acceptodds
Under review as a conference paper at ICLR 2027

SpikeMind: Physiologically Grounded Spiking Concepts for EEG Emotion Recognition and Rationale Generation

Abstract

Electroencephalography (EEG) emotion recognition often relies on latent features whose physiological meaning is difficult to compare across subjects or trace through language rationales. SpikeMind uses 12 differential-entropy (DE)-grounded temporal concepts as shared variables for recognition and rationale generation. Their targets describe regional activity and contrasts computed from training-standardized DE features. A coordinate-preserving spiking encoder and fixed band–region readouts learn trajectories aligned with these targets. Reliability-weighted source-subject prototypes classify their temporal summaries and expose predicted-versus-runner-up margin evidence. A planner defines supervision and evaluation facts, while Qwen3-4B organizes the complete continuous concept record into a rationale. On SEED and SEED-IV, weighted accuracy reaches and under subject-dependent evaluation, and and under source-only leave-one-subject-out evaluation. The latter exceed EEG Conformer under the same source-prototype rule by and percentage points. Concept-marker correlations range from to , and language evidence-retention F1 ranges from to . This shared record makes marker agreement, decision sensitivity, and language evidence retention separately testable within one recognition system.

Primary area:
interpretability and explainable AI
Keywords:
EEG emotion recognition, spiking neural networks, physiological concepts, cross-subject generalization, language rationales

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ZQTgOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

SpikeMIND is a method using differential entropy (DE) concepts based on physiology (electrode placement and spectral bands) to improve emotion classification across subjects and generate rationale for it. The model is tested across 2 datasets, SEED and SEED-IV, with 2 settings: subject-dependent and leave-one-subject-out (LOSO), where it outperforms baselines based on diversity of model (graph, transformer spiking net, …).

Critical strengths
  • The system does not only focus on performance, from its design it uses physiological inductive bias and provides explainability with natural language rationale when classifying emotion.
Critical weaknesses
  • The evaluation is based only on 2 datasets spanning 30 participants in total while focussing on probing cross-subject generalization.
  • The system pipeline is complex, hard to understand in the current version of the manuscript, and seems specialized for emotion recognition.

Note: The code is not shared, even if the supplement material provides details for reproducibility.

Questions for authors

I would lean for a rejection of the paper. The paper method is specialized in EEG emotion recognition, which limits its significance especially with an evaluation limited to 2 datasets spanning 30 participants. The method evaluation would benefit from more dataset diversity, larger scale, and adaptation to other EEG tasks.

Questions ordered by importance:

  • Q1: Could you explain the motivation to select SEED and SEED-IV as the datasets to evaluate SpikeMIND? Are there some of their characteristics that make them the best for probing properties, such as cross-subject generalization or rationale generation? To strengthen the evaluation, additional datasets in the EEG emotion recognition task family could be considered.
    • SEED-V and SEED-VII provide harder classification tasks (more class) and more subjects that might leave more room for a model to outperform other methods in the LOSO setting.
    • Outside the SEED dataset family, many other datasets: DEAP, DREAMER, FACED, MPED exist. Would those datasets fit the SpikeMIND framework or would they need adaptation outside the current method's scope (e.g., different electrode geometry). If possible, within the rebuttal time, could SpikeMIND be tested on at least 2 additional datasets?
  • Q2: Could you explain if the SpikeMIND framework could be extended to other tasks like BCI or if the method is specialized in EEG emotion recognition? Testing the framework on another EEG task where the cross-subject problem is imporant would make the method more significant.
  • Q3: Could you clarify if:
    • the results reported are the ones from the baseline' paper or are they from rerunning the methods?
    • the evaluation protocols match between SpikeMIND and the other original evaluation baselines?
  • Q4: Could you clarify some points about the language generation rationale and its evaluation?
    • Is the LLM only used to generate a rationale and provide explainability, or is it also a component improving emotion classification?
    • What is the baseline in the language rating (Fig 3.c)?
    • Could you provide in the supplementary materials some qualitative example of the rationale generated from the LLM and the baseline?
  • Q5: Why are the weighted average (WA) and unweighted average (UA) metrics used? It seems that results are most of the time within 1 point. Are the SEED datasets unbalanced?
  • Q6: Could you add some references and contrast SpikeMIND to the EEG foundation model approaches that try to tackle the cross-subject problem with a different philosophy?

