FACE: Few-shot Adapter with Cross-view Fusion for Cross-subject EEG Emotion Recognition

arXiv:2503.18998 · eess.IV, cs.AI, cs.CV · Submitted 2025-03-24 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "FACE: Few-shot Adapter with Cross-view Fusion for Cross-subject EEG Emotion Recognition".

Dev: Here's the gist of the paper. Cross-subject EEG emotion recognition is challenged by significant inter-subject variability and intricately entangled intra-subject variability.

Rosa: This article introduces the few-shot adapter with a crossview fusion method called FACE for cross-subject EEG emotion recognition.

Dev: Problem Context Existing works have primarily addressed these challenges through domain adaptation or generalization strategies.

Rosa: Domain-adaptation methods require large amounts of unlabeled target subject data during training, which restricts their applicability in scenarios with scarce or non-specific target subject data.

Paper summary: Dev: So looking at the paper "FACE: Few-shot Adapter with Cross-view Fusion for Cross-subject EEG Emotion Recognition" and the authors Haiqi Liu, C. L. Philip Chen, and Tong Zhang Senior Member, IEEE, what are we taking away from this work?

Rosa: The core idea is that by using a cross-view fusion module to merge global brain connectivity with local patterns, and then adding the few-shot adapter module for rapid subject-specific emotional feature adjustment, you get a method that works well across different subjects.

Taro: It suggests that we don't have to rely on massive domain adaptation methods or domain generalization techniques when we only have a few labeled samples for a new subject.

Dev: Essentially, it means we can get strong emotion recognition performance even in few-shot scenarios where the target subject data is scarce, because the model is initialized smartly.

Rosa: It moves the focus from needing huge amounts of unlabeled target data during training to learning a subject-agnostic initialization that can adapt quickly.

Taro: The implication for real-world applications, like in a BCI interface or any deployment where subjects change, is that we can deploy systems that are much more robust to those individual differences.

Dev: It shows the framework has strong generalization capability, maintaining a performance level comparable to domain adaptation and generalization based approaches even though it uses different strategies.

Rosa: And the results consistently show that both components of FACE—the cross-view fusion and the adapter structure—are actually needed together for achieving those top scores.

Taro: That’s important because it tells us that neither feature alone is sufficient; you need that combined approach to get the best result.

Conclusion: Rosa: So we've been looking at this paper called FACE by Liu and Chen and Zhang Senior Member, IEEE today, which is all about using a cross-view fusion module and a few-shot adapter for recognizing emotions in EEG data across different people.

Dev: Right, the authors are focusing on tackling that big headache of inter-subject variability where it's really hard to build models that work reliably between subjects.

Taro: What they actually did is propose this joint feature fusion thing—merging global connectivity with local patterns—and then adding this adapter structure to make it fast for new subjects.

Rosa: So the title itself, FACE, suggests a "Face" or a strong connection point for this fusion idea that lets you adapt quickly.

Dev: It's about taking those two views and putting them together in a shared space so the model can handle the subject-specific details better without needing tons of new data.

Taro: And what they show is that this combination actually beats other cross-subject methods, even when you only give it a few labels for the new person.

Rosa: The main implication here is that we might not need those huge databases of labeled EEG from every single subject just to get good emotion recognition out of it.

Dev: It means the initialization phase becomes way more important, because they optimize a starting point that can handle any new subject with just a few steps of adjustment.

Taro: But there are still some caveats—the authors show it works well on SEED and SEED-IV but we gotta see how it handles the really messy, real-world data outside the lab.

Rosa: Exactly, so while this framework is strong for few-shot scenarios right now, the next thing we need to look at is how robust that whole system stays when things get unpredictable.

Dev: We'll talk more about those real-world failure modes next when we discuss how this kind of rapid adaptation actually runs in a live system.

South China University of Technology

eess.IV, cs.AI, cs.CV

Submitted: 2025-03-24

Updated: 2026-11-12

Comments: Early Access Version

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 84/100

The gist: Cross-subject EEG emotion recognition is challenged by significant inter-subject variability and intricately entangled intra-subject variability.

Key concepts

Cross-View Fusion (CVF)
This module integrates two different ways of looking at the EEG data: one extracting global brain connectivity patterns and another identifying region-specific emotional information. It projects these two views into a shared space and fuses them using meta-learning to create a unified representation, improving feature quality.
Few-Shot Adapter (FSA)
The FSA module uses an adapter structure combined with meta-learning to allow the model to quickly adjust its emotional features when presented with only a few labeled samples from a new subject. This enables rapid, subject-specific calibration without requiring extensive training for every new person.
Meta-Learning
Meta-learning is used during the meta-training phase to optimize parameters that allow the model to learn how to adapt quickly. It helps create a subject-agnostic parameter initialization, meaning the model starts with a strong foundation that can be rapidly fine-tuned for any new subject using minimal labeled data.
Cross-Subject Variability
This refers to the differences in EEG emotion recognition performance that arise because subjects are different individuals. These differences stem from variations in brain structure and cognitive processes, making it hard for models trained on one group to perform well on a completely new subject.

