CortexFM — A Lightweight Multimodal Foundation Model for Spike + EMG BCI

A 5.04 M-parameter multimodal Transformer foundation model that jointly learns spike trains and surface EMG envelopes from public DANDI motor-cortex data, evaluated on the FALCON M1 benchmark.

CortexFM은 약 5.04 M 파라미터 규모의 다중모달 Transformer 파운데이션 모델로, 공개 DANDI 운동피질 데이터로부터 스파이크와 EMG 포락선을 공동 학습하고 FALCON M1 벤치마크 위에서 평가된 경량 BCI 백본이다.

Model description

CortexFM is a small, public, and fully reproducible foundation model for invasive brain–computer interface (BCI) decoding. It targets the regime where private million-hour pretraining data and 45 M – 350 M parameter backbones are not available, and asks how far we can push neural-decoding quality with ~3.85 hours of public data and a ~5 M-parameter model trained in about six minutes on a single consumer GPU.

CortexFM은 (i) 단위(per-unit)/근육(per-muscle) 정체성 보존, (ii) 공개 데이터·공개 벤치마크만으로의 재현, (iii) FALCON 표준 정렬을 세 가지 설계 원리로 둔다. 백본은 10-layer × 6-head × d=192 PreNorm Transformer (4.45 M params, FLASH SDPA), 헤드는 spike Poisson NLL 재구성, EMG MSE 재구성, cross-modal InfoNCE 대조 학습의 세 갈래로 구성된다.

Architecture summary

Component	Configuration
Backbone	PreNorm Transformer, 10 layers, 6 heads, d_model = 192, FFN = 768, GELU
Attention	SDPA with FLASH / EFFICIENT backends (PyTorch 2.10)
Backbone params	4,449,024 (≈ 4.45 M)
Spike tokenizer	Per-unit learned embedding ⊕ log(1 + α · count) + temporal positional embedding
EMG tokenizer	Per-muscle learned embedding ⊕ scalar-to-vector MLP + temporal positional embedding
Heads	Spike recon (Poisson NLL), EMG recon (MSE), Contrastive projector (d_p = 128)
Total params	5,044,994 (≈ 5.04 M)
Bin size	20 ms (FALCON official)
Context length T	64 bins (1.28 s) → 1,088 tokens
Mixed precision	BF16 (InfoNCE softmax promoted to FP32)

Intended uses

Research-grade neural decoding of primate motor cortex (M1) spike trains into 16-channel surface/intramuscular EMG envelopes.
Backbone for downstream BCI probes: as a frozen feature extractor with a thin (~3 K-param) per-session output-space affine adapter, CortexFM enables session-1 adaptation to held-out recording days.
Cross-modal pretraining baseline for studies that compare per-unit tokenization against patch-tokenized BCI foundation models (e.g., NDT-3) at a 1/9 – 1/69 parameter ratio.
Educational reference for compact (~5 M-param) foundation-model training from public data on a single consumer GPU.

Out-of-scope uses

Clinical or assistive deployment. This is a research checkpoint trained on a single non-human primate (MonkeyL, DANDI 000941). It is not intended for human BCI control or medical decision-making.
Cross-subject generalization. The pretraining set is one subject; cross-subject transfer (e.g., MonkeyN, MC_Maze, human cortex) has not been validated.
Direct kinematic decoding. The model outputs EMG envelopes; downstream kinematic readouts require an additional decoding stage.
Real-time control without calibration. Held-out sessions require a brief (≥ ~8 s) per-session affine calibration to enter the positive-R² regime.

Training data

Dataset	DOI	Subject	Modality	Duration
DANDI:000941 (Rouse & Schieber 2018)	10.48324/dandi.000941/0.211015.0907	MonkeyL (1 NHP)	M1 spikes (64 units) + intramuscular EMG (16 muscles)	11 sessions total

Pretraining uses the four held-in calibration sessions of DANDI 000941 (sessions 20120924, 20120926, 20120927, 20120928), totaling 3 h 38 min of paired spike + EMG recordings. The remaining 7 sessions (4 minival + 3 held-out calibration) are reserved for FALCON M1 evaluation and OOD session-1 adaptation.

License: CC-BY-4.0 (DANDI public release).

