acceptodds
Withdrawn from ICLR 2027

Learning What to Remember: Selective Memory for Long-Context EEG Representation Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

15 views

Abstract

Long EEG recordings contain brief informative events within extended background activity. Learning from longer inputs therefore requires a representation that preserves useful evidence as context accumulates. We introduce SPARK, an encoder combining recurrent memory across temporal chunks, selective compression, and utility-guided controllers. Self-supervised pretraining learns to propagate and summarize context; subsequent utility adaptation refines the representation using curated diagnostic labels and physiological objectives. Under matched 120-second pretraining, SPARK reaches recording-level balanced accuracy on CHB-MIT, compared with for a near-parameter-matched dense-attention control. Utility adaptation raises performance to , exceeding a cross-entropy control with shared auxiliary objectives and matched adaptation data and updates (). Evaluation on three recording-level datasets and ten short-window benchmarks shows the strongest adaptation gain on CHB-MIT, alongside task-dependent context preferences and transfer behavior. These results identify recurrent memory and utility adaptation as complementary contributors to long-context EEG representation learning.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG representation learning, long-context modeling, selective memory, recurrent memory, self-supervised learning, utility-guided adaptation, EEG foundation models

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_wUUyOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies long-context EEG representation learning, focusing on how to retain information useful for downstream tasks amid extensive background activity. The authors propose SPARK, which combines multiresolution time–frequency features, gated recurrent memory, and importance-weighted compression and replay. The model first learns to accumulate information across chunks through self-supervised pretraining, then adapts its representations and controllers using a supervised utility objective. Experiments cover three recording-level classification datasets and short-window tasks involving clinical EEG, sleep, affect, and motor imagery.

Critical strengths
  1. The problem is practically meaningful, and the method is designed around it. Important information in long EEG recordings often appears in only a small number of segments. Retaining that information in a compressed representation is a worthwhile problem. The paper separates information accumulation across chunks from subsequent utility adaptation. The overall design has a clear motivation and goes beyond simply adding a classification module to an existing encoder.
  2. The controlled comparisons are valuable. Under the same pretraining data and update budget, Table 17 compares Stage 1 with Dense-Attn-120, which has a similar parameter count, and reports an 11.48-percentage-point improvement in CHB-MIT recording-level BAcc. Table 20 further introduces a CE control with shared auxiliary objectives, using the same adaptation data and update budget; full SPARK still performs 1.57 percentage points better. These comparisons help examine the contributions of recurrent memory and utility adaptation, respectively. Provided that the reported numbers can be verified, I consider them important evidence supporting the paper’s contributions.
Critical weaknesses
  1. I cannot currently reconcile the reported CHB-MIT scores, which substantially affects my assessment. Section G.3 on page 31 reports 71 validation recordings from chb21–22 and 52 test recordings from chb23–24. However, the official file manifest contains only 64 and 31 original EDF files for these groups, respectively. There may be a reasonable explanation for the larger counts. For example, splitting the test EDF files into one-hour segments and retaining the final partial segments does yield 52 sub-recordings. However, I could not find an explanation of such processing in the paper. More importantly, the evaluation unit determines which metric values are attainable. As a simple example, ordinary accuracy evaluated on 10 samples can only take values of 0%, 10%, 20%, and so on. Although balanced accuracy is calculated differently, its possible values are also constrained by the numbers of positive and negative samples. If all original EDF files are retained, each EDF is treated as one evaluation sample, and any file containing a seizure annotation is labeled positive, the official manifest and annotations give positive/negative counts of 3/6 for chb23 and 12/10 for chb24. Following Equation (27), computing BAcc separately for each subject and then averaging equally across the two subjects gives: BAcc (%) = (5/12) × (20a + 10b + 5c + 6d). Here, a and b are the numbers of correctly classified positive and negative recordings for chb23, while c and d are the corresponding counts for chb24. All four must be integers within their valid ranges. The final score must therefore be an integer multiple of 5/12 (approximately 0.4167) percentage points, although not every multiple is attainable. For example, the two grid points surrounding 93.10% are 92.9167% and 93.3333%; neither rounds to 93.10% at two decimal places. I would like the authors to clarify the actual recording partition and labeling rules, and provide the confusion matrices for each test subject under all three seeds, so that readers can reproduce these scores using Equation (27).

  2. The current experiments more directly support the usefulness of long-context pretraining than memory retention across complete long recordings. The model is pretrained on 120-second inputs, but downstream CHB-MIT evaluation uses independent 10-second windows, resets memory between windows, and averages the predicted probabilities to obtain recording-level scores. Thus, during evaluation, the model does not continuously update its memory throughout an entire long recording. This does not diminish the potential value of long-context pretraining, but two capabilities should be distinguished: learning better short-window representations from longer pretraining inputs, and retaining earlier important events through extended periods of background activity.

