Neuroscience Decoding Encoding Models
AI & ML interests
Neuroscience, AI, maths
Recent Activity
The NDEM community provides pretrained models along with their checkpoint with the purpose of:
- Studying the learning dynamics of the models
- Studying how well these learning dynamics match brain learning dynamics
Models are pretrained on the Jean-Zay state-owned supercluster: This work was granted access to the HPC resources of IDRIS under the allocations 2023-AD011014524 and 2022-AD011013176R1 made by GENCI (P.Orhan).
Models currently available are:
- Wav2vec2 base model (https://huggingface.co/facebook/wav2vec2-base), but pretrained (no fine-tuning) on Librispeech (English speech), FMA (music), subset of audioset, or all of them together. It also includes a model pretrained on VoxPopuli french dataset.
- Wav2vec2 tiny model, where we used only 3 transformer layers. Models' performances are surprisingly high.
Scientific papers using the models provided in this repository: Orhan, P., Boubenec, Y., & King, J.-R. (2024). Algebraic structures emerge from the self-supervised learning of natural sounds. https://doi.org/10.1101/2024.03.13.584776
Models are pretrained using HuggingFace's trainer. These models pretraining are often shorter (100,000 steps compared to 400 000) than original pretraining because of resource scarcity. In my experience, most emergences I studied had happened before 100 000 steps.
Known version compatibility issues for Wav2vec2: Some Wav2vec2 models are trained with a torch <=2.0.1, while other are trained with torch>2.1.1 This can create critical error when loading the model, at the step of loading the Wav2Vec2PositionalConvEmbedding, which uses different module for the weight_norm depending on the pretraining version. This is quite unfortunate. Consequently I recommend checking for the error message of HuggingFace with from_pretrained(), and changing your torch version if the weights can't be properly loaded.
Models trained with 2.0.1 and prior version:
outputs_XXX_short_X
outputs_XXX_tiny_X
Models trained with 2.1.1 and following version:
outputs_XXX_large_X
outputs_XXX_mini_X