Preprocessing pipeline

EMG: 60/180/200/300/400 Hz notch → 4th-order Butterworth high-pass at 65 Hz → rectify → 99 % clip → 95 % normalize → polyphase resample (1 kHz → 50 Hz) → re-rectify → 10 Hz low-pass envelope.
Spike: 20 ms bin counts per unit on the same time grid.
Output: Zarr store with /emg/envelope, /spike/counts, /eval_mask, and trial markers. Spike/EMG share a common 20 ms bin axis (FALCON official invariant).

Training procedure

Objectives

Joint loss with three components:

$\mathcal{L}_{\text{total}} = w_{\text{spike}} \cdot \mathcal{L}_{\text{spike}} + w_{\text{emg}} \cdot \mathcal{L}_{\text{emg}} + w_{\text{cont}} \cdot \mathcal{L}_{\text{cont}}$

$\mathcal{L}_{\text{spike}}$: Poisson NLL over per-unit spike counts.
$\mathcal{L}_{\text{emg}}$: MSE on per-muscle EMG envelopes.
$\mathcal{L}_{\text{cont}}$: Symmetric InfoNCE on pooled cross-modal embeddings (FP32-promoted), temperature τ = 0.1.

Loss weights $(w_{\text{spike}}, w_{\text{emg}}, w_{\text{cont}}) = (1.0, 1.0, 0.5)$.

Masking

Spike and EMG tokens are independently masked at 50 % per bin. Either modality must be reconstructed from the unmasked complement of itself and the (independently masked) cross-modal signal.

Optimization

Hyperparameter	Value
Optimizer	AdamW
Learning rate	3 × 10⁻⁴
Weight decay	0.01
LR schedule	Linear warmup (500 steps) → cosine decay
Batch size	8
Context length T	64 bins
Mixed precision	BF16 (InfoNCE softmax in FP32)
Gradient clip	1.0
Max epochs	50 (early-best at epoch 28)

Training environment

Item	Value
GPU	NVIDIA RTX 5080 (16 GB GDDR7, sm_120) — single consumer card
OS / runtime	WSL2 Ubuntu 24.04
Framework	PyTorch 2.10.0 + cu128, PyTorch Lightning
Wall-clock training time	≈ 6 minutes for 30 epochs
Best checkpoint	`epoch28-0.2599.ckpt` (60.7 MB, val_loss = 0.2599)
External cloud GPU	None — fully on-device

Train/val gap stayed below 0.03 throughout, so no early stopping was applied and the lowest-validation-loss checkpoint was kept verbatim.

Evaluation

FALCON M1 (held-in calibration sessions, variance-weighted R² over 16 muscles)

Setting	Params (used)	Per-session R² (mean ± std)	Pooled R²	NL
POYO-1 zero-EMG floor	15.47 M (0 used)	−1.273 ± 0.299	—	2.4e-5
POYO-1 frozen + per-session affine	15.47 M + ~2 K/sess	+0.451 ± 0.112	+0.498	—
CortexFM zero-shot	5.04 M	−1.035 ± 0.234	—	0.131
CortexFM frozen + Ridge linear probe	5.04 M + ~3 K	−0.258 ± 0.327	+0.125	—
CortexFM + EMG-head FT 200 step (ZS)	5.04 M (FT 37 K = 0.75 %)	−0.038 ± 0.063	—	—
CortexFM frozen + per-session affine	5.04 M + ~3 K/sess	+0.484 ± 0.102	+0.529	—

NL = FALCON normalized latency (inference time / data duration).

Auxiliary co-bps (CortexFM only)

Mean 0.756 ± 0.128 bits/spike above per-unit mean-rate baseline on the four held-in calibration files.

Held-out OOD calibration sessions (DANDI 000941, days +6 to +30, 3 sessions)

Variance-weighted R², calibration ≈ 640 bins:

Session	CortexFM + affine	POYO-1 + affine	Δ (POYO-1 − CortexFM)
20121004	+0.4443	−0.0209	−0.4652
20121017	+0.2730	−0.2326	−0.5056
20121024	+0.4046	+0.1824	−0.2222
Per-session mean ± std	+0.374 ± 0.073	−0.024 ± 0.169	−0.398
Pooled R²	+0.387	−0.008	−0.395

The decisive separation between CortexFM and POYO-1 emerges in the OOD held-out sessions: the held-in gap is small (Δ = +0.031 in CortexFM's favor), but the held-out gap reaches Δ = +0.395 pooled R² — a gap attributable to backbone representation quality rather than the affine adapter recipe (both backbones use the identical per-session output-space affine).

