- Rating
- 2(+1)
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 1
Read review
Paper summary
This work introduces SPOT (Selective Prediction Over Targets) a pretraining objective for EEG models that does not rely on signal reconstruction but rather signal selection. It also introduces CORTEX that uses SPOT to train a biosignal tokenizer. The authors have conducted a series of experiments with clean and noisy data to test their hypotheses.
Critical strengths
The paper is very well-written. The proposed methods are oriented around the hypothesis that reconstruction models use much of their capacity to reconstruct noise present at EEG signals. The main reason behind it is that MSE reconstruction allocates capacity over the directions with the most variance, regardless of whether this variance comes from signal or noise. They pretrain CORTEX using SPOT (the two proposed methodologies). I like the noise analysis with the synthetic data. The SPOT methodology is also model agnostic which makes it efficient for many other EEG FM architectures.
Critical weaknesses
However, I find the evidence not convincing enough. First of all, the analysis on synthetic data does not follow on real data (as discussed by the authors) or it is very over-simplified. Regarding the novelty, I do not see how SPOT is effectively different than any other codebook-based model (e.g LaBraM, NeuroRVQ). The objective is different that is true (InfoNCE instead of cross-entropy) but these models do not actually do reconstruction. They predict the closest codebook (in the latent space) for the masked token during pre-training. So I understand the authors' claim about that a reconstruction tokenizer (stage 1 at LaBraM) learns also noise but both approaches learn a specific set of latent vectors and it's the model's pre-training job to decide which ones to use, essentially choosing from nearby content. To that argument the authors have an ablation with LaBraM but I don't see the effectiveness (see questions below): LaBraM-recon is better than LaBraM-SPOT. I also have concerns about the results on Table 1. Cortex's pre-training corpus task distribution matches the downstream tasks while other EEG FMs have a broader pre-training corpus. This alignment can inflate performance. In addition, the authors only did probing which if not done properly can severely affect results - e.g. discussed in "Are EEG Foundation Models Worth It? Comparative Evaluation with Traditional Decoders in Diverse BCI Tasks" (ICLR2026). There is a lack of full-fine tuning results and benchmarks outside the pre-training corpus distribution.
Questions for authors
- How were the LaBraM-recon and LaBraM-spot trained ? Did the authors use the pre-trained models ? Did they re-train both tokernizer and LaBraM model from scratch using both settings ? How do the authors justify the results ?
- Could the authors perform a comparison where some FMs are pre-trained on the exact same corpus as CORTEX to alleviate distribution match concerns ? And evaluation on tasks outside clinical (sleep and seizure) ?
- About novelty, how is the method fundamentally/theoretically different besides the technical implementation than codebook-based EEG FMs ? The authors' claim about capturing noise applies only to the tokenisers.
- How do the authors create the positive and negative pairs ? If they are random maybe the model can use shortcuts like subject-id etc.
- How about 4096 or the InfoNCE temperature choices ? Any ablations on those parameters ?
5 comments
LaBraM training and same corpus.
Reply to Reviewer 3 (2/5). Part A, your questions. Questions: how were LaBraM-recon and LaBraM-SPOT trained, and can some models be pretrained on the same corpus as CORTEX?
How the two LaBraM arms were trained. Neither arm uses a pretrained LaBraM checkpoint. Both were trained from scratch on our pooled corpus with the same settings: batch size 64, 100 epochs, AdamW, the same data and sampling. Their stored training arguments differ only in the model name and the output paths. Both use LaBraM's encoder (5,818,976 parameters, identical state_dict keys and shapes). LaBraM-recon is LaBraM's stage-1 VQ-NSP tokenizer (reconstructs Fourier amplitude and phase through an 8192-entry codebook and a decoder; no masking). LaBraM-SPOT is the same encoder trained with SPOT (75% masking, EMA teacher with decay 0.999, residual FSQ with levels 8^8 and depth 5, InfoNCE with up to 4096 candidates; no decoder, no reconstruction term). Neither arm trains a stage 2; both are probed directly from the final-epoch checkpoint with the same frozen probe. Masking, a teacher and a quantizer come with predicting a code from context, so this comparison is supporting evidence; the single-factor test is the 2x2 on CORTEX (reply 1/5). Details: appendix section "The LaBraM arms".
