], [ "optimizer", "regularization", "sparse", "lasso", "regularized", "clustering", "tensors", "softmax", "minimization", "algorithms" ], [ "autoencoders", "imagenet", "benchmarks", "bottleneck", "cnn", "rnns", "cnns", "networks", "neural", "optimized" ], [ "imagenet", "generative", "cnn", "autoencoder", "autoencoders", "gans", "adversarially", "adversarial", "gan", "cnns" ], [ "lasso", "optimizer", "optimize", "optimization", "minimization", "minimize", "regularization", "optimizing", "optimized", "optimizes" ], [ "supervised", "learns", "priors", "autoencoders", "backpropagation", "probabilistic", "rnns", "softmax", "generative", "models" ], [ "imagenet", "cnn", "cnns", "recognition", "softmax", "supervised", "autoencoders", "neural", "segmentation", "recognizing" ], [ "recognizing", "supervised", "convolutions", "imagenet", "cnns", "recognition", "cnn", "autoencoders", "convolutional", "classifiers" ] ]
569.045129
all-MiniLM-L6-v2
0.480952
-0.2564
0.172465
0.846093
ArXiv ML Papers
Top2Vec
45
40
[ [ "learning", "learns", "learnt", "reinforcement", "bandit", "bandits", "ai", "planning", "learnable", "softmax" ], [ "softmax", "distributed", "learns", "federated", "adversarially", "learnt", "learning", "supervised", "rnns", "rnn" ], [ "adversary", "adversarial", "adversarially", "cnns", "classifiers", "softmax", "cnn", "exploiting", "attacks", "adversaries" ], [ "biases", "bias", "fairness", "adversarially", "classifiers", "adversarial", "supervised", "discrimination", "discriminate", "classifier" ], [ "rnns", "lstm", "learns", "autoencoders", "softmax", "corpus", "autoencoder", "supervised", "classifiers", "cnn" ], [ "softmax", "learning", "backpropagation", "autoencoders", "modeling", "rnns", "learns", "neural", "models", "rnn" ], [ "imagenet", "networks", "regularization", "neural", "neuron", "backpropagation", "cnns", "cnn", "softmax", "adversarial" ], [ "classifier", "classifiers", "classifying", "classification", "classify", "supervised", "softmax", "ensemble", "boosting", "ensembles" ], [ "predict", "forecasts", "forecasting", "forecast", "softmax", "predicts", "rnns", "prediction", "predicting", "lstm" ], [ "classifier", "classification", "lstm", "classifiers", "classifying", "softmax", "classify", "supervised", "recognition", "neural" ], [ "classifiers", "predicting", "softmax", "supervised", "datasets", "classifying", "classification", "classifier", "dataset", "classify" ], [ "networks", "nodes", "graphs", "graph", "cnn", "cnns", "embeddings", "softmax", "vertex", "supervised" ], [ "gan", "gans", "generative", "adversarial", "autoencoders", "adversarially", "cnn", "imagenet", "autoencoder", "cnns" ], [ "optimization", "minimization", "optimizer", "optimizing", "lasso", "minimize", "regularization", "optimal", "softmax", "optimize" ], [ "softmax", "autoencoders", "supervised", "models", "rnns", "generative", "backpropagation", "probabilistic", "priors", "learns" ], [ "autoencoders", "imagenet", "benchmarks", "bottleneck", "cnn", "rnns", "cnns", "networks", "neural", "optimized" ], [ "cnn", "imagenet", "autoencoders", "neural", "supervised", "softmax", "cnns", "recognition", "segmentation", "recognizing" ], [ "cnns", "recognition", "imagenet", "cnn", "convolutions", "recognizing", "supervised", "autoencoders", "softmax", "classifiers" ], [ "maps", "cnns", "cnn", "imagenet", "3d", "recognition", "recognizing", "convolutions", "supervised", "vision" ], [ "supervised", "metrics", "classifiers", "classifying", "softmax", "classification", "classifier", "regularization", "metric", "similarity" ], [ "regularized", "lasso", "sparse", "regularization", "minimization", "softmax", "tensors", "optimizer", "clustering", "tensor" ] ]