Why zero-shot R² is negative

Three factors documented in the thesis (Chapter 6):

Objective mismatch: pretraining minimizes joint masked-recon + InfoNCE, whereas FALCON M1 measures EMG-only regression.
Inference-time input shift: EMG is the prediction target at evaluation time, so the EMG tokenizer is fed zeros — out of pretraining distribution.
Absence of per-session linear correction: standard FALCON pipelines fit a shallow regressor per session; CortexFM zero-shot does not.

The Ridge linear probe resolves factor (3) (pooled R² entering the positive regime at +0.125); EMG-head fine-tuning resolves factors (1)+(2) (per-session R² up to −0.038); per-session affine resolves all three jointly (pooled R² = +0.529).

Limitations

Single-subject pretraining. Pretraining is restricted to MonkeyL (DANDI 000941). Cross-subject transfer to MonkeyN, MC_Maze, or human cortex is not validated.
n = 3 OOD sessions. The held-out evaluation uses three sessions; effect sizes are large but formal Holm-corrected statistical power is limited.
Calibration dependence on OOD. With fewer than ~400 calibration bins (< 8 s) on a held-out session, OOD R² becomes unstable. Real-time deployment therefore requires a brief calibration cycle per session.
EMG-only readout. The model decodes 16-channel EMG envelopes, not kinematics directly. A downstream kinematic stage is needed for end-effector control.
No clinical validation. The model is research-grade. It has not been evaluated for safety, robustness, or efficacy in any clinical BCI setting and must not be used as such.

How to use

import torch
from cortex_fm.training import CortexFMPretrainModule

# Load checkpoint
module = CortexFMPretrainModule.load_from_checkpoint(
    "epoch28-0.2599.ckpt",
    map_location="cuda",
    strict=True,
)
module.eval()

# Inference: spike counts -> EMG envelope
# spike_counts: (B, T=64, N=64) int
# emg_placeholder: (B, T=64, M=16) float (zeros at inference)
spike_counts = torch.zeros(1, 64, 64, dtype=torch.long, device="cuda")
emg_placeholder = torch.zeros(1, 64, 16, device="cuda")

with torch.no_grad():
    out = module(spike_counts, emg_placeholder)

emg_pred = out["emg_pred"].view(1, 64, 16)[:, -1, :]  # (B, 16) at last bin
log_rate = out["log_rate"]  # (B, T, 64) Poisson log-rates

For FALCON M1 evaluation, see benchmark_wrapper/ and the CortexFMFalconDecoder reference implementation.

Citation

@mastersthesis{shin2026cortexfm,
  author       = {Shin, Jaeguk},
  title        = {{CortexFM}: A Lightweight Multimodal Foundation Model for Spike--EMG Decoding on Public Brain--Computer Interface Data},
  school       = {Dong-eui University},
  type         = {{M.S.} thesis},
  year         = {2026},
  month        = jun,
  address      = {Busan, Republic of Korea},
}

If you also use the FALCON benchmark, please cite Karpowicz et al. 2024.

Ethical considerations

CortexFM is a research artifact. The following points apply:

Animal data. Pretraining data come from a single non-human primate recorded under the original Rouse & Schieber 2018 protocols (DANDI 000941, CC-BY-4.0). No additional animal experiments were conducted for this release.
No human data. The released checkpoint has not been trained or evaluated on human neural recordings.
Dual-use awareness. Invasive BCI decoding can in principle inform assistive devices or surveillance / commercial neuro-monitoring systems. The author releases this checkpoint to support open scientific reproduction and lightweight benchmarking; downstream users are responsible for ensuring their applications respect informed consent, neural privacy, and applicable medical-device regulation.
No clinical claims. CortexFM has not been evaluated against clinical-grade BCIs and must not be deployed in patient-facing systems without full regulatory validation.

Model details

Developed by: Jaeguk Shin (신재국), Dong-eui University, Department of Artificial Intelligence — M.S. thesis (June 2026), advised by faculty of Dong-eui University AI Department.
Model type: Multimodal Transformer foundation model (spike + EMG).
Language: N/A (the inputs are neural signals; the model card is bilingual EN/KO).
License: MIT (see LICENSE).
Finetuned from: Trained from scratch.
Related links: Thesis full text and reproducibility scripts to be released at the GitHub companion repository.

Downloads last month: -; Downloads are not tracked for this model. How to track