Skip to content

Deep Learning Decoders

After 2018, deep learning took over BCI decoding across the board. This chapter presents the transition in chronological order: from task-specific CNNs like EEGNet, to dynamical modeling such as LFADS, to neural foundation models like NDT / POYO / CEBRA. The endpoint of this trajectory is a large model pre-trained on massive neural data that can few-shot transfer to new subjects and new tasks.

Relationship to Chapter 04. Classical decoding (Chapter 04) solves "precise decoding for one subject, one task." Deep-learning decoders solve "generalization across subjects, tasks, and electrode arrays." This is the biggest paradigm shift in BCI over the past five years: where every new patient used to require months of recalibration, foundation models such as CEBRA / POYO now show meaningful zero-shot or few-shot transfer between monkey and human, between different Utah arrays, and across spike / LFP / ECoG modalities. If this path matures, BCI gains the same "pre-train once, deploy everywhere" engineering footing that NLP gained after GPT-2.

Recommended reading order. Read in three stages — task-specific → dynamics → cross-domain foundation models: start with EEGNet and CNN Methods to enter the design space of compact CNNs on EEG BCI; then move to LFADS and Dynamics Modeling to understand the pivotal turn in Pandarinath 2018, which brought latent dynamics into decoding; next, NDT Series and Transformer tracks the evolution NDT1 → 2 → 3; CEBRA and Contrastive Learning showcases the elegance of behavior-aligned embedding; and finally, Neural Foundation Model POYO closes the chapter by surveying how POYO+ / Neuroformer / BrainBERT converge toward a "GPT for neural data."

In this chapter:

评论 #