Additional feedback not affecting the decision assessment:

  • Adding to the supplementary material some context on the DE framework and concept methods (definition of the notions such as concept, prototype, evidence) would help readers non-familiar with them.
  • To improve the paper's clarity, figures 1 and 2 would benefit from more detailed captions.
  • The language ratings would benefit from more blind raters even if they would grade fewer paired records.
  • In the text some notations are not introduced, for example:
    • LL line 154
    • XfX^f, HH, and WW line 158
    • R\mathcal R line 194
    • I\mathbb{I} line 198
    • QkQ_k line 199
Reviewer_f2B5OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SpikeMind, a framework for EEG emotion recognition and rationale generation. The model encodes differential entropy (DE) features via a dual-branch SNN, aligning 12 physiological concept trajectories with predefined targets to perform reliability-weighted, prototype-based classification. A fine-tuned Qwen3-4B then translates the structured concept evidence into natural-language rationales, with evaluations demonstrated across subject-dependent and cross-subject protocols on SEED and SEED-IV.

Critical strengths
  1. The adoption of subject-balanced prototype matching provides a structured, interpretable classification rule in the concept space rather than relying on a standard linear head.

  2. The paper includes extensive empirical analyses, such as grounding metrics, ablation studies, and classifier interventions, to inspect the alignment between the learned SNN representations and the heuristic target markers.

  3. The manuscript is clearly written and well-structured, with intuitive framework diagrams and visualizations that facilitate understanding of the overall pipeline.

Critical weaknesses

W1. Limited architectural novelty and unsubstantiated computational benefits.

The encoder heavily borrows its dual-branch design from existing work, combining paradigms established in DBPNet[1] and directly adapting its spiking blocks and fusion from S²M-Former[2], rendering the architectural novelty incremental. Furthermore, SpikeMind's estimated arithmetic energy (194.6/251.9 μJ) already exceeds that of the dense EEG Conformer baseline (177.56 μJ), even without factoring in the non-trivial overhead of DE preprocessing and downstream Qwen3-4B inference. Without empirical wall-clock speedups or physical energy savings, the added complexity of a spiking architecture is unjustified.

W2. Limited methodological novelty and narrow scope of physiological grounding.

Guiding EEG representation learning via auxiliary physiological supervision is an established paradigm: prior works such as Neuro-KE [3] and DeeperBrain [4] explore knowledge-guided EEG representations, with DeeperBrain incorporating richer neurodynamic targets spanning functional connectivity, cross-frequency coupling, and signal complexity.

W3. Insufficient justification of prototype construction and temporal aggregation.

The subject-balanced prototype construction may offer limited benefits on SEED and SEED-IV, given their relatively balanced recording protocols, yet no comparison with conventional sample-level class prototypes is provided. Additionally, the temporal aggregation mechanism assumes that larger absolute concept responses indicate greater importance, which may not reflect actual discriminative relevance. The resulting weighted temporal center may also fall in a low-activity interval when multiple peaks exist, making its interpretation as an important temporal location questionable.

W4. Insufficient justification for LLM-based rationale generation.

The classification pipeline already produces a complete, deterministic evidence record that can be straightforwardly rendered into readable explanations using rule-based templates or well-prompted off-the-shelf models. More critically, the LLM adds no new physiological insights or classification utility; its role is strictly confined to textual reformatting, which introduces latency overhead and risks hallucinations or omissions of the extracted evidence.

W5. Limited dataset diversity and insufficient generalizability.

The experiments are restricted to SEED and SEED-IV. Such limited evaluation is insufficient to establish the robustness and generalizability of the proposed framework across different recording conditions, electrode montages, and emotion elicitation paradigms. The authors should conduct additional experiments on more diverse EEG emotion recognition datasets, such as FACED [5], REFED [6], DEAP [7], and EmoEEG-MC [8], to provide stronger evidence of generalizability.

