AI summary · DOI: 10.64898/2026.11.05.764629 · bioRxiv: Neuroscience

Probing EEG Foundation Model Attention Maps for Lightweight Interpretable Emotion Recognition

L. N. Arvanitopoulos, A. Dragomir, C. A. Fidas, A. Bezerianos
Original paper Published on 06 Nov. 2026 bioRxiv preprint License: CC BY 4.0

How to read this summary? Illustrated guide

The essentials

The paper uses REVE attention maps as lightweight connectivity features for emotion recognition, achieving competitive performance with much lower computational cost.

Based on the abstract only: check the details in the original paper.

Question

Can attention maps from a pretrained EEG foundation model provide interpretable, low-cost features for emotion recognition without fine-tuning? The study also asks whether a label-free symmetry criterion can identify informative attention heads.

Method

The authors aggregate raw first-layer attention weights from individual REVE heads across the full duration of SEED-IV trials into channel-to-channel adjacency matrices, used instead of encoder representations. They rank heads with a label-free symmetry-based criterion and train classical machine-learning classifiers using the top-ranked head; the abstract reports comparisons with fully fine-tuned EEG foundation models but gives no exact scores or dataset counts.

Findings

  • Result Attention-derived connectivity features achieved balanced accuracy and F1 scores competitive with several fully fine-tuned EEG foundation models.
    See the passage in the paper
    Classical machine learning classifiers are then trained on the single most informative head's attention-derived FC matrix and achieve balanced accuracy and F1 scores on the SEED-IV emotion-recognition dataset that are competitive with several fully fine-tuned state-of-the-art EEG Foundation Models, requiring at least two orders of magnitude fewer calculations than the standard use of the encoder.
  • Result Using attention-derived matrices required at least two orders of magnitude fewer calculations than standard encoder use.
    See the passage in the paper
    Classical machine learning classifiers are then trained on the single most informative head's attention-derived FC matrix and achieve balanced accuracy and F1 scores on the SEED-IV emotion-recognition dataset that are competitive with several fully fine-tuned state-of-the-art EEG Foundation Models, requiring at least two orders of magnitude fewer calculations than the standard use of the encoder.
  • Result The most discriminative connectivity edges showed a significant over-representation of frontal-region connections.
    See the passage in the paper
    Subsequent cortical region analysis revealed a significant over-representation of frontal region connections by ranking the most discriminative FC matrix connectivity edges, validating our methods as consistent with prior affective-computing literature.
  • Authors’ interpretation The authors propose linear probing of attention maps as a resource-efficient route to deploying EEG foundation models.
    See the passage in the paper
    These results suggest that interpretable, computationally inexpensive linear probing of attention maps is a viable route to deploying EEG Foundation Models in resource-constrained settings.

Limitations

  • The evidence described is limited to emotion recognition on the SEED-IV dataset, so generalization to other tasks or datasets is not established in the abstract.
  • The reported approach relies on the pretrained REVE model and does not establish whether the same results hold for other EEG foundation models.

Why it matters

The authors argue that attention maps can serve both as interpretable connectivity features and as a computationally inexpensive alternative to standard encoder representations. This could make EEG foundation models more practical in resource-constrained settings.

Written by AI from the paper. Passages are checked word for word against the text we analysed; the rest may contain errors.

AI-generated keywords: EEG foundation models attention maps emotion recognition functional connectivity computational efficiency

