acceptodds
Withdrawn from ICLR 2027

Learning EEG Representations via Topology-Aware Privileged Distillation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Audio and visual signals offer rich affective cues for emotion recognition, but they are identity-sensitive and raise privacy concerns at deployment. EEG is an appealing inference modality because it does not directly expose facial appearance or voice content, yet its low signal-to-noise ratio often limits recognition performance. We therefore adopt a cross-subject privileged learning setup, where audio and video are used only during training, enabling an EEG-only model at inference. However, existing cross-modal distillation objectives often overlook the region-level interaction structure inherent in EEG, even though affective processing is reflected in inter-regional dependencies. To address this, we propose Topology-Aware Relational Distillation (TRD): the teacher induces a region-dependency graph from audio-visually refined EEG region features, and the student aligns its induced graph by matching the global edge distribution rather than directly imitating features. Meanwhile, the modality gap remains a barrier because EEG representations can still be semantically misaligned with the teacher’s multimodal decision space. To mitigate this gap, we further introduce Masked Decision-Space Alignment (MDA). MDA injects student EEG region features into a frozen teacher decision pathway with privileged modalities masked, and distills the resulting predictions so that the student can inherit multimodal semantics without requiring privileged modalities at inference. Under cross-subject EEG-only inference on EAV and MDMER, our method consistently improves the same student backbone trained without distillation, yielding up to +2.09 pp accuracy and +4.46 pp Macro-F1.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Knowledge Distillation, Privileged Learning, EEG

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

This work addresses cross-subject EEG-based emotion recognition, incorporating audio and video during training while requiring only EEG at inference. The method uses Topology-Aware Relational Distillation to transfer region-dependency structures from a multimodal teacher to the student, and Masked Decision-Space Alignment to align EEG representations with multimodal semantics. Experiments on EAV and MDMER demonstrate improvements in EEG-only emotion recognition.

Critical strengths

1.The use of audio-visual information to guide EEG learning during training, while requiring only EEG at inference, has practical value. The method aligns the student with the teacher at both the regional-relation level and the decision-space level, potentially helping the student acquire richer emotion-related information.

2.Rather than directly matching teacher and student features, TRD aligns their region-dependency structures. This may reduce the effects of cross-modal feature discrepancies and inter-subject variability on distillation.

Critical weaknesses

1.The experimental evaluation needs to be more comprehensive. For example, why was LaBraM chosen rather than REVE, which was pretrained on more data? Similarly, why was iTransformer selected rather than more recent state-of-the-art sequence models? Dataset coverage is also limited, with no evaluation on other multimodal EEG datasets compatible with this setting, such as MAHNOB-HCI.

  1. The methodological validation needs greater rigor. In the Cross-Modal EEG Refinement module, it is unclear whether the audio-visually refined features genuinely preserve region-specific information. Audio-visual information may simply be broadcast across regions, making them appear similar because of a shared cross-modal component. This could alter existing EEG semantics rather than provide meaningful regional information, and the paper does not appear to include an ablation examining this possibility. Moreover, because TRD explicitly optimizes the student graph to match the teacher graph, the graph alignment shown in Figure 4 is, to some extent, an expected consequence of the training objective. I recommend comparing teacher graphs constructed with and without audio-visual refinement and reporting results across multiple random seeds to strengthen the evidence.

  2. Sections 4.4.1 and 4.4.2 could place less emphasis on reiterating the method and algorithmic design, and provide more analysis of what the model learns and how this supports the claimed benefits.

Questions for authors

1.Equation (2) does not specify the architecture of the Cross-Modal EEG Refinement function (f_{\mathrm{ref}}). How does audio-visual information refine individual EEG regions, and how does the module avoid broadcasting an identical cross-modal component to every region? Could the authors compare teacher graphs before and after refinement to establish whether the gains come from meaningful multimodal relational knowledge?

  1. Is there a standardized data-splitting protocol for the compared methods on these datasets? If not, the authors’ choice of a uniform 6:2:2 or 8:1:1 split requires further explanation, as the results may depend on the particular partition. Please provide more justification for this choice or conduct experiments using multiple splits.