569.015667
all-MiniLM-L6-v2
0.442857
-0.252986
0.175005
0.842255
ArXiv ML Papers
Top2Vec
46
40
[ [ "learning", "ai", "planning", "reinforcement", "bandit", "learnt", "learns", "bandits", "learnable", "softmax" ], [ "adversarially", "distributed", "softmax", "federated", "learning", "learns", "learnt", "supervised", "rnns", "rnn" ], [ "adversary", "adversarial", "adversarially", "cnns", "adversaries", "cnn", "classifiers", "attacks", "softmax", "attacker" ], [ "regularization", "adversarially", "adversarial", "privacy", "private", "randomized", "adversary", "normalization", "regularized", "softmax" ], [ "fairness", "adversarially", "bias", "biases", "classifiers", "discrimination", "adversarial", "discriminate", "supervised", "classifier" ], [ "softmax", "autoencoder", "lstm", "supervised", "rnns", "embeddings", "learns", "corpus", "autoencoders", "cnn" ], [ "learns", "rnns", "neural", "backpropagation", "learning", "softmax", "autoencoders", "models", "rnn", "modeling" ], [ "forecasts", "forecasting", "forecast", "prediction", "predict", "predicting", "predicts", "softmax", "classifiers", "rnns" ], [ "softmax", "networks", "neural", "backpropagation", "cnns", "imagenet", "regularization", "cnn", "neuron", "rnns" ], [ "classify", "classifying", "eeg", "supervised", "classifiers", "classifier", "classification", "recognition", "lstm", "svm" ], [ "networks", "graphs", "nodes", "graph", "cnn", "softmax", "cnns", "supervised", "vertex", "embeddings" ], [ "dataset", "classifying", "classifiers", "predicting", "datasets", "softmax", "supervised", "classification", "classify", "rnns" ], [ "classifiers", "ensemble", "softmax", "boosting", "supervised", "classification", "classifying", "classify", "classifier", "ensembles" ], [ "imagenet", "cnns", "cnn", "supervised", "softmax", "neural", "recognition", "autoencoders", "segmentation", "recognizing" ], [ "regularization", "classifiers", "metrics", "classifier", "classifying", "classification", "softmax", "supervised", "similarity", "metric" ], [ "sparse", "softmax", "clustering", "minimization", "optimizer", "lasso", "algorithms", "regularized", "regularization", "supervised" ], [ "gan", "autoencoders", "adversarially", "adversarial", "autoencoder", "generative", "imagenet", "gans", "cnns", "cnn" ], [ "cnns", "rnns", "cnn", "imagenet", "autoencoders", "bottleneck", "benchmarks", "networks", "softmax", "neural" ], [ "imagenet", "recognizing", "recognition", "convolutions", "cnns", "cnn", "supervised", "autoencoders", "convolutional", "softmax" ], [ "optimization", "minimization", "optimizer", "optimizing", "optimize", "minimize", "regularization", "lasso", "optimized", "optimizes" ], [ "autoencoders", "learns", "rnns", "generative", "softmax", "models", "supervised", "backpropagation", "priors", "autoencoder" ] ]
584.353214
all-MiniLM-L6-v2
0.452381
-0.262968
0.171184
0.844559
ArXiv ML Papers
Top2Vec
43
50
[ [ "adversarially", "classifiers", "discrimination", "bias", "biases", "fairness", "supervised", "adversarial", "discriminate", "classifier" ], [ "softmax", "distributed", "learns", "federated", "adversarially", "learnt", "learning", "supervised", "rnns", "rnn" ], [ "adversary", "security", "private", "adversarial", "softmax", "regularization", "classifiers", "randomized", "adversarially", "privacy" ], [ "adversary", "cnns", "attacks", "adversarial", "adversarially", "cnn", "classifiers", "adversaries", "softmax", "attacker" ], [ "learnt", "learns", "ai", "bandit", "bandits", "learnable", "reinforcement", "learning", "planning", "softmax" ], [ "backpropagation", "rnns", "learns", "softmax", "neural", "modeling", "models", "autoencoders", "rnn", "learning" ], [ "interpretability", "classifiers", "adversarially", "explanations", "softmax", "adversarial", "supervised", "neural", "cnns", "cnn" ], [ "personalized", "factorization", "recommender", "recommendation", "softmax", "recommendations", "embeddings", "retrieval", "supervised", "ranking" ], [ "lstm", "corpus", "learns", "cnn", "autoencoders", "autoencoder", "rnns", "softmax", "neural", "cnns" ], [ "neural", "cnns", "softmax", "backpropagation", "networks", "regularization", "rnns", "imagenet", "autoencoders", "cnn" ], [ "softmax", "forecasts", "predicts", "predict", "prediction", "forecasting", "predicting", "forecast", "classifiers", "rnns" ], [ "classify", "classifying", "neural", "classification", "classifier", "softmax", "lstm", "classifiers", "autoencoders", "recognition" ], [ "classifiers", "classification", "supervised", "lasso", "sparse", "softmax", "classifying", "regularization", "svm", "regularized" ], [ "nodes", "graphs", "graph", "networks", "cnn", "cnns", "vertex", "embeddings", "softmax", "supervised" ], [ "supervised", "predicting", "dataset", "datasets", "classifying", "classify", "classifiers", "classifier", "classification", "softmax" ], [ "gan", "gans", "adversarial", "generative", "autoencoders", "adversarially", "cnn", "imagenet", "autoencoder", "cnns" ], [ "imagenet", "cnn", "cnns", "supervised", "softmax", "recognition", "autoencoders", "neural", "segmentation", "recognizing" ], [ "bottleneck", "benchmarks", "autoencoder", "networks", "rnns", "imagenet", "autoencoders", "cnns", "cnn", "optimized" ], [ "convolutions", "autoencoders", "recognizing", "supervised", "imagenet", "cnns", "recognition", "cnn", "convolutional", "softmax" ], [ "supervised", "softmax", "generative", "autoencoders", "rnns", "probabilistic", "learns", "backpropagation", "models", "priors" ], [ "optimization", "optimizer", "regularization", "lasso", "minimize", "minimization", "optimizing", "optimize", "optimized", "optimal" ] ]
498.800022
all-MiniLM-L6-v2
0.452381
-0.264192
0.17378
0.850448
ArXiv ML Papers
Top2Vec
44
50
[ [ "distributed", "softmax", "federated", "learnt", "learns", "learning", "supervised", "adversarially", "rnns", "rnn" ], [ "classifiers", "adversary", "cnns", "cnn", "attacks", "adversarial", "adversarially", "softmax", "exploiting", "adversaries" ], [ "fairness", "biases", "classifiers", "classifier", "adversarial", "discrimination", "discriminate", "adversarially", "supervised", "bias" ], [ "supervised", "leveraging", "inference", "estimating", "interventions", "causal", "models", "priors", "generalization", "observational" ], [ "ai", "reinforcement", "bandit", "planning", "learning", "learns", "learnt", "bandits", "learnable", "softmax" ], [ "backpropagation", "rnns", "softmax", "neural", "autoencoders", "learns", "models", "learning", "modeling", "rnn" ], [ "softmax", "neural", "classifying", "classify", "autoencoders", "classifiers", "classifier", "classification", "supervised", "lstm" ], [ "graphs", "softmax", "supervised", "networks", "cnn", "nodes", "graph", "cnns", "embeddings", "rnns" ], [ "cnns", "cnn", "imagenet", "softmax", "regularization", "neural", "networks", "backpropagation", "autoencoders", "rnns" ], [ "predicting", "forecasts", "forecasting", "forecast", "prediction", "softmax", "rnns", "classifiers", "predicts", "predict" ], [ "embeddings", "recommender", "recommendations", "recommendation", "supervised", "softmax", "factorization", "personalized", "ranking", "retrieval" ], [ "autoencoder", "softmax", "corpus", "learns", "rnns", "cnn", "neural", "autoencoders", "lstm", "cnns" ], [ "classifier", "classifying", "boosting", "supervised", "classify", "ensemble", "classification", "classifiers", "ensembles", "softmax" ], [ "classification", "metrics", "classifiers", "supervised", "softmax", "metric", "similarity", "classifying", "classifier", "regularization" ], [ "optimizer", "regularization", "sparse", "lasso", "regularized", "clustering", "tensors", "softmax", "minimization", "algorithms" ], [ "autoencoders", "imagenet", "benchmarks", "bottleneck", "cnn", "rnns", "cnns", "networks", "neural", "optimized" ], [ "imagenet", "generative", "cnn", "autoencoder", "autoencoders", "gans", "adversarially", "adversarial", "gan", "cnns" ], [ "lasso", "optimizer", "optimize", "optimization", "minimization", "minimize", "regularization", "optimizing", "optimized", "optimizes" ], [ "supervised", "learns", "priors", "autoencoders", "backpropagation", "probabilistic", "rnns", "softmax", "generative", "models" ], [ "imagenet", "cnn", "cnns", "recognition", "softmax", "supervised", "autoencoders", "neural", "segmentation", "recognizing" ], [ "recognizing", "supervised", "convolutions", "imagenet", "cnns", "recognition", "cnn", "autoencoders", "convolutional", "classifiers" ] ]
530.921954
all-MiniLM-L6-v2
0.480952
-0.2564
0.172465
0.845788
ArXiv ML Papers
Top2Vec
45
50
[ [ "learning", "learns", "learnt", "reinforcement", "bandit", "bandits", "ai", "planning", "learnable", "softmax" ], [ "softmax", "distributed", "learns", "federated", "adversarially", "learnt", "learning", "supervised", "rnns", "rnn" ], [ "adversary", "adversarial", "adversarially", "cnns", "classifiers", "softmax", "cnn", "exploiting", "attacks", "adversaries" ], [ "biases", "bias", "fairness", "adversarially", "classifiers", "adversarial", "supervised", "discrimination", "discriminate", "classifier" ], [ "rnns", "lstm", "learns", "autoencoders", "softmax", "corpus", "autoencoder", "supervised", "classifiers", "cnn" ], [ "softmax", "learning", "backpropagation", "autoencoders", "modeling", "rnns", "learns", "neural", "models", "rnn" ], [ "imagenet", "networks", "regularization", "neural", "neuron", "backpropagation", "cnns", "cnn", "softmax", "adversarial" ], [ "classifier", "classifiers", "classifying", "classification", "classify", "supervised", "softmax", "ensemble", "boosting", "ensembles" ], [ "predict", "forecasts", "forecasting", "forecast", "softmax", "predicts", "rnns", "prediction", "predicting", "lstm" ], [ "classifier", "classification", "lstm", "classifiers", "classifying", "softmax", "classify", "supervised", "recognition", "neural" ], [ "classifiers", "predicting", "softmax", "supervised", "datasets", "classifying", "classification", "classifier", "dataset", "classify" ], [ "networks", "nodes", "graphs", "graph", "cnn", "cnns", "embeddings", "softmax", "vertex", "supervised" ], [ "gan", "gans", "generative", "adversarial", "autoencoders", "adversarially", "cnn", "imagenet", "autoencoder", "cnns" ], [ "optimization", "minimization", "optimizer", "optimizing", "lasso", "minimize", "regularization", "optimal", "softmax", "optimize" ], [ "softmax", "autoencoders", "supervised", "models", "rnns", "generative", "backpropagation", "probabilistic", "priors", "learns" ], [ "autoencoders", "imagenet", "benchmarks", "bottleneck", "cnn", "rnns", "cnns", "networks", "neural", "optimized" ], [ "cnn", "imagenet", "autoencoders", "neural", "supervised", "softmax", "cnns", "recognition", "segmentation", "recognizing" ], [ "cnns", "recognition", "imagenet", "cnn", "convolutions", "recognizing", "supervised", "autoencoders", "softmax", "classifiers" ], [ "maps", "cnns", "cnn", "imagenet", "3d", "recognition", "recognizing", "convolutions", "supervised", "vision" ], [ "supervised", "metrics", "classifiers", "classifying", "softmax", "classification", "classifier", "regularization", "metric", "similarity" ], [ "regularized", "lasso", "sparse", "regularization", "minimization", "softmax", "tensors", "optimizer", "clustering", "tensor" ] ]