References [1] Ni, Q., et al. (2024). DBPNet: Dual-Branch Parallel Network with Temporal-Frequency Fusion for Auditory Attention Detection. IJCAI. https://doi.org/10.24963/ijcai.2024/345 [2] Wang, J., et al. (2025). S²M-Former: Spiking Symmetric Mixing Branchformer for Brain Auditory Attention Detection. NeurIPS. https://proceedings.neurips.cc/paper_files/paper/2025/hash/b8330f5b70b3c53172417deac6f057b1-Abstract-Conference.html [3] Neuro-KE: Scaling EEG Foundation Models via Label-Free Knowledge Integration. OpenReview. https://openreview.net/forum?id=y7vaftM0NX [4] Wang, J., et al. (2026). DeeperBrain: A Neuro-Grounded EEG Foundation Model Towards Universal BCI. arXiv:2601.06134. https://arxiv.org/abs/2601.06134 [5] Chen, J., et al. (2023). A Large Finer-grained Affective Computing EEG Dataset. Scientific Data. https://doi.org/10.1038/s41597-023-02650-w [6] Ning, X., et al. (2025). REFED: A Subject Real-time Dynamic Labeled EEG-fNIRS Synchronized Recorded Emotion Dataset. NeurIPS (Datasets and Benchmarks). https://proceedings.neurips.cc/paper_files/paper/2025/hash/2bf0ed7c35d9d84128e7f7c72ab76402-Abstract-Datasets_and_Benchmarks_Track.html [7] Koelstra, S., et al. (2012). DEAP: A Database for Emotion Analysis Using Physiological Signals. IEEE Transactions on Affective Computing. https://doi.org/10.1109/T-AFFC.2011.15 [8] Xu, X., et al. (2025). A Multi-Context Emotional EEG Dataset for Cross-Context Emotion Decoding. Scientific Data. https://doi.org/10.1038/s41597-025-05349-2

Questions for authors

Q1.Could the authors clarify the main methodological contributions beyond existing dual-branch SNN encoders and physiology-guided EEG representation learning approaches? Given the reported energy estimates, what practical advantages does the spiking architecture offer over conventional non-spiking alternatives?

Q2.How much does subject-balanced prototype construction improve performance compared with conventional sample-level prototypes? Additionally, how should the weighted temporal center be interpreted when concept trajectories exhibit multiple separated peaks rather than a single salient interval?

Q3.Could the authors compare fine-tuned Qwen against well-prompted frozen LLMs or lightweight alternatives in terms of explanation quality, evidence retention, and inference cost?

Q4.How well does SpikeMind generalize beyond SEED and SEED-IV? Additional evaluations on datasets such as FACED, REFED, DEAP, or EmoEEG-MC would help assess its robustness across different recording configurations and emotion elicitation paradigms.

Reviewer_tfFqOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on EEG-based emotion recognition, emphasizing the use of the differential-entropy temporal concept as shared variables for recognition and rationale generation. A coordinate-preserving spiking encoder and fixed band-region readouts were proposed to learn trajectories and construct source-subject prototypes. Evaluation with SEED and SEED-IV showed good performance on both within-subject and cross-subject emotion classificatoin.

Critical strengths
  1. This work involved 12 DE features, which are related to different brain regions, to improve EEG-based emotion recognition and show effective results.
  2. The method had improvement compared to other traditional works and give intrepretability of the decoding.
Critical weaknesses
  1. The method includes too many modules. It's hard to know which parts contribute to the overall performance in practice.
  2. Some concept and language evidence are given, while further explanation and comarison with findings from related works are necessary to make it persuasive.
Questions for authors
  1. How did you choose the EEG concepts, and was there any evidence across subjects found with the current analysis?
  2. It's would be benificial to give a clear ablation or analysis to show which part of the methods really contributed to the overall performance.
  3. How the language inovoled into the model, could it help capture effective information?
  4. Did the spike network perform better that commonly neural networks?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.