  2. Could the authors clarify why only LaBraM was considered among pretrained EEG foundation models, without including additional models such as REVE,CBraMod? What motivated the selection of iTransformer (2024) as the time-series baseline, and were more recent state-of-the-art alternatives considered? Were other multimodal EEG datasets, such as MAHNOB-HCI, considered alongside EAV and MDMER, and what criteria guided the dataset selection? I consider these issues major concerns regarding the rigor of the experimental evaluation.

  3. Table 4 reports only point estimates, without clarifying whether it uses the same random seeds as the main experiments, and provides neither variability estimates nor significance tests. Given that Shuffled Graph also improves Macro-F1 from 38.57 to 39.28. Therefore, additional experimental results seem necessary to support this conclusion.In particular, the graph alignment shown in Figure 4 is, to some extent, a direct consequence of optimizing the training objective. More convincing evidence is needed to demonstrate that this alignment translates into meaningful benefits from multimodal information.

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

in this work, authors claim that affect using audio-visual can leak user identity. both modalities can be combined with EEG modalities which does not have access to audio-visual features of users. authors therefore propose to use EEG modality as a primary modality while audio-visual as privileged information (PI). the test is done via EEG modality only. a teacher-student framework is used in parallel with knowledge distillation (KD) that distills audio-visual into EEG. authors propose to use graph-based KD, in particular graph edge KD over refined multimodal features into EEG features. they also used logit KD. results are reported over 2 datasets.

Critical strengths
  • overall, the writing is good.
  • the paper provides a form of KD for EEG from audio-visual.
  • results are reported.
Critical weaknesses

1- general direction. the authors claimed that audio and visual modality can leak participant identity in affect applications, therefore, they shouldnt be used at test time. instead, authors propose to use EEG modality during train and test while considering audio-visual modality as privileged information. authors then use KD to distill audio-visual into EEG. the privacy issue is not clear here. is it: a- the identity of the users at TRAINSET can be leaked if one have access to the model. therefore, at training time, the model should not see modalities that can leak user identity. b- the TEST USER (at inference), does not want to provide their audio-visual because it REVEALS their IDENTITY (? but they have or given access to expensive EEG sensors and they are willing to wear them so one can assess their affect???) or their data will be transferred through internet and they are concerned that an unauthorized actor can access to it and learn the test user identity.

if it is the a-case, swapping modalities (by making EEG as main modality and audio-visual a privileged ones) is an absurd choice. more specifically, swapping cheap accessible modalities (audio-visual) with expensive, less accessible modality (EEG) that requires specialized hardware to be worn by user is strange choice. using EEG at test time because privacy issues in case-a is not a good direction. typically, one prefers to design learning methods that: uses audio-visual as primary and EEG as privileged modality while suppressing identity leakage from audio-visual as commonly done. interchanging modalities to end up using the expensive modality at test time is not a good direction.

if it is the b-case, what is a common application that requires all these constraints for affect prediction? it is a real application?

in the paper, it is not very clear which case the authors intended. but, it seems case-a. however, there is no discussion or support in the paper that EEG does not leak user-identity.

all this to say that using EEG as primary (with poor performance in affect tasks) modality and audio-visual as privileged modalities is not justified in this work. the final modal, as designed in this paper, is impractical. this gives no reason for this direction to be pursued, as well as this paper. yes, one can design different KD/PI setups with different configurations. but, they should be practical. as it is, the paper does not provide any clear and direct application to the KD/PI chosen here.

2- method:

  • it is not clear what is a refinement function f_ref presented in sec.3.1. it takes the 3 modalities, and predicts somehow only EEG modality. it seems like a fusion module. but it is not clear why it produces only EEG modality. there is no details about this.
  • sec.3.2: why KD via graph instead of features? this was not justified.
  • sec.3.2: it is not clear/justified why the need for what it is called 'modulation' to compute student region dependency graph over EEG only; nor, why it is done the way it is presented (eq.7,8). these were given as facts.
  • sec.3.2: no justification to edge definition (eq.11) nor to why do KD over graph edge (eq.12).
  • sec.3.2: it is not clear why the KD over graph edges is not enough. why the need to use a second KD over logits? and why the need to use the teacher frozen head? how that helps? and how masking audio-visual inputs into this teacher frozen head 'enables the student to inherit multimodal semantics'? isnt there a concern here that the teacher frozen head will leak identity information since it is used at test time as well as it is part of student decision?
  • sec.3.2 - line-264: "During teacher training, ea (a) and ev (v) are replaced by 0 with a probability": is this mean that suppression occurs randomly (on/off)? what is the probabilities? and how this affects the students at test time where both modalities are off all time?
  • sec.3.2 - line 278: l_cls is not defined. it is the supervised classification loss at the frozen teacher head?
  • fig.2: all the teacher modules should be indicated as frozen. for the student model: the student head defined in the green box should be replaced by frozen teacher as there is no student head. indicate the 3 losses on the figure. l_cls is missing.
  • sec.2: last paragraph: the proposed method is poorly aligned with existing works. it is not clear what novelty it brings in terms of KD.

3- experiments:

  • sec.4.1 - line 297: how N is determined for both datasets to the said values?
  • sec.4.2: tab.1,2: it is not clear how other methods were trained? are their performance is the one reported publicly or did the authors retrain them? also, why 'ours (cls-only)' is better than others over EAV datset in terms of accuracy in tab.1? is the comparison of this method (trained over multimodal) vs others trained with EEG only fair (tab.1)?
  • sec.4.3: the ablation in tab.3.a: ablation should be done for each element separately (MDA, TRD) and not incremental. the effect of each one is not clear. in tab.3.b, the justification of using graph kd over feature kd here seems to stem from the empirical results. there is no method justification.

notations:

  • sec.3, line-151: why X, a, v for modalities? you can use consistent notation.

overall, nor the direction, nor the method in this work are justified.

please stop using ai.

Questions for authors

please see weaknesses.

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

The paper studies cross-subject emotion recognition with privileged modalities: audio and video are available during training, while inference uses EEG only. It proposes Topology-Aware Relational Distillation (TRD) to align teacher–student region-dependency graphs, and Masked Decision-Space Alignment (MDA) to inject student EEG features into a frozen teacher pathway with privileged inputs masked.

Critical strengths

The training design matches the deployment constraint. Privileged modalities are used only for training, and the final inference path contains only the EEG student.

The controlled distillation comparisons and component ablations are useful. Relative to the same Cls-only backbone, the full method improves Accuracy/Macro-F1 by 1.15/2.97 percentage points on EAV and 2.09/4.46 points on MDMER. TRD and MDA also each improve both metrics when used separately.

Critical weaknesses

The “topology-aware” claim is not sufficiently isolated. The graph is a fully connected similarity matrix over predefined electrode groups, but there is no comparison with random electrode-to-region assignments preserving the number and sizes of regions. Therefore, the results establish the benefit of relational distillation more clearly than the benefit of scalp topology.

The evidence for cross-subject generalization is limited. Both datasets use only one fixed subject split, while the three runs measure seed variation rather than variation across subject splits. This matters especially for EAV, which has 42 subjects and shows only a 1.15-point Accuracy gain over Cls-only; no paired significance test is reported.

Questions for authors

Assessment-critical experiment: What performance does TRD obtain with random electrode-to-region assignments that preserve the number and sizes of regions? This would clarify whether the gain comes from the scalp topology or from generic relational distillation.

Assessment-critical experiment: Can the authors report results over multiple independent subject-wise splits or LOSO evaluation, together with paired significance tests between Cls-only and the full method?

Clarification: Will the exact subject IDs in the fixed train/validation/test splits be released? The paper currently reports only the 6:2:2 split for EAV and 8:1:1 split for MDMER.

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

Related papers

Discussion (0)

Sign in to comment.