Generalizability Under Sensor Failure: Tokenization + Transformers Enable More Robust Latent Spaces
This addresses robustness to sensor failure for neuroscience experiments, which is an incremental improvement over prior work focusing on sessions and subjects.
The paper tackled the problem of neural data generalization under sensor failure, a prevalent issue in neuroscience, and found that the TOTEM tokenizer-transformer model outperformed or matched the widely used EEGNet across all generalizability cases, including sessions, subjects, and sensors.
A major goal in neuroscience is to discover neural data representations that generalize. This goal is challenged by variability along recording sessions (e.g. environment), subjects (e.g. varying neural structures), and sensors (e.g. sensor noise), among others. Recent work has begun to address generalization across sessions and subjects, but few study robustness to sensor failure which is highly prevalent in neuroscience experiments. In order to address these generalizability dimensions we first collect our own electroencephalography dataset with numerous sessions, subjects, and sensors, then study two time series models: EEGNet (Lawhern et al., 2018) and TOTEM (Talukder et al., 2024). EEGNet is a widely used convolutional neural network, while TOTEM is a discrete time series tokenizer and transformer model. We find that TOTEM outperforms or matches EEGNet across all generalizability cases. Finally through analysis of TOTEM's latent codebook we observe that tokenization enables generalization.