How we justify the results. On clean signals the results depend on the task (LaBraM-recon higher on HSP staging, HMC and arousal; LaBraM-SPOT higher on TUH and Siena), and we do not claim an advantage on clean signals. Under 12 dB EMG, LaBraM-SPOT is higher on every task above the trivial floor: HSP staging 33.9 against 27.8, HMC 27.4 against 0.4, arousal 22.6 against 17.1, TUH 55.8 against 22.4, Siena 27.3 against 2.8. Our claim is about robustness, and it holds on both backbones.
Pretraining on the same corpus. We agree that this is the key control. Three of our comparisons hold the corpus fixed exactly, so corpus alignment cannot explain them:
- CORTEX-recon against CORTEX-SPOT, three pretraining seeds each: SPOT is ahead under 12 dB EMG in 9 of 9 task-by-seed comparisons (+8.7 +- 5.0) and equal on clean signals (-1.0 +- 1.6).
- LaBraM-recon against LaBraM-SPOT (above).
- The 2x2 test of target codes and loss (reply 1/5).
Corpora never seen in pretraining. Four of our eight evaluations use corpora that are not in our pretraining data. CORTEX-SPOT has the best frozen score at 12 dB on all four, by a clear margin over the next frozen model:
| Corpus (metric) | CORTEX-SPOT | Next best frozen model |
|---|---|---|
| Sleep-EDF (kappa) | 67.2 | 58.9 (NeuroRVQ) |
| HMC (kappa) | 51.7 | 43.7 (NeuroRVQ) |
| CHB-MIT (AUPRC) | 46.6 | 24.5 (BIOT) |
| Siena (AUPRC) | 30.2 | 9.2 (REVE) |
On clean signals it is best on two of the four (Sleep-EDF and Siena). The overlap also runs in both directions: LaBraM and CBraMod were themselves pretrained on data that include the TUH EEG corpus. Pretraining all six published models on our corpus is not possible within the discussion period; the Limitations section now states that they are evaluated from their released checkpoints.
Questions: positives and negatives, number of negatives and temperature.
Reply to Reviewer 3 (3/5). Part A, your questions. Questions: how are the positive and negative pairs created, can the model use shortcuts such as the subject, and what about 4096 negatives and the temperature?
Positives and negatives. The positive for each masked cell is the EMA teacher's code for that cell, computed from the unmasked recording. The negatives are the teacher codes of 4096 other masked cells in the batch, drawn without regard to subject, so they can come from the same recording (Section 2.1).
Shortcuts such as subject identity. A subject-identity shortcut could help against negatives from other subjects, but it cannot help against negatives from the same recording. We therefore trained CORTEX-SPOT with negatives restricted to the same recording, and with negatives restricted to other recordings (one pretraining run and one probe seed each).
| Model | TUH, clean signals | TUH, 12 dB | HMC, clean signals | HMC, 12 dB | Arousal, clean signals | Arousal, 12 dB |
|---|---|---|---|---|---|---|
| All other masked cells (main) | 71.1 | 70.4 | 52.8 | 51.6 | 57.6 | 54.4 |
| Same recording only | 69.1 | 68.8 | 49.8 | 47.3 | 56.4 | 54.8 |
| Other recordings only | 69.4 | 68.7 | 53.4 | 51.8 | 56.0 | 52.8 |
Performance on clean signals stays within 3 points of the main setting in either case. We also tried to decode identity from the frozen embeddings. Session identity (114 classes) is decoded with 65.2% accuracy from SPOT and 70.4% from reconstruction, and patient identity (42 classes) with 85.0% and 87.1%. Both encoders carry identity information, and SPOT carries slightly less. We find no evidence that SPOT relies on a subject-identity shortcut. This is in the appendix section "Negative source and identity probes".