Read further

  • Comprehensive Summary
  • Layman's Summary
  • Blog article
  • Original abstract
The paper investigates whether attention maps from a pretrained EEG foundation model can serve as useful, interpretable features for emotion recognition without the computational cost of running or fine-tuning the model’s full encoder. The authors focus on REVE and the SEED-IV emotion-recognition dataset. Their motivation is that, although EEG foundation models aim to learn representations that generalize across people and datasets, they have not consistently outperformed task-specific methods. Their large representations can also be expensive to compute and store. The proposed approach uses attention weights as a lightweight proxy for weighted functional connectivity. The authors extract raw attention weights from a single head in the model’s first layer, aggregate them over each trial, and construct weighted adjacency matrices linking EEG channels. They use these matrices as classification features instead of the model’s usual encoder representations, without fine-tuning the foundation model. To select a useful head without relying on labels, the authors propose a symmetry-based ranking criterion. Motivated by near-symmetric attention patterns observed in trained encoder-decoder transformers, the criterion aims to identify heads whose attention contains more discriminative information. Classical machine-learning classifiers are then trained using the matrix from the selected head. On SEED-IV, the resulting classifiers achieve balanced accuracy and F1 scores that the authors describe as competitive with those of several fully fine-tuned EEG foundation models. The method also requires at least two orders of magnitude fewer calculations than standard use of the encoder. The abstract does not provide exact score values, identify the specific classifiers compared, or detail the computational measurements. The authors also examine which connections are most discriminative. Their cortical-region analysis finds a significant over-representation of connections involving frontal regions. They interpret this finding as consistent with prior affective-computing research and as evidence of the proposed features’ interpretability. The study concludes that attention maps from a pretrained EEG model can be repurposed as compact features for emotion recognition. This approach may offer a practical alternative when computational resources are limited, retaining task-relevant information in matrices that are easier to inspect than large encoder representations. Avoiding fine-tuning and substantially reducing computation could make the method useful for lightweight probing or deployment of EEG foundation models. However, the available information is limited to the title and abstract. It does not report details needed to assess the strength and scope of the results, including dataset splits, participant-level evaluation, uncertainty estimates, statistical tests of classification performance, exact scores, or comparisons with other lightweight baselines. The reported results concern SEED-IV and do not establish generalization to other datasets, subjects, or EEG tasks. The abstract also describes the symmetry-based head-selection criterion only briefly, leaving its implementation and robustness unclear. Overall, the paper presents a promising approach to efficiency and interpretability, but the available information is insufficient to independently assess its quantitative performance or broader generalizability.

Was this synthesis useful?

Score: 0

Why do we need votes?

Votes are used to determine whether we need to re-run our summarizing tools. If the count reaches -10, our tools can be restarted.

Ask questions about this paper to our AI assistant

How to chat with this paper? Illustrated guide

You can also chat with multiple papers at once here.

Figures and plots

Ask about a figure by number or use its question button to request visual analysis. Up to two figures and four image panels per question. Values read from plots are approximate; check the original data.

AI assistant instructions?

Tools for this paper

Explain it differently 1 free rewrite left today
Save to a collection Group papers in collections to compare them, chat with them and export them. Continue with Google Create a free account Log in

Advanced tools Premium

For researchers who follow a field closely: citation alerts, journal-club slides, version comparison and Zotero export. Discover Premium

Zotero Zotero guide
Citation alerts Citation alerts guide
Who works on it?

Among the 23 papers around this one, the name that comes back most: Philipp Thölke (2 papers)

2 more people sign several of them. The full map adds their institutions and approaches, from the papers of the last years.

Continue with Google Create a free account: your first map is free
Weekly letter The latest letter of its field, "Brain and behaviour this week" of 16 November 2026: Brain and behaviour this week: timing, repair and the origins of attraction · Receive one by email for free

Papers around this one

See them on a map

Ranked by how close their title and abstract are to this paper’s. Select some to read them together.

Closest papers

  1. Very close Review Cited by this paper
    Foundation Models for Neural Signal Decoding: EEG‐Centered Perspectives Toward Unified Representations
    Jii Kwon, You-Min Shin · 2025 · European Journal of Neuroscience · cited 5 times · Not summarized yet

    This perspective surveys EEG-centered foundation models for neural-signal decoding, including their generalization and interpretability challenges.

    Abstract

    Neural signals such as EEG, ECoG, and intracortical recordings offer a valuable window into brain dynamics but remain difficult to decode due to high dimensionality, nonstationarity, and substantial interindividual variability. Traditional machine learning and deep learning models often show limited generalizability and insufficient interpretability in these settings. Foundation models (FMs)—large‐scale architectures pretrained on diverse datasets—have recently emerged as a promising paradigm for building robust, transferable, and physiologically grounded neural representations. Among these modalities, EEG currently serves as the most practical and representative platform for FM development …

  2. Very close Broader context
    Representational and Functional Robustness to Electrode Montages in EEG Foundation Models
    J. Steglich, Justus Meyer zu Bexten, Shakiba Moradi et al. · 2026 · arXiv.org · preprint, not peer reviewed · cited 1 time · Not summarized yet

    This study analyzes how electrode montages affect EEG foundation-model representations and predictions, rather than probing attention for emotion recognition.