Questions for authors

What exactly does “recording” mean in Section G.3? Please explain how the 71 validation recordings and 52 test recordings are derived from the original EDF files, and provide the mapping from source EDF files to evaluation recordings, together with the recording-level labeling rules. This would help determine whether the discrepancy in recording counts simply reflects undocumented preprocessing. Could you verify the three CHB-MIT BAcc values for full SPARK in Table 19? Please provide TP, TN, FP, and FN for each of the two test subjects under each seed, along with the validation-selected thresholds. If the metric is not calculated by computing BAcc from hard predictions separately for each subject and then averaging equally across subjects, please explain how the implementation differs from Equation (27). This clarification is the most likely to change my current assessment. How were the 51,695 ten-second Siena windows generated? The EDF headers of the 41 official files indicate approximately 141.02 hours of data in total. Extracting complete, non-overlapping ten-second windows separately from each file yields 50,749 windows. Please clarify the window stride, handling of incomplete final windows, and whether repeated sampling was used. Which long-context capability is directly supported by the current experiments? Downstream evaluation uses independent ten-second windows and resets memory between them. An improvement in recording-level scores therefore does not, by itself, demonstrate retention of earlier important events throughout a complete long recording. Do any existing controls or analyses provide more direct evidence of this retention capability? If not, could you confirm that the current experiments primarily establish the transfer value of long-context pretraining? Do the CHB-MIT gains hold for both test subjects? Please report BAcc separately for chb23 and chb24 for Stage 1, the CE control, and full SPARK under the three existing seeds. I would like to understand whether full SPARK improves over both controls on both subjects, or whether the average gain is mainly contributed by one subject. This would affect my assessment of the robustness of the results.

Reviewer_LtFwOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The work presents a model named SPARK, an EEG encoder that aims to learn long-context representations by combining multi-resolution spectral tokenization, recurrent memory over 6s chunks, and importance-weighted compression. The model is trained in two stages: self-supervised pretraining via masked reconstruction, followed by supervised utility adaptation using a multi-objective reward encompassing diagnostic margins, physiological feature preservation, consistency, and boundary smoothness. Evaluation was carried out on multiple EEG decoding tasks, including CHB-MIT, TUAB, and TUSZ.

Critical strengths
  • The idea of framing long-context EEG learning as a selective retention problem and addressing the "reconstruction trap" is sensible and provides an interesting motivation.

  • The experimental setup shows real effort at matched controls. The inclusion of near-parameter-matched architecture baselines (e.g., Dense-Attn), a cross-entropy control with shared auxiliaries, and matched pretraining updates helps support the technical contributions.

  • I appreciate the authors' transparency in fixing the chb01/chb21 same-person data leak and candidly reporting negative transfer (such as on motor imagery tasks) rather than claiming universal wins. The limitations appendix and the held-out audit honestly showing proxy trade-offs—such as the composite reward increasing while consistency and diagnostic BAcc drop—are also very welcome.

Critical weaknesses
  • The approach to training long 120s contexts seems disconnected from the inference phase. While pretraining uses 120s windows, downstream inference resets memory between independent 10-12s windows and relies on mean-pooling for recording scores. Recurrence barely runs downstream, and mean-pooling is arguably the worst pooling strategy for rare events.

  • The utility objective appears under-justified. Unbounded logit-margin can grow simply via scale rather than better discrimination, and boundary smoothing might inadvertently erase true abrupt onsets. Furthermore, relying on 5 band-powers does not perfectly equate to physiological fidelity, and the audit trade-offs themselves indicate that composite reward does not always align with downstream quality.

  • I am not convinced by the two-subject results on CHB-MIT. The test set relies on only two subjects (chb23-24), meaning the reported standard deviations over 3 fine-tuning seeds measure only fine-tune noise, not patient or pretrain variance. The +4.47pp gain over Stage 1 reduces to just +1.57pp over the CE baseline on these same subjects, and paired differences on TUAB and TUSZ cross zero.

  • The controls, while useful, are not entirely causal. Comparing SPARK with Flat-120 and Dense-Attn confounds recurrence with bottleneck/aggregation, as SPARK retains M=8 tokens plus query-replay while the controls compress to per-chunk summaries. Similarly, the CE adaptation control changes the schedule and drops the consistency machinery, so the performance gap cannot be purely attributed to the margin objective.

  • The evaluation relies on "separate campaigns" where model configurations (controller modes, adaptation cycles, context length) are heavily tailored per task. For example, 120s pretraining is best on CHB-MIT, but TUAB and TUSZ perform better at 60s. Additionally, claims of "aligned" downstream settings seem overclaimed given the dataset-specific modifications (e.g., SEED 62-to-19 adapter, Sleep-EDFx 2ch/100Hz padding) compared to baselines like REVE or CBraMod.

  • There are several numerical anomalies in the reported results that warrant explanation. The seeds in Table 19 appear unnaturally regular, and the AUROC/PR-AUC scores in Table 26 look mathematically derived directly from BAcc. The paired standard deviations are implausibly tight, and some baseline values appear copied exactly despite claims of different data splits. Finally, the writing tone is heavily LLM-polished, and while Appendix A discloses LLM assistance, this is secondary to the core issues of the 2-subject evaluation, the unused long-context at inference, and the unfair supervised comparisons.

Questions for authors
  • Could you directly test memory retention through event-position sweeps, distractor duration, chunk-shuffle, or deletion, and compare mean pooling against max/top-k or attention MIL pooling?