651.461712
all-MiniLM-L6-v2
0.442857
-0.252986
0.175005
0.843242
ArXiv ML Papers
Top2Vec
46
50
[ [ "learning", "ai", "planning", "reinforcement", "bandit", "learnt", "learns", "bandits", "learnable", "softmax" ], [ "adversarially", "distributed", "softmax", "federated", "learning", "learns", "learnt", "supervised", "rnns", "rnn" ], [ "adversary", "adversarial", "adversarially", "cnns", "adversaries", "cnn", "classifiers", "attacks", "softmax", "attacker" ], [ "regularization", "adversarially", "adversarial", "privacy", "private", "randomized", "adversary", "normalization", "regularized", "softmax" ], [ "fairness", "adversarially", "bias", "biases", "classifiers", "discrimination", "adversarial", "discriminate", "supervised", "classifier" ], [ "softmax", "autoencoder", "lstm", "supervised", "rnns", "embeddings", "learns", "corpus", "autoencoders", "cnn" ], [ "learns", "rnns", "neural", "backpropagation", "learning", "softmax", "autoencoders", "models", "rnn", "modeling" ], [ "forecasts", "forecasting", "forecast", "prediction", "predict", "predicting", "predicts", "softmax", "classifiers", "rnns" ], [ "softmax", "networks", "neural", "backpropagation", "cnns", "imagenet", "regularization", "cnn", "neuron", "rnns" ], [ "classify", "classifying", "eeg", "supervised", "classifiers", "classifier", "classification", "recognition", "lstm", "svm" ], [ "networks", "graphs", "nodes", "graph", "cnn", "softmax", "cnns", "supervised", "vertex", "embeddings" ], [ "dataset", "classifying", "classifiers", "predicting", "datasets", "softmax", "supervised", "classification", "classify", "rnns" ], [ "classifiers", "ensemble", "softmax", "boosting", "supervised", "classification", "classifying", "classify", "classifier", "ensembles" ], [ "imagenet", "cnns", "cnn", "supervised", "softmax", "neural", "recognition", "autoencoders", "segmentation", "recognizing" ], [ "regularization", "classifiers", "metrics", "classifier", "classifying", "classification", "softmax", "supervised", "similarity", "metric" ], [ "sparse", "softmax", "clustering", "minimization", "optimizer", "lasso", "algorithms", "regularized", "regularization", "supervised" ], [ "gan", "autoencoders", "adversarially", "adversarial", "autoencoder", "generative", "imagenet", "gans", "cnns", "cnn" ], [ "cnns", "rnns", "cnn", "imagenet", "autoencoders", "bottleneck", "benchmarks", "networks", "softmax", "neural" ], [ "imagenet", "recognizing", "recognition", "convolutions", "cnns", "cnn", "supervised", "autoencoders", "convolutional", "softmax" ], [ "optimization", "minimization", "optimizer", "optimizing", "optimize", "minimize", "regularization", "lasso", "optimized", "optimizes" ], [ "autoencoders", "learns", "rnns", "generative", "softmax", "models", "supervised", "backpropagation", "priors", "autoencoder" ] ]
640.067949
all-MiniLM-L6-v2
0.452381
-0.262968
0.171184
0.850129
ArXiv ML Papers
GMM
43
10
[ [ "kernel", "clustering", "matrix", "rank", "series", "time", "regression", "proposed", "tensor", "algorithm" ], [ "gnns", "link", "node", "embedding", "graphs", "graph", "nodes", "structure", "networks", "social" ], [ "reward", "rl", "reinforcement", "policy", "agent", "agents", "policies", "games", "action", "control" ], [ "speech", "text", "language", "languages", "attention", "word", "words", "sequence", "task", "natural" ], [ "images", "gan", "gans", "image", "generative", "latent", "generation", "generator", "variational", "vae" ], [ "convergence", "bounds", "gradient", "stochastic", "convex", "optimization", "algorithm", "bound", "complexity", "sample" ], [ "fairness", "ml", "label", "classifier", "machine", "classification", "decision", "prediction", "their", "research" ], [ "networks", "deep", "memory", "architectures", "layer", "systems", "quantum", "neural", "network", "time" ], [ "segmentation", "images", "object", "cnn", "convolutional", "image", "deep", "detection", "video", "visual" ], [ "perturbations", "against", "robustness", "robust", "attack", "attacks", "defense", "adversarial", "privacy", "examples" ] ]
2.576472
all-MiniLM-L6-v2
0.94
0.053246
0.1459
0.813589
ArXiv ML Papers
GMM
44
10
[ [ "speech", "recognition", "speaker", "acoustic", "asr", "audio", "music", "signal", "end", "quality" ], [ "posterior", "variational", "inference", "estimation", "bayesian", "distribution", "latent", "gaussian", "series", "processes" ], [ "classifier", "classification", "machine", "label", "prediction", "metric", "accuracy", "datasets", "ml", "features" ], [ "reinforcement", "reward", "agent", "agents", "rl", "action", "control", "policy", "games", "policies" ], [ "attack", "defense", "perturbations", "privacy", "attacks", "adversarial", "robustness", "against", "robust", "examples" ], [ "graphs", "node", "graph", "nodes", "embedding", "link", "structure", "gnns", "embeddings", "representation" ], [ "algorithm", "convex", "convergence", "regret", "bound", "stochastic", "bounds", "algorithms", "optimization", "gradient" ], [ "memory", "layer", "network", "networks", "neural", "accuracy", "architectures", "deep", "architecture", "input" ], [ "image", "segmentation", "object", "visual", "images", "gan", "convolutional", "dataset", "deep", "detection" ], [ "sequence", "sentences", "word", "natural", "language", "text", "task", "words", "attention", "semantic" ] ]
2.836519
all-MiniLM-L6-v2
0.98
0.06128
0.149774
0.825734
ArXiv ML Papers
GMM
45
10
[ [ "detection", "segmentation", "visual", "gan", "image", "images", "object", "convolutional", "supervised", "medical" ], [ "examples", "against", "adversarial", "defense", "perturbations", "attack", "attacks", "privacy", "robust", "robustness" ], [ "agent", "reinforcement", "reward", "rl", "policy", "agents", "regret", "action", "policies", "games" ], [ "text", "speech", "attention", "audio", "language", "languages", "speaker", "word", "translation", "recurrent" ], [ "quantum", "posterior", "parameters", "variational", "bayesian", "physics", "uncertainty", "dynamics", "inference", "equations" ], [ "series", "classifier", "prediction", "classification", "machine", "time", "accuracy", "selection", "label", "regression" ], [ "convergence", "algorithm", "convex", "matrix", "optimization", "rank", "algorithms", "linear", "problems", "gradient" ], [ "neural", "layer", "network", "networks", "deep", "architectures", "accuracy", "memory", "hardware", "layers" ], [ "user", "items", "language", "recommendation", "item", "knowledge", "label", "recommender", "topic", "metric" ], [ "nodes", "graph", "graphs", "node", "link", "gnns", "embedding", "structure", "embeddings", "representation" ] ]
2.766196
all-MiniLM-L6-v2
0.97
0.024243
0.165378
0.838997
ArXiv ML Papers
GMM
46
10
