Core Concepts
Tokenization and transformers outperform traditional CNN models in creating more generalizable latent spaces for neural data representations.
Abstract
The content discusses the challenges faced in neuroscience regarding generalizability across sessions, subjects, and sensor failure. It introduces two models, EEGNet and TOTEM, comparing their performance in various scenarios. TOTEM's tokenization approach proves to be more effective in creating generalizable representations. The study also delves into the analysis of TOTEM's latent codebook to understand its capabilities further.
Stats
We collect four 128-channel EEG sessions: A1, A2, B1, B2, each with 600 trials.
For each session our subject sat in front of a monitor and was instructed to fixate on a center point that randomly changed to ◀, ▶, ▲, and ▼.
We recorded 600 trials per session with most datasets having fewer than a couple hundred trials per session.
We simulate sensor failure by randomly zeroing-out X% of test set sensors where X∈ {0, 10, 20,...100}.
Hyperparameters were selected for optimal performance on within-session data and kept consistent across all modeling experiments.
Quotes
"We find that tokenization + transformers are a promising approach to modeling time series neural data." - Sabera Talukder
"TOTEM's tokenization + transformers outperform in numerous generalization cases." - Yisong Yue
"These models are also ripe for interpretability analysis which can uncover new findings about time series neural data." - Bingni W. Brunton