4096 negatives and the temperature. Yes, we ran both ablations (table in the shared-results comment, and in the appendix section "Ablations of the SPOT objective").
| Model | TUH, clean signals | TUH, 12 dB | HMC, clean signals | HMC, 12 dB |
|---|---|---|---|---|
| 256 negatives | 63.3 | 63.5 | 46.4 | 44.7 |
| 1024 negatives | 66.9 | 58.9 | 50.0 | 48.7 |
| 4096 (main), learned temperature | 71.1 | 70.4 | 52.8 | 51.6 |
| Temperature 0.02 | 68.4 | 69.0 | 52.5 | 51.1 |
| Temperature 0.2 | 71.5 | 69.9 | 51.5 | 48.8 |
More negatives improve performance on clean signals, and the result changes little over the two temperatures. For reference, CORTEX-recon scores 59.4 on TUH and 47.7 on HMC at 12 dB. SPOT is ahead of it at 4096 negatives and at both temperatures. With fewer negatives it is ahead on one of the two tasks and behind on the other. The temperature of the main setting is now stated in the pretraining table of the appendix.
Questions: novelty.
Reply to Reviewer 3 (1/5). Part A, your questions. Question: how is SPOT different from codebook models such as LaBraM and NeuroRVQ?
Thank you for the detailed review and for the pointer to the comparison with traditional decoders. We answer your questions first (replies 1/5 to 3/5), and then the weaknesses (replies 4/5 and 5/5). We start with novelty because several other points depend on it. This comment also answers the novelty point in your weaknesses.
The difference is in how the code targets are built, not in the loss. We agree that stage 2 of LaBraM and NeuroRVQ predicts codes and does not reconstruct. But those codes come from a tokenizer that was first trained to reconstruct the signal, and stage 2 can only choose among them. In SPOT the targets come from an EMA teacher of the context encoder, trained only through the prediction loss, with no reconstruction term anywhere. So our claim is about where the codes come from, and this is also why it applies to codebook models even though their stage 2 does not reconstruct.
The theoretical difference. Theorem 1 is about exactly this step. A reconstruction-trained tokenizer allocates its capacity by input variance (part ii), so when noise dominates some directions, part of every code describes that noise. Stage 2 must then learn to tell apart codes that differ only in noise the context cannot predict, which is impossible to do from context and wastes the target vocabulary. A predictive code allocates capacity by what the context predicts (part i). SPOT moves the predictive principle from stage 2 into the tokenizer, so the codes are shaped by predictability from the start.
A test that separates the two. On the CORTEX backbone we crossed (1) the source of the target codes, a finished reconstruction model ("recon codes", as in LaBraM-style models) or a finished SPOT model ("SPOT codes"), with (2) the loss used to predict them, cross-entropy (as in LaBraM-style models) or InfoNCE (as in SPOT). The row "recon codes, cross-entropy" is therefore the LaBraM and NeuroRVQ recipe on our backbone. In every run the student is a fresh CORTEX with the main SPOT recipe, the teacher is frozen, and both kinds of code go through the same fixed quantizer. Each combination was trained twice (two student seeds). Scores are under 12 dB EMG, mean of the two runs.
SPOT codes score higher in all six comparisons, and in each run separately (smallest single-run gain +1.8). Replacing recon codes by SPOT codes raises the score by 8 points on average (3.5 to 13.6). Changing the loss with the codes fixed matters less, about 3 points on average, and with SPOT codes 0.3 points or less. So most of the difference comes from the codes, not the loss, which is the novelty we claim. One limit: our recon codes come from a reconstruction-trained encoder and a fixed quantizer, so this tests the idea and not LaBraM's exact learned codebook. Both runs are in the appendix section "Target codes versus loss".