Large Cognition Model (LCM)

EEG is cheap, non-invasive, and abundant, yet EEG models rarely transfer. Recordings have a low signal-to-noise ratio, vary enormously between subjects, and — most awkwardly for pretraining — come from incompatible montages, with different electrode counts and placements across every lab and headset. A model trained on one dataset usually cannot even ingest another without discarding channels.

The Large Cognition Model (LCM), from Neuro Industry and the Georgia Institute of Technology, attacks the montage problem directly. Before any transformer sees the signal, a learnable channel mapping projects an arbitrary electrode configuration into a common channel space, and a learnable channel embedding preserves electrode-specific identity. Only then is the recording cut into spatio-temporal patches and encoded. This makes it possible to pretrain across datasets recorded on different hardware for different paradigms — motor imagery, steady-state visual evoked potentials, and emotion — and to fine-tune the result on downstream BCI tasks.

Pretraining combines two self-supervised objectives borrowed from vision: a momentum contrastive alignment between an online encoder and an EMA-updated target encoder, and a masked reconstruction loss over hidden patches. LCM belongs to the same wave of EEG foundation models as LaBraM, and is evaluated against it directly.

#Key Features

• Cross-montage channel mapping: A trainable transformation matrix plus a learnable channel embedding align differing electrode layouts into one latent space, so datasets from different rigs can be pooled for pretraining.
• Dual self-supervised objectives: A layer-normalized MSE alignment loss between the online and momentum encoders is combined with a masked patch reconstruction loss, so representations must be both stable across views and complete enough to inpaint.
• Spatio-temporal patching: The signal is tokenized into patches spanning both time and electrodes, letting attention integrate across the two axes rather than treating channels independently.
• Strong architecture prior: The encoder is competitive on downstream benchmarks even when trained from scratch without any pretraining phase.

#Technical Details

LCM is a 33.9M-parameter transformer encoder. The online encoder is updated by gradient descent while the target encoder follows an exponential moving average with momentum in [0.996, 1.0]; the total loss is the alignment term plus a weighted masked reconstruction term. Pretraining uses three public corpora: PhysioMI (109 subjects, motor imagery and execution), TSU (35 subjects, SSVEP), and SEED (15 subjects, emotion). All recordings are segmented into four-second windows, average re-referenced, scaled, and resampled to 256 Hz. Training ran 200 epochs at batch size 1,024 with AdamW, a peak learning rate of 1.5e-4, a 10-epoch warmup, and cosine annealing; downstream fine-tuning ran 100 epochs.

Evaluation covers two motor imagery benchmarks. On BCIC-2A (10 subjects, 4 classes), pretrained LCM reaches 0.6166 balanced accuracy, 0.4619 Cohen's kappa, and 0.5932 weighted F1, against 0.5846 / 0.4462 / 0.5715 for EEGPT and 0.5613 / 0.4151 / 0.5520 for LaBraM. On BCIC-2B (10 subjects, 2 classes) it reaches 0.7523 balanced accuracy and 0.8244 AUROC versus 0.7212 and 0.8059 for EEGPT. Without pretraining, LCM still scores 0.5263 on BCIC-2A, ahead of the pretrained BIOT and BENDR baselines.

#Applications

LCM targets brain-computer interface and clinical EEG settings where labelled data is scarce and hardware is heterogeneous: motor imagery decoding for assistive control, cognitive state monitoring, neurofeedback, and disease classification from clinical recordings. The channel-mapping design is the practically useful part — a lab with a 32-channel cap can start from weights pretrained on 64-channel data without rewriting its pipeline, which is the usual barrier to reusing published EEG models.

#Impact

LCM's contribution is the demonstration that a learnable montage alignment, rather than channel-wise tokenization or fixed electrode subsets, is enough to pool heterogeneous EEG corpora for self-supervised pretraining. Its reported gains over EEGPT and LaBraM are modest but consistent across metrics on both benchmarks. Two caveats matter for anyone planning to build on it. The evaluation is narrow: only two motor imagery datasets, both with 10 subjects, back the broader claims about cognitive-state and disease decoding. And reproducibility is unresolved — the original training code and pretrained checkpoint were not preserved, and a later re-implementation released by the first author reaches roughly 0.36 balanced accuracy on BCIC-2A, about 20 points below the published figure, with the gap still under investigation. The preprint has not been peer reviewed.