acceptodds
Under review as a conference paper at ICLR 2027

Exact Temporal Readouts Expose and Test a Two-Boundary Decision Pattern in Imagined-Speech EEG

Abstract

Temporal aggregation can determine an imagined-speech EEG prediction, yet its role is usually opaque. We introduce an exact temporal readout framework built around a bidirectional exponential-memory mixture (BEM): each trial logit is a normalized, analytically recoverable kernel over frame-level class evidence anchored at both trial boundaries. On BCI Competition 2020 Track 3, BEM raises development accuracy from 77.07 ± 1.01% to 85.73 ± 1.57% and gains 7.11 participant-paired percentage points on a training-excluded historical audit. Its exact kernels place 89.67% of their mass in the outer fifths. Distilling this pattern into endpoint-fifths pooling yields no detected difference from BEM; applying the rule to the same frozen mean-trained evidence tensors gains 4.13 points on development and 2.98 on audit, isolating aggregation from encoder retraining. Boundary-masking and filtering controls further support reliance on broad two-edge structure. In a separate NEMAR on007591 calibration-to-online stress test, models were locally frozen before EEG or decoder-outcome inspection, although event labels and official quality-control metadata were known. The declared raw-accuracy gate passes: all three participants favor BEM, with a 4.67-point mean gain. However, balanced accuracy falls for all three, and a disclosed post-outcome same-evidence endpoint transfer loses 2.67 points. Exact readouts therefore turn a temporal preference into a distillable, intervenable, and externally falsifiable hypothesis, while the mixed stress test rules out a general two-boundary law.

Primary area:
learning on time series and dynamical systems
Keywords:
imagined-speech EEG, temporal evidence integration, exponential memory, bidirectional sequence modeling, brain–computer interfaces

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_d7eJOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes Bidirectional Exponential-memory Mixture (BEM) to aggregate CNN-generated class evidence across an EEG trial. The main finding is that emphasizing trial beginnings and ends improves imagined-speech classification on BCI2020. The authors investigate this pattern through kernel analysis, fixed pooling controls, and external experiments. Although the primary BCI2020 dataset shows substantial improvements, I have serious concerns about the scientific significance, interpretability, generalizability, and presentation of the work.

Critical strengths
  1. The authors did honestly report negative results and acknowledge that the observed boundary preference may not generalize across datasets.
  2. The authors analyzed and evaluated different temporal aggregations by comparing different readouts.
Critical weaknesses
  1. I find it extremely hard to read. The paper has an unnecessarily complicated presentation and unclear figures. There are extensive mathematical terminology for what is fundamentally a temporal weighted-average operation, making the contribution difficult to follow. In Figure 1, there are unexplained graphical elements and terms such as "diagnose," "distill," and "same-evidence swap" without sufficiently intuitive explanations. In the Simplex mixture box, it’s unclear what it means that different lines (dashed and solid) connect dots. Figure 2C contains nearly overlapping curves. Figure 4 does not clearly communicate the statistical evidence, while Figure 5 combines multiple experimental settings without sufficiently clarifying their respective purposes. The complete model-performance comparison is relegated to the appendix rather than presented alongside the main findings.

  2. Limited methodological novelty. The proposed BEM is essentially a constrained temporal pooling method. The model is therefore structurally biased toward boundary emphasis. More importantly, based on the reported initialization, it seems that approximately 91.92% of the initial kernel mass already lies in the outer fifths, compared with the reported 89.67% after training. This substantially weakens the interpretation that the model discovers a new temporal pattern. Furthermore, a simple fixed endpoint-pooling rule performs comparably to BEM, questioning the necessity of the proposed parameterization.

  3. The most serious issue, in my opinion, is its unclear scientific significance and physiological validity. The central finding is that weighting the beginning and end of an EEG trial improves classification on BCI2020. However, the authors do not establish whether the discriminative information reflects imagined-speech activity or experimental confounds, such as cue-related responses, eye movements, muscular artifacts, or trial transitions. This is particularly concerning given the unusually high classification accuracy and the much weaker results reported in recent controlled imagined-speech benchmarks (e.g., Van Dyck and Van Hulle, 2026). Moreover, the study has limited itself to only one main dataset and encoder configuration. The external online experiment includes only three participants, showing no statistically significant raw-accuracy improvement. On ASU, BEM performs substantially worse than mean pooling, while the advantage disappears with an alternative encoder. These results provide little evidence that the proposed temporal weighting strategy generalizes.