Terminology

Summary

Cross-subject EEG emotion recognition is challenged by significant inter-subject variability and intricately entangled intra-subject variability. p. 0 p. 0 This article introduces the few-shot adapter with a crossview fusion method called FACE for cross-subject EEG emotion recognition. p. 0 p. 0 p. 0

Problem Context

Existing works have primarily addressed these challenges through domain adaptation or generalization strategies. p. 0 Domain-adaptation methods require large amounts of unlabeled target subject data during training, which restricts their applicability in scenarios with scarce or non-specific target subject data. p. 0 In contrast, domain-generalization methods often struggle to generalize to unseen target domains due to their dependence on source data diversity. p. 0 Few-shot learning approaches attempt to address this issue by leveraging the FSL method to learn a whole subject-agnostic initialization that can quickly adapt to new subjects using only a few labeled samples. p. 0 However, the scarcity of labeled EEG data from target subjects often leads to catastrophic overfitting when performing whole model parameter updates with limited training steps. p. 0

Proposed Architecture Components

The FACE framework comprises two key components: a cross-view fusion (CVF) module and a few-shot adapter (FSA) module. The CVF module extends the conventional GCN view by introducing a cross-view architecture, where the GCN view extracts global connectivity patterns and a complementary local view identifies region-specific emotional information. p. 1 This module projects two view representations into a shared space and performs subject-specific complementary fusion with meta-learning. p. 1 The FSA module incorporates the adapter structure with the meta-learning paradigm to enable rapid calibration for unseen subjects. p. 1

Methodology Details

The model is optimized by minimizing the cross-entropy loss as follows: min θ′ E(x,y)∼Dst que [L(fθ′ (x), y)]. The dynamic GCN processes X through two-layers GCN to derive high-level representations which can be formally expressed as H = BN (Θ2 ∗ σ (BN (Θ1 ∗ X))). p. 3 The cross-view fusion mechanism captures cross-view dependencies Z△ between Zfs and Zg through a multi-head self-attention mechanism. p. 4 The few-shot adapter enhances conventional residual adapters through strategic integration of batch normalization layers to statistically regularize inter-subject variability. p. 4

Experimental Validation

Comprehensive evaluations on three widely used benchmarks (SEED, SEED-IV, and SEED-V) demonstrate that FACE outperforms state-of-the-art cross-subject EEG emotion recognition methods. Table I presents a comprehensive comparison of various methods on SEED and SEED-IV datasets. p. 6 FACE achieves an impressive accuracy of 96. p. 6 p. 672% on SEED and 95. p. 1 p. 195% on SEED-IV under the 10-shot setting. p. 5 p. 5

Key Findings

The CVF module effectively combines global brain connectivity and local patterns, utilizing meta-learned cross-view fusion to dynamically provide a unified representation. The FSA module incorporates the adapter structure and meta-learning paradigm to enable rapid subject-specific emotional feature adjustment for unseen subjects. p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 Comprehensive evaluations on three widely used benchmarks demonstrate that FACE outperforms state-of-the-art cross-subject EEG emotion recognition methods. p. 1 The proposed framework yields a subject-agnostic parameter initialization through Ne episodes of bi-level optimization during the meta-training phase. p. 4 The model achieves an emotion classification accuracy of up to 98% when fine-tuning with approximately 10 samples per class from a new subject. p. 8

Future Directions

We anticipate this work will inspire further research into few-shot cross-subject BCI interface tasks. Future investigations will focus on few-shot multimodal EEG emotion recognition and extend the few-shot learning paradigm to broader affective BCI applications. p. 2 The results consistently highlight the superiority and strong generalization of the proposed method in few-shot learning scenarios. p. 7 p. 7 The SEED-V dataset achieves superior accuracy and exhibits relatively low variance, which may be attributed to its higher data quality. p. 8 The number of attention layers does not produce significant performance variations across the datasets and shot conditions. p. 9 The proposed framework demonstrates efficient adaptation capability through its metalearned parameter initialization. p. 5 p. 5 The model requires a few labeled samples from the target subject to perform rapid parameter adaptation via inner-loop optimization. p. 5 The results demonstrate that while semi-supervised methods benefit from unlabeled data, our approach achieves exceptional performance even with limited labeled data. p. 7 The optimal results are achieved when both CVF and FSA modules are employed. p. 7 p. 7 p. 7 The CVF module demonstrates superior performance by effectively capturing complementary multi-view features, whereas selfattention and cross-attention mechanisms exhibit performance degradation due to inter-view feature distribution incompatibility. p. 8 The FSA module enables more effective subject-specific calibration, resulting in superior performance improvements. p. 8 The consistency of activated regions under the same emotions across different datasets validates the effectiveness of the FSA module. p. