image
imagewidth (px)
28
28
label
class label
10 classes
text_label
stringclasses
874 values
p_label
sequence
is_ambiguous
bool
1 class
77
p(3)=0.41, p(7)=0.59
[ 0, 0, 0, 0.4070048, 0, 0, 0, 0.59299517, 0, 0 ]
true
99
p(1)=0.43, p(9)=0.57
[ 0, 0.42807725, 0, 0, 0, 0, 0, 0, 0, 0.5719228 ]
true
99
p(0)=0.38, p(9)=0.62
[ 0.37826502, 0, 0, 0, 0, 0, 0, 0, 0, 0.62173504 ]
true
11
p(1)=0.46, p(6)=0.54
[ 0, 0.45579028, 0, 0, 0, 0, 0.5442097, 0, 0, 0 ]
true
77
p(6)=0.43, p(7)=0.57
[ 0, 0, 0, 0, 0, 0, 0.43084446, 0.5691556, 0, 0 ]
true
11
p(1)=0.50, p(6)=0.50
[ 0, 0.50238645, 0, 0, 0, 0, 0.4976135, 0, 0, 0 ]
true
22
p(0)=0.48, p(2)=0.52
[ 0.47757974, 0, 0.5224203, 0, 0, 0, 0, 0, 0, 0 ]
true
33
p(3)=0.48, p(4)=0.52
[ 0, 0, 0, 0.4796488, 0.52035123, 0, 0, 0, 0, 0 ]
true
0no label
p(0)=0.62, p(2)=0.38
[ 0.6161282, 0, 0.38387182, 0, 0, 0, 0, 0, 0, 0 ]
true
88
p(0)=0.50, p(8)=0.50
[ 0.50370896, 0, 0, 0, 0, 0, 0, 0, 0.49629098, 0 ]
true
88
p(3)=0.45, p(8)=0.55
[ 0, 0, 0, 0.44907367, 0, 0, 0, 0, 0.5509263, 0 ]
true
33
p(1)=0.42, p(3)=0.58
[ 0, 0.41939825, 0, 0.5806018, 0, 0, 0, 0, 0, 0 ]
true
0no label
p(0)=0.51, p(8)=0.49
[ 0.51421547, 0, 0, 0, 0, 0, 0, 0, 0.48578456, 0 ]
true
99
p(2)=0.50, p(9)=0.50
[ 0, 0, 0.4993801, 0, 0, 0, 0, 0, 0, 0.5006198 ]
true
88
p(4)=0.49, p(8)=0.51
[ 0, 0, 0, 0, 0.49415827, 0, 0, 0, 0.50584173, 0 ]
true
11
p(1)=0.38, p(3)=0.62
[ 0, 0.38217163, 0, 0.61782837, 0, 0, 0, 0, 0, 0 ]
true
77
p(3)=0.59, p(7)=0.41
[ 0, 0, 0, 0.5880495, 0, 0, 0, 0.4119505, 0, 0 ]
true
22
p(1)=0.46, p(2)=0.54
[ 0, 0.45873684, 0.54126316, 0, 0, 0, 0, 0, 0, 0 ]
true
88
p(2)=0.62, p(8)=0.38
[ 0, 0, 0.6156755, 0, 0, 0, 0, 0, 0.38432446, 0 ]
true
88
p(1)=0.40, p(8)=0.60
[ 0, 0.40358528, 0, 0, 0, 0, 0, 0, 0.59641474, 0 ]
true
44
p(4)=0.59, p(8)=0.41
[ 0, 0, 0, 0, 0.5888328, 0, 0, 0, 0.4111672, 0 ]
true
77
p(2)=0.41, p(7)=0.59
[ 0, 0, 0.41265386, 0, 0, 0, 0, 0.58734614, 0, 0 ]
true
77
p(0)=0.42, p(7)=0.58
[ 0.4236934, 0, 0, 0, 0, 0, 0, 0.57630664, 0, 0 ]
true
99
p(1)=0.39, p(9)=0.61
[ 0, 0.3867687, 0, 0, 0, 0, 0, 0, 0, 0.6132313 ]
true
55
p(5)=0.59, p(7)=0.41
[ 0, 0, 0, 0, 0, 0.58657503, 0, 0.413425, 0, 0 ]
true
22
p(2)=0.43, p(5)=0.57
[ 0, 0, 0.42855716, 0, 0, 0.57144284, 0, 0, 0, 0 ]
true
66
p(0)=0.38, p(6)=0.62
[ 0.38483214, 0, 0, 0, 0, 0, 0.61516786, 0, 0, 0 ]
true
99
p(5)=0.54, p(9)=0.46
[ 0, 0, 0, 0, 0, 0.54087263, 0, 0, 0, 0.45912743 ]
true
55
p(4)=0.51, p(5)=0.49
[ 0, 0, 0, 0, 0.51396924, 0.48603076, 0, 0, 0, 0 ]
true
88
p(4)=0.60, p(8)=0.40
[ 0, 0, 0, 0, 0.59648204, 0, 0, 0, 0.40351793, 0 ]
true
0no label
p(0)=0.51, p(1)=0.49
[ 0.5095867, 0.49041328, 0, 0, 0, 0, 0, 0, 0, 0 ]
true
77
p(2)=0.60, p(7)=0.40
[ 0, 0, 0.597877, 0, 0, 0, 0, 0.40212294, 0, 0 ]
true