References

Van Dyck, Bob. "The limits of imagined-speech decoding from EEG: A cross-dataset comparison and critical evaluation." (2026).

Questions for authors
  1. What evidence shows that the discriminative information near trial boundaries reflects imagined speech rather than cue-related activity or ocular/muscular artifacts? Can the authors provide time-resolved, spatial, and spectral temporal analyses, including experiments restricted to the actual imagery interval? More interestingly, would this finding only holds to imagined speech dataset or also to cue-based motor imagery dataset, visual ERP datasets: SSVEP, P300, etc.
  2. How does the learned temporal kernel differ from its initialization? What genuinely new temporal structure is learned during training?
  3. Why is BEM necessary when fixed endpoint pooling achieves comparable performance? Do attention (e.g., EEG FM can have attention pooling) and freely learned temporal pooling (e.g., CNN based pooling) baselines also learn to emphasize both boundaries? Given the negative or inconclusive external results, what evidence supports the practical usefulness of BEM beyond the original BCI2020 pipeline?
  4. Could the authors clarify the degree of independence of the historical audit and provide the complete model-comparison results in the main text?
Reviewer_nweVOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates the temporal aggregation (readout) layer in imagined-speech EEG decoding. The authors propose a Bidirectional Exponential-Memory Mixture (BEM) readout, which learns a normalized mixture of exponential decays from both trial boundaries. On BCI Competition 2020 Track 3, BEM improves accuracy over mean pooling and puts ~90% of its weights on the first and last 1/5 of the trial. The authors then distill this into a simple "endpoint-fifths" averaging rule, and further test it on a compact encoder (EEGNet) and three external datasets (on007591, ASU, FEIS), where the two-boundary advantage disappears or reverses.

Critical strengths

None

Critical weaknesses
  1. Trivial method and repackaging failure as contribution: The proposed BEM is just simple exponential moving average from two ends, and Propositions 1-3 are elementary textbook math (Appendix A.2 even uses a toy T=3T=3 example to fill space). The authors' own experiment shows BEM has no real difference from simply averaging the first and last 1/5 steps (endpoint-fifths). More importantly, when changing to a standard compact encoder (EEGNet) or testing on external datasets (ASU, FEIS, on007591), the method completely fails or drops performance a lot. Instead of admitting the method just overfits to one specific pipeline, the authors twist the narrative and package "falsifying the hypothesis" as a main contribution.
  2. Lack of basic EEG domain knowledge and signal analysis: The paper claims ~90% decision weights are in the outer 20% boundaries of BCI2020 trials, but never plots a single real EEG waveform, ERP, spectrogram, or scalp topography. In BCI experiments, trial start and end boundaries commonly contain cue onset/offset evoked potentials or EOG/EMG artifacts. Treating EEG purely as abstract tensors and only doing masking/cropping on logits without checking the actual physiological signals gives very little insight to the EEG community.
  3. Strong traces of automated AI generation (Auto-Research) in figures and writing:
    • The AI Use Statement (Line 617-624) admits generative AI was used across the entire pipeline from ideation, experiment design, coding, result interpretation, to drafting.
    • Figure 1 is obviously drawn by Python script using fake sine waves with irregular low-frequency fluctuations rather than real EEG signals, inside crude matplotlib boxes with text symbols (☑).
    • The text heavily relies on repetitive LLM defensive patterns ("not X, but Y" / "rather than" appears over 20 times in 9 pages), while Discussion and Conclusion are compressed into just 16 lines of semicolon-separated summaries.
  4. Questionable sandbox logs and timestamp errors in Appendix: Appendix B.3 and B.4 read like automated agent execution logs (putting raw SHA-256 hashes bcf3203c... in the text and using a separate validator script to re-verify saved predictions). Line 807 even states "A provenance note corrects two impossible manually typed times and records the filesystem chronology", which strongly suggests an LLM agent hallucinated timestamps when generating freeze records and was later patched by a file-system check.
Questions for authors

I have no question for the AI behind the fake author.