[ [ "convex", "matrix", "algorithm", "convergence", "algorithms", "linear", "kernel", "problems", "rank", "optimization" ], [ "nodes", "graph", "link", "node", "embedding", "gnns", "graphs", "networks", "structure", "embeddings" ], [ "regret", "reinforcement", "agent", "agents", "control", "reward", "policies", "rl", "policy", "games" ], [ "variational", "uncertainty", "distribution", "parameters", "dynamics", "bayesian", "neural", "posterior", "latent", "estimation" ], [ "language", "natural", "word", "text", "sentences", "sequence", "words", "task", "representations", "semantic" ], [ "object", "images", "segmentation", "image", "visual", "medical", "detection", "gan", "supervised", "convolutional" ], [ "attack", "robustness", "attacks", "adversarial", "privacy", "against", "perturbations", "defense", "examples", "robust" ], [ "audio", "speaker", "recognition", "acoustic", "speech", "asr", "music", "signals", "signal", "end" ], [ "series", "machine", "time", "classification", "classifier", "prediction", "label", "fairness", "ml", "regression" ], [ "neural", "memory", "network", "hardware", "architectures", "accuracy", "networks", "deep", "layer", "energy" ] ]
2.882266
all-MiniLM-L6-v2
0.98
0.054135
0.155939
0.826401
ArXiv ML Papers
GMM
43
20
[ [ "algorithm", "rank", "tensor", "matrix", "low", "kernel", "clustering", "sparse", "linear", "subspace" ], [ "graph", "link", "embedding", "node", "gnns", "nodes", "graphs", "structure", "embeddings", "representation" ], [ "bound", "online", "regret", "bandits", "bandit", "arm", "sqrt", "reward", "agents", "algorithm" ], [ "user", "recommendation", "language", "text", "word", "item", "recommender", "users", "items", "words" ], [ "image", "generator", "gans", "images", "generative", "gan", "latent", "vae", "variational", "generation" ], [ "convex", "stochastic", "gradient", "optimization", "convergence", "descent", "sgd", "optimal", "function", "problems" ], [ "communication", "devices", "iot", "distributed", "traffic", "vehicle", "wireless", "federated", "computing", "server" ], [ "equations", "physics", "quantum", "neural", "forecasting", "dynamics", "series", "systems", "time", "parameters" ], [ "segmentation", "medical", "cancer", "image", "imaging", "ct", "images", "patients", "covid", "net" ], [ "fair", "federated", "private", "privacy", "user", "fairness", "sensitive", "differential", "local", "differentially" ], [ "series", "ml", "day", "software", "time", "machine", "classification", "study", "health", "research" ], [ "images", "detection", "scene", "video", "cnn", "object", "3d", "image", "visual", "objects" ], [ "self", "features", "deep", "supervised", "feature", "domain", "classification", "dataset", "explanations", "attention" ], [ "sequence", "languages", "translation", "recurrent", "text", "language", "attention", "task", "sentences", "word" ], [ "classification", "metric", "bias", "class", "label", "prediction", "labels", "classifier", "machine", "generalization" ], [ "speech", "asr", "music", "speaker", "acoustic", "audio", "recognition", "signal", "enhancement", "quality" ], [ "distributions", "inference", "posterior", "bayesian", "distribution", "variational", "gaussian", "estimation", "processes", "probabilistic" ], [ "pruning", "hardware", "architectures", "networks", "nas", "accuracy", "neural", "memory", "architecture", "layers" ], [ "robustness", "attacks", "adversarial", "attack", "perturbations", "against", "examples", "defense", "robust", "attacker" ], [ "environment", "reward", "control", "agent", "agents", "action", "policy", "rl", "reinforcement", "policies" ] ]
4.033339
all-MiniLM-L6-v2
0.9
0.016575
0.153815
0.857305
README.md exists but content is empty. Use the Edit dataset card button to edit it.
Downloads last month
75
Edit dataset card