    Abstract

    EEG foundation models (EEG-FMs) are intended to generalize across different datasets by learning representations that, ideally, are invariant to dataset-specific EEG configurations such as electrode montages. However, EEG-FMs that accept different montages as input do not guarantee that representations and predictions remain stable across different electrode configurations, especially outside the training setting. In this work, we investigate the effects of different electrode montages through a joint functional and representational analysis of four EEG foundation models selected to span distinct montage-handling designs. We evaluate embeddings on cross-subject resting-state eyes-open/closed…

  3. Very close Groundwork for this paper Cited by this paper Used in depth by this paper
    REVE: A Foundation Model for EEG - Adapting to Any Setup with Large-Scale Pretraining on 25,000 Subjects
    Y. El Ouahidi, Jonathan Lys, Philipp Thölke et al. · 2025 · Neural Information Processing Systems · cited 150 times · Not summarized yet

    REVE is the pretrained EEG foundation model whose first-layer attention maps this paper uses as emotion-recognition features.

    Abstract

    Foundation models have transformed AI by reducing reliance on task-specific data through large-scale pretraining. While successful in language and vision, their adoption in EEG has lagged due to the heterogeneity of public datasets, which are collected under varying protocols, devices, and electrode configurations. Existing EEG foundation models struggle to generalize across these variations, often restricting pretraining to a single setup, resulting in suboptimal performance, in particular under linear probing. We present REVE (Representation for EEG with Versatile Embeddings), a pretrained model explicitly designed to generalize across diverse EEG signals. REVE introduces a novel 4D positi…

  4. Very close Review
    A Survey on Attention-Based Deep Learning Methods for EEG Emotion Recognition
    Zhi-Xuan Qi · 2026 · Applied and Computational Engineering · Not summarized yet

    This survey reviews channel, temporal, and self-attention methods for EEG emotion recognition.

    Abstract

    Electroencephalography (EEG)-based emotion recognition has attracted increasing attention in recent years because of its potential applications in brain-computer interfaces, intelligent healthcare, and affective computing. With the rapid development of deep learning, attention mechanisms have gradually become an important component of EEG emotion recognition models, since they are able to improve feature representation by focusing on more informative EEG signals. This paper reviews recent attention-based deep learning methods for EEG emotion recognition, with particular attention to channel attention, temporal attention, and self-attention implemented in Transformer-based architectures. Thei…

  5. Close Another approach to the problem
    Multi-Scale Attention-Enhanced BiLSTM Network for EEG Emotion Recognition
    Yong-Bin Zhang, Run-Yi Li, Xiu-Li Fu et al. · 2026 · IEEE Access · Not summarized yet

    MS-BiLSTM-Attention recognizes EEG emotions with a lightweight recurrent attention network, rather than classifiers on pretrained-model attention maps.

    Abstract

    Over the last decade, significant progress has been made in detecting human emotions via electroencephalographic (EEG) recordings. Nevertheless, contemporary EEG-driven affect recognition techniques continue to encounter multiple hurdles, including insufficient feature extraction capability, inadequate modeling of local patterns, low feature extraction efficiency, high model complexity, and difficulty in practical deployment. To address these issues, we introduce an EEG-specific lightweight architecture for affective computing titled multi-scale Bidirectional Long Short-Term Memory with Attention (MS-BiLSTM-Attention). This model uses DE features, which are embedded into the EEG electrode to…

  6. Close Another approach to the problem
    Multi-scale feature fusion and neural architecture search for electroencephalography-based affective-state recognition in public benchmark datasets
    Chi-Xun Guan, Yun-Han Wang · 2026 · Frontiers in Medicine · Not summarized yet

    MFANet classifies EEG affective states using multiscale feature fusion and neural architecture search, rather than attention-derived connectivity features.