11
p(1)=0.44, p(6)=0.56
[ 0, 0.4445085, 0, 0, 0, 0, 0.5554915, 0, 0, 0 ]
true
11
p(0)=0.43, p(1)=0.57
[ 0.42579526, 0.5742047, 0, 0, 0, 0, 0, 0, 0, 0 ]
true
22
p(0)=0.53, p(2)=0.47
[ 0.5329073, 0, 0.46709272, 0, 0, 0, 0, 0, 0, 0 ]
true
11
p(1)=0.49, p(2)=0.51
[ 0, 0.48685473, 0.51314527, 0, 0, 0, 0, 0, 0, 0 ]
true
88
p(0)=0.49, p(8)=0.51
[ 0.48576495, 0, 0, 0, 0, 0, 0, 0, 0.514235, 0 ]
true
22
p(1)=0.38, p(2)=0.62
[ 0, 0.3802728, 0.61972713, 0, 0, 0, 0, 0, 0, 0 ]
true
99
p(1)=0.45, p(9)=0.55
[ 0, 0.45123187, 0, 0, 0, 0, 0, 0, 0, 0.54876816 ]
true
55
p(2)=0.55, p(5)=0.45
[ 0, 0, 0.5471103, 0, 0, 0.45288965, 0, 0, 0, 0 ]
true
11
p(1)=0.42, p(4)=0.58
[ 0, 0.42055926, 0, 0, 0.5794407, 0, 0, 0, 0, 0 ]
true
22
p(2)=0.60, p(3)=0.40
[ 0, 0, 0.60222054, 0.3977795, 0, 0, 0, 0, 0, 0 ]
true
99
p(1)=0.47, p(9)=0.53
[ 0, 0.46597242, 0, 0, 0, 0, 0, 0, 0, 0.5340275 ]
true
33
p(3)=0.39, p(9)=0.61
[ 0, 0, 0, 0.38931885, 0, 0, 0, 0, 0, 0.6106811 ]
true
66
p(6)=0.49, p(7)=0.51
[ 0, 0, 0, 0, 0, 0, 0.4920458, 0.50795424, 0, 0 ]
true
77
p(2)=0.50, p(7)=0.50
[ 0, 0, 0.49990892, 0, 0, 0, 0, 0.500091, 0, 0 ]
true
55
p(5)=0.61, p(7)=0.39
[ 0, 0, 0, 0, 0, 0.60824573, 0, 0.3917543, 0, 0 ]
true
33
p(1)=0.40, p(3)=0.60
[ 0, 0.3996877, 0, 0.6003123, 0, 0, 0, 0, 0, 0 ]
true
22
p(2)=0.45, p(8)=0.55
[ 0, 0, 0.45297113, 0, 0, 0, 0, 0, 0.5470288, 0 ]
true
88
p(4)=0.55, p(8)=0.45
[ 0, 0, 0, 0, 0.55043864, 0, 0, 0, 0.44956142, 0 ]
true
99
p(5)=0.53, p(9)=0.47
[ 0, 0, 0, 0, 0, 0.5299525, 0, 0, 0, 0.47004747 ]
true
22
p(2)=0.58, p(4)=0.42
[ 0, 0, 0.575402, 0, 0.42459795, 0, 0, 0, 0, 0 ]
true
11
p(1)=0.44, p(6)=0.56
[ 0, 0.4367434, 0, 0, 0, 0, 0.5632566, 0, 0, 0 ]
true
55
p(5)=0.44, p(6)=0.56
[ 0, 0, 0, 0, 0, 0.443231, 0.556769, 0, 0, 0 ]
true
77
p(5)=0.41, p(7)=0.59
[ 0, 0, 0, 0, 0, 0.41450867, 0, 0.5854914, 0, 0 ]
true
77
p(5)=0.58, p(7)=0.42
[ 0, 0, 0, 0, 0, 0.576222, 0, 0.42377797, 0, 0 ]
true
22
p(2)=0.57, p(7)=0.43
[ 0, 0, 0.5702413, 0, 0, 0, 0, 0.42975876, 0, 0 ]
true
99
p(0)=0.44, p(9)=0.56
[ 0.4422371, 0, 0, 0, 0, 0, 0, 0, 0, 0.5577629 ]
true
11
p(0)=0.46, p(1)=0.54
[ 0.46397096, 0.53602904, 0, 0, 0, 0, 0, 0, 0, 0 ]
true
88
p(0)=0.41, p(8)=0.59
[ 0.41236994, 0, 0, 0, 0, 0, 0, 0, 0.58763003, 0 ]
true
33
p(1)=0.42, p(3)=0.58
[ 0, 0.42286626, 0, 0.5771338, 0, 0, 0, 0, 0, 0 ]
true
66
p(4)=0.40, p(6)=0.60
[ 0, 0, 0, 0, 0.40103728, 0, 0.5989627, 0, 0, 0 ]
true
88
p(0)=0.54, p(8)=0.46
[ 0.53830785, 0, 0, 0, 0, 0, 0, 0, 0.46169215, 0 ]
true
0no label
p(0)=0.62, p(8)=0.38
[ 0.6168924, 0, 0, 0, 0, 0, 0, 0, 0.3831076, 0 ]
true
88
p(8)=0.60, p(9)=0.40
[ 0, 0, 0, 0, 0, 0, 0, 0, 0.60115737, 0.39884257 ]
true
33
p(1)=0.40, p(3)=0.60
[ 0, 0.3990372, 0, 0.60096276, 0, 0, 0, 0, 0, 0 ]
true
44
p(0)=0.49, p(4)=0.51
[ 0.4927888, 0, 0, 0, 0.5072112, 0, 0, 0, 0, 0 ]
true
44
p(4)=0.41, p(6)=0.59
[ 0, 0, 0, 0, 0.4148215, 0, 0.58517843, 0, 0, 0 ]
true
77
p(2)=0.47, p(7)=0.53
[ 0, 0, 0.46618918, 0, 0, 0, 0, 0.5338108, 0, 0 ]
true
55
p(5)=0.53, p(7)=0.47
[ 0, 0, 0, 0, 0, 0.5300644, 0, 0.46993563, 0, 0 ]
true
0no label
p(0)=0.50, p(1)=0.50
[ 0.49987543, 0.50012463, 0, 0, 0, 0, 0, 0, 0, 0 ]
true
44
p(4)=0.43, p(6)=0.57
[ 0, 0, 0, 0, 0.42906788, 0, 0.5709321, 0, 0, 0 ]
true
66
p(2)=0.49, p(6)=0.51
[ 0, 0, 0.49284253, 0, 0, 0, 0.5071575, 0, 0, 0 ]
true
11
p(0)=0.53, p(1)=0.47
[ 0.5300049, 0.46999514, 0, 0, 0, 0, 0, 0, 0, 0 ]
true
22
p(0)=0.47, p(2)=0.53
[ 0.46749356, 0, 0.53250647, 0, 0, 0, 0, 0, 0, 0 ]
true
0no label
p(0)=0.47, p(2)=0.53
[ 0.47282144, 0, 0.5271785, 0, 0, 0, 0, 0, 0, 0 ]
true
11
p(0)=0.49, p(1)=0.51
[ 0.49483973, 0.5051602, 0, 0, 0, 0, 0, 0, 0, 0 ]
true
77
p(6)=0.42, p(7)=0.58
[ 0, 0, 0, 0, 0, 0, 0.41634953, 0.5836505, 0, 0 ]
true
55
p(1)=0.54, p(5)=0.46
[ 0, 0.53951657, 0, 0, 0, 0.46048334, 0, 0, 0, 0 ]
true
22
p(1)=0.50, p(2)=0.50
[ 0, 0.49525434, 0.5047457, 0, 0, 0, 0, 0, 0, 0 ]
true
66
p(1)=0.38, p(6)=0.62
[ 0, 0.38075843, 0, 0, 0, 0, 0.6192416, 0, 0, 0 ]
true
88
p(0)=0.41, p(8)=0.59
[ 0.4072794, 0, 0, 0, 0, 0, 0, 0, 0.59272057, 0 ]
true
44
p(0)=0.38, p(4)=0.62
[ 0.38404495, 0, 0, 0, 0.61595505, 0, 0, 0, 0, 0 ]
true
55
p(5)=0.43, p(7)=0.57
[ 0, 0, 0, 0, 0, 0.42985925, 0, 0.5701407, 0, 0 ]
true
99
p(2)=0.49, p(9)=0.51
[ 0, 0, 0.4920304, 0, 0, 0, 0, 0, 0, 0.5079696 ]
true
22
p(2)=0.49, p(7)=0.51
[ 0, 0, 0.49241677, 0, 0, 0, 0, 0.5075832, 0, 0 ]
true
66
p(1)=0.39, p(6)=0.61
[ 0, 0.38711947, 0, 0, 0, 0, 0.6128805, 0, 0, 0 ]
true
99
p(2)=0.62, p(9)=0.38
[ 0, 0, 0.6197099, 0, 0, 0, 0, 0, 0, 0.38029012 ]
true
77
p(2)=0.62, p(7)=0.38
[ 0, 0, 0.6219566, 0, 0, 0, 0, 0.37804344, 0, 0 ]
true
77
p(3)=0.55, p(7)=0.45
[ 0, 0, 0, 0.55263036, 0, 0, 0, 0.4473696, 0, 0 ]
true
44
p(3)=0.54, p(4)=0.46
[ 0, 0, 0, 0.535004, 0.464996, 0, 0, 0, 0, 0 ]
true
66
p(2)=0.38, p(6)=0.62
[ 0, 0, 0.38459733, 0, 0, 0, 0.6154027, 0, 0, 0 ]
true
44
p(0)=0.43, p(4)=0.57
[ 0.42750177, 0, 0, 0, 0.5724982, 0, 0, 0, 0, 0 ]
true
66
p(0)=0.46, p(6)=0.54
[ 0.46319106, 0, 0, 0, 0, 0, 0.536809, 0, 0, 0 ]
true
55
p(2)=0.62, p(5)=0.38
[ 0, 0, 0.6155121, 0, 0, 0.3844879, 0, 0, 0, 0 ]
true
88
p(3)=0.39, p(8)=0.61
[ 0, 0, 0, 0.38902405, 0, 0, 0, 0, 0.6109759, 0 ]
true
99
p(6)=0.43, p(9)=0.57
[ 0, 0, 0, 0, 0, 0, 0.42756817, 0, 0, 0.5724318 ]
true
11
p(1)=0.41, p(6)=0.59
[ 0, 0.40667504, 0, 0, 0, 0, 0.59332496, 0, 0, 0 ]
true
66
p(2)=0.44, p(6)=0.56
[ 0, 0, 0.4360127, 0, 0, 0, 0.56398726, 0, 0, 0 ]
true
55
p(5)=0.57, p(6)=0.43
[ 0, 0, 0, 0, 0, 0.5684987, 0.43150142, 0, 0, 0 ]
true

Mnist-Ambiguous

This dataset contains mnist-like images, but with an unclear ground truth. For each image, there are two classes which could be considered true. Robust and uncertainty-aware DNNs should thus detect and flag these issues.

Features

Same as mnist, the supervised dataset has an image (28x28 int array) and a label (int).

Additionally, the following features are exposed for your convenience:

  • text_label (str): A textual representation of the probabilistic label, e.g. p(0)=0.54, p(5)=0.46
  • p_label (list of floats): Ground-Truth probabilities for each class (two nonzero values for our ambiguous images)
  • is_ambiguous (bool): Flag indicating if this is one of our ambiguous images (see 'splits' below)

Splits

We provide four splits:

  • test: 10'000 ambiguous images
  • train: 10'000 ambiguous images - adding ambiguous images to the training set makes sure test-time ambiguous images are in-distribution.
  • test_mixed: 20'000 images, consisting of the (shuffled) concatenation of our ambiguous test set and the nominal mnist test set by LeCun et. al.,
  • train_mixed: 70'000 images, consisting of the (shuffled) concatenation of our ambiguous training and the nominal training set.

Note that the ambiguous test images are highly ambiguous (i.e., the two classes have very similar ground truth likelihoods), the training set images allow for more unbalanced ambiguity. This is to make the training set more closely connected to the nominal data, while still keeping the test set clearly ambiguous.

For research targeting explicitly aleatoric uncertainty, we recommend training the model using train_mixed. Otherwise, our test set will lead to both epistemic and aleatoric uncertainty. In related literature, such 'mixed' splits are sometimes denoted as dirty splits.

Assessment and Validity

For a brief discussion of the strength and weaknesses of this dataset, including a quantitative comparison to the (only) other ambiguous datasets available in the literature, we refer to our paper.

Paper

Pre-print here: https://arxiv.org/abs/2207.10495

Citation:

@misc{https://doi.org/10.48550/arxiv.2207.10495,
  doi = {10.48550/ARXIV.2207.10495},
  url = {https://arxiv.org/abs/2207.10495},
  author = {Weiss, Michael and Gómez, André García and Tonella, Paolo},
  title = {A Forgotten Danger in DNN Supervision Testing: Generating and Detecting True Ambiguity},
  publisher = {arXiv},
  year = {2022}
}

License

As this is a derivative work of mnist, which is CC-BY-SA 3.0 licensed, our dataset is released using the same license.

Downloads last month
212
Edit dataset card