  • Would it be possible to perform leave-subjects-out cross-validation on all 23 CHB-MIT subjects and report event-level metrics (e.g., sensitivity, false alarms per hour) to better demonstrate true generalization?

  • How can we interpret the adaptation improvements if the CE baseline is evaluated with a strictly matched schedule, and plain CE is used for all baselines with the exact same Stage-2 labeled data?

  • Could you clarify the exact recording/chunking definitions, address the mismatch between reported recording counts and EDF files, and release the code, split manifests, and per-seed prediction logs to clear up the numerical anomalies?

Reviewer_AiyrOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper attempts to address the problem that long EEG records hide informative events in long background, which are not prioritized in reconstruction-based pretraining. The authors proposes SPARK. SPARK is an encoder that carries recurrent memory across 6 chunks and then compressed according to importance. They first self-supervised on EEG data and then adapted on 'diagnostic margin', 'band-power preservation', 'consistency', 'boundary terms'. The authors provide comparison against BIOT, LaBraM, CBraMod, CSBrain and REVE respectively on 3 recording levels, proving that recurrent memory plus utility-guided adaptation improves long context EEG.

Critical strengths
  1. Very clear framing of problem on 'reconstruction trap'
  2. Careful leakage control
  3. The authors compare their work with a variety of benchmarks
Critical weaknesses
  1. The paper emphasizes improvements on CHB-MIT, but it is only tested on chb23 and chb24; it makes me worry about the variability and utility of the given method
  2. Baselines gets less supervision: baselines don't have a supervised stage. I will suggest also run supervised training on baselines for better comparison
  3. Small or negative gains except CHB-MIT: except for the gains described in 1. gains in TUAB TUSZ are small. And as shown in Table, although SPARK outperforms baselines in many evaluations, the gain is small, some within confidence interval.
Questions for authors
  1. What are the results on full set of CHB-MIT subjects, not just two data points? It would help to show a bigger picture on the most performant evaluation you did.
  2. The paper mentions recurrent memory, there are many recent work on SSM-style baselines that excel on long context, how do they perform under your problem framing?
  3. Are TUAB, TUSZ and TUEV test subjects excluded from Stage 1 pretraining?
  4. I would like to flag out one set of numbers that looks abnormal to me. REVE in Table 1 has the same variance across evaluations. So I checked the REVE results in table 19, the results across seeds are: | Dataset | Seed 0 | Seed 1 | Seed 2 | Mean | |------:|-------:|-------:|-------:|-----:| | CHB-MIT | 79.84 | 80.49 | 80.93 | 80.42 | | TUAB | 80.70 | 81.35 | 81.79 | 81.28 | | TUSZ | 62.10 | 62.75 | 63.19 | 62.68 | In the three evaluations, the gaps across seeds are exactly (-0.58, +0.07, +0.51), and that's why they have the same VAR; are those number resulting from rounding?
Reviewer_BY5uOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SPARK, an EEG representation-learning framework that combines recurrent memory, importance-weighted compression, and selective replay. Self-supervised pretraining learns to process temporal chunks, followed by supervised utility adaptation through four differentiable controllers. Experiments cover three recording-level datasets and ten short-window benchmarks, with matched architecture controls, a supervised adaptation control, and reward-component analyses. The strongest improvements are on CHB-MIT, while preferred pretraining context and controller use vary across tasks. Recording-level predictions average independently processed windows, so the experiments assess transfer from longer pretraining contexts rather than persistent memory across an entire recording.

Critical strengths
  1. The problem of retaining informative events within long EEG recordings is clearly motivated.
  2. The evaluation includes architecture controls, supervised adaptation comparisons, and reward-component ablations.
  3. The paper acknowledges task-dependent gains and limitations rather than claiming uniform improvements.
Critical weaknesses
  1. Limited interpretation of context-length effects. Section 3.2 should explicitly reference Figure 3(b) and explain why CHB-MIT benefits from longer contexts while TUAB and TUSZ peak at 60 seconds. Since effective batch size also changes with context length, further analysis is needed to distinguish task-specific effects from optimization differences.

  2. Insufficient isolation of compression and replay. The current comparisons do not quantify the separate gains from importance-weighted summaries and replay. Ablations removing replay or chunk-importance weighting while retaining query-based attention would clarify their contributions. Comparing learned top-k selection with random or uniform selection under the same replay budget would further test the value of selective retrieval.

Questions for authors
  1. What explains the different context-length trends in Figure 3(b)? Do they persist when batch size is varied separately from context length or when pretraining is repeated with another seed?

  2. Could you separately remove replay and the chunk-importance prior, keeping the remaining architecture and learned query attention unchanged? At the same replay budget, how does learned top-k selection compare with random or uniformly spaced selection?

  3. Could you report the CHB-MIT results for each test subject for Stage 1, full SPARK, and the CE control?

  4. The utility pool explicitly excludes downstream validation and test subjects. Does the Stage 1 TUH pretraining manifest also exclude subjects and recordings used in the TUAB, TUSZ, and TUEV validation/test sets?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.