    Abstract

    Electroencephalography (EEG)-based affective-state recognition may provide an objective neurophysiological complement to self-report scales and behavioral observations, but robust modeling remains difficult because EEG signals are non-stationary, temporally heterogeneous, and strongly subject-dependent. This study evaluated MFANet, a multi-scale feature fusion and neural architecture search-enhanced framework for binary classification of valence, arousal, and dominance. MFANet combines parallel one-dimensional convolutional branches with different receptive fields, channel-wise attention for adaptive feature recalibration, and a neural architecture search-guided long short-term memory module…

  7. Close Same method, another problem
    Lightweight Semantic EEG Foundation Model for Frozen Cross-Disorder Transfer
    Rita Huan-Ting Peng, Nhat Bui · 2026 · Not summarized yet

    USE-FM freezes a pretrained EEG encoder for downstream transfer, while this paper uses frozen REVE attention maps for emotion recognition.

    Abstract

    Large-scale EEG foundation models have demonstrated promising transferability across neurological disorders, but often require millions of parameters and substantial computational resources. In this paper, we present the Universal Semantic EEG Foundation Model (USE-FM), a lightweight EEG foundation model that learns transferable neural representations through self-supervised signal reconstruction on the Temple University Hospital EEG Corpus (TUEG). After pretraining, the encoder is frozen and evaluated on two clinically distinct downstream tasks, abnormal EEG detection (TUAB) and epileptic seizure recognition (TUEP), using a unified frozen-transfer protocol against recent EEG foundation mode…

  8. Close Broader context Cited by this paper
    CoMET: A Contrastive-Masked Brain Foundation Model for Universal EEG Representation
    Ang Li, Zikai Wang, Liu-Yin Yang et al. · 2025 · cited 8 times · Not summarized yet

    CoMET develops a universal EEG foundation model, a neighboring approach to the pretrained model whose attention this paper probes.

    Abstract

    Electroencephalography (EEG) is a non-invasive technique for recording brain activity, widely used in brain-computer interfaces, clinic, and healthcare. Traditional EEG deep models typically focus on specific dataset and task, limiting model size and generalization. Recently, self-supervised brain foundation models have emerged and been applied to various downstream tasks. Nevertheless, these models still have limitations: current SOTA models typically rely on masked reconstruction strategy; however, EEG features of adjacent channels are highly correlated, which causes the pre-training to overly focus on low-dimensional signal-similarity features in local regions and neglect the global discr…

Prior work it builds on

  1. Close Broader context
    AdaBrain-Bench: Benchmarking Brain Foundation Models for Brain-Computer Interface Applications
    Jia-Min Wu, Zichen Ren, Jun-Yu Wang et al. · 2025 · arXiv.org · preprint, not peer reviewed · cited 22 times · Not summarized yet

    AdaBrain-Bench benchmarks brain foundation models for BCI applications, providing a broader evaluation context for this paper's EEG-model comparison.

    Abstract

    Non-invasive Brain-Computer Interfaces (BCI) offer a safe and accessible means of connecting the human brain to external devices, with broad applications in home and clinical settings to enhance human capabilities. However, the high noise level and limited task-specific data in non-invasive signals constrain decoding capabilities. Recently, the adoption of self-supervised pre-training is transforming the landscape of non-invasive BCI research, enabling the development of brain foundation models to capture generic neural representations from large-scale unlabeled electroencephalography (EEG) signals with substantial noises. However, despite these advances, the field currently lacks comprehens…

  2. Close Review
    Systematic review of self-supervised foundation models for brain network representation using electroencephalography
    Hannah Portmann, Y. Morishima · 2026 · cited 1 time · Not summarized yet

    This systematic review surveys self-supervised EEG foundation models that learn whole-brain representations.