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

The paper proposes a readout for trial-level classification of imagined-speech EEG that replaces mean pooling of frame-level class evidence by a normalized temporal kernel. For each class, the kernel is a learned probability measure over exponential decays anchored at the trial start and at the trial end, plus a uniform component. The authors call this readout the bidirectional exponential-memory mixture (BEM). The primary dataset is BCI Competition 2020 Track 3, with 15 participants, five classes and 64 channels, and the encoder is a convolutional network trained end to end. BEM raises accuracy from 77.1% to 85.7% on the checkpoint-selection run, and from 72.4% to 79.5% on a third run whose labels had been inspected during development. The learned kernels place 90% of their mass in the first and last fifths of the trial. A fixed rule that averages the evidence in the first and last fifths matches BEM. Applying this rule to the frozen evidence tensors of models trained with mean pooling raises accuracy by 4.1 and 3.0 percentage points on the selection run and the third run, respectively. The authors use this swap to separate the effect of aggregation from that of encoder retraining. Masking the trial edges reduces BEM's accuracy more than masking the center of the trial. The gain persists when the 50 Hz notch filter is removed or applied causally. The effect does not transfer to other encoders or datasets. A compact EEGNet-style encoder shows no gain, mean pooling outperforms BEM by 13 to 17 percentage points on the ASU corpus, and no model exceeds chance on the FEIS corpus. The authors also froze an analysis on a three-participant online dataset (on007591) before inspecting its outcomes. In this test, raw accuracy rises by 4.7 percentage points, balanced accuracy falls for every participant, and the endpoint rule applied to frozen evidence lowers accuracy by 2.7 percentage points. The authors conclude that the two-boundary pattern is a property of the primary pipeline and do not claim a general law.

Critical strengths

The paper's contribution is an exact decision kernel, showing it can be replaced by a distillation of the kernel into a simple and fixed rule, which can be applied to frozen evidence tensors. This rule replacement separates the effects of aggregation from co-adaptation of the encoder. The authors find that the replacement supports their hypothesis sometimes (BCI 2020) and sometimes contradicts the hypothesis (on007591).

Critical weaknesses
  1. The paper does not examine what content is in the the trial edges that the kernels focus in on; without knowing this, it is hard to assess whether the finding is of interest. Trial windows begin and end at experimenter-defined times. Class-informative signal at the edges may therefore come from cue-evoked potentials, from eye or muscle activity at cue onset and offset, or from the epoching itself, rather than from imagined speech. The encoder is a second possible source. Its convolutions use zero padding (kernel widths 11 and 9, the latter with dilation 2), so the first and last evidence frames are computed partly from padded values and differ systematically from interior frames. A kernel that weights edge frames heavily can thereby apply a position-specific decision rule through an evidence head that is the same at every time step. The paper tests neither of the two possible sources. Suitable tests include examining the channel topography and spectrum of the edge-discriminative evidence, comparing class-conditional evoked responses at the edges, and retraining with reflect padding or with the edges cropped before the encoder. Without such tests, the result shows only that a classifier on one dataset weights edge frames heavily for unknown reasons. The failures to transfer suggest that the reasons are specific to this pipeline.

  2. The held-out evidence is weak. The development accuracy is the checkpoint-selection criterion; the labels of the third run were inspected during development; the compact encoder shows no effect; the ASU corpus reverses the effect by 13 to 17 percentage points; and the external online test runs at or below the majority-class rate (24% to 33% accuracy, with 20% chance), where no comparison of readouts is informative. The remaining result is a within-pipeline effect on one benchmark. Together with a 137-parameter readout and three elementary propositions, this result is a slender contribution for this venue.

Questions for authors
  1. Clarification: What do the edge frames encode? For the primary pipeline, please report the class-conditional evoked responses in the first and last 0.6 s of the trial, and the channel topography of the edge-discriminative evidence. Please also retrain BEM and mean pooling with reflect or replicate padding, or after cropping 0.1 s from each edge before the encoder. If the gain disappears under non-zero padding or cropping, the finding is an artifact of the encoder.

  2. Clarification: where does the 795-sample window lie relative to the cue in the BCI 2020 protocol, and does the first or last fifth of the window overlap a cue onset or offset?

  3. Clarification: how do the primary results compare with the published competition results and with EEGNet or EEG Conformer trained on the same runs? The paper avoids rank claims, but a reader needs to know whether 77% accuracy with mean pooling is a competitive starting point.

  4. Clarification: why does BEM lose 13 to 17 percentage points to mean pooling on the ASU corpus? With about 190 trials per participant, does BEM overfit its 137 readout parameters within nested cross-validation, or does the task have a different temporal structure?

  5. Clarification: performance on on007591 is at or below the majority-class rate. Given this, what does the raw-accuracy criterion establish that the balanced-accuracy result does not refute?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.