    Abstract

    Automated analysis of electroencephalography (EEG) has recently undergone a paradigm shift. The introduction of transformer architectures and self-supervised pretraining (SSL) has led to the development of EEG foundation models. These models are pretrained on large amounts of unlabeled data and can be adapted to a range of downstream tasks. This systematic review summarizes recent SSL-trained EEG foundation models that learn whole-brain representations from multichannel EEG rather than representations derived from a single channel. We searched PubMed, IEEE Xplore, Scopus, and arXiv through July 21, 2025. Nineteen preprints and peer-reviewed articles met inclusion criteria. We extracted infor…

  3. Close Broader context
    CBraMod: A Criss-Cross Brain Foundation Model for EEG Decoding
    Jiquan Wang, Sha Zhao, Zhiling Luo et al. · 2024 · International Conference on Learning Representations · cited 322 times · Not summarized yet

    CBraMod proposes a general EEG foundation model for decoding, whereas this paper studies interpretable emotion features from REVE attention.

    Abstract

    Electroencephalography (EEG) is a non-invasive technique to measure and record brain electrical activity, widely used in various BCI and healthcare applications. Early EEG decoding methods rely on supervised learning, limited by specific tasks and datasets, hindering model performance and generalizability. With the success of large language models, there is a growing body of studies focusing on EEG foundation models. However, these studies still leave challenges: Firstly, most of existing EEG foundation models employ full EEG modeling strategy. It models the spatial and temporal dependencies between all EEG patches together, but ignores that the spatial and temporal dependencies are heteroge…

  4. Close Broader context
    CSBrain: A Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding
    Yuchen Zhou, Jia-Min Wu, Zichen Ren et al. · 2025 · Neural Information Processing Systems · cited 87 times · Not summarized yet

    CSBrain develops a cross-scale EEG foundation model for decoding, rather than probing an existing model's attention maps.

    Abstract

    Understanding and decoding brain activity from electroencephalography (EEG) signals is a fundamental challenge in neuroscience and AI, with applications in cognition, emotion recognition, diagnosis, and brain-computer interfaces. While recent EEG foundation models advance generalized decoding via unified architectures and large-scale pretraining, they adopt a scale-agnostic dense modeling paradigm inherited from NLP and vision. This design neglects a core property of neural activity: cross-scale spatiotemporal structure. EEG task patterns span a wide range of temporal and spatial scales, from short bursts to slow rhythms, and from localized cortical responses to distributed interactions. Ign…

  5. Close Same question
    EmotionMeter: A Multimodal Framework for Recognizing Human Emotions
    Wei-Long Zheng, Wei Liu, Yifei Lu et al. · 2019 · IEEE Transactions on Cybernetics · cited 1,233 times · Not summarized yet

    EmotionMeter recognizes human emotions using EEG and eye movements, while this paper targets emotion recognition from EEG-derived attention features.

    Abstract

    In this paper, we present a multimodal emotion recognition framework called EmotionMeter that combines brain waves and eye movements. To increase the feasibility and wearability of EmotionMeter in real-world applications, we design a six-electrode placement above the ears to collect electroencephalography (EEG) signals. We combine EEG and eye movements for integrating the internal cognitive states and external subconscious behaviors of users to improve the recognition accuracy of EmotionMeter. The experimental results demonstrate that modality fusion with multimodal deep neural networks can significantly enhance the performance compared with a single modality, and the best mean accuracy of 8…

How are these papers found?

We compare the title and abstract of this paper with those of every paper summarized on this site, and with the papers Semantic Scholar recommends, those it cites and those citing it (for a much-cited paper, its most recent citations and, from OpenAlex, the most cited ones). Among the prior work and the follow-ups, the papers that use each other in depth, a method it uses and the much-cited follow-ups come a little higher. The levels (very close, close, related) come from the similarity of the texts, not from a judgment of quality; * marks a paper without an abstract, compared by its title. The sentence under a paper is written by an AI from the two abstracts: check it in the papers. A paper marked not relevant by three readers leaves the list.

Updated on 20 November 2026.

The illustrated guide to the papers around a paper

Disclaimer: The AI-based summarization tool and virtual assistant provided on this website may not always provide accurate and complete summaries or responses. We encourage you to carefully review and evaluate the generated content to ensure its quality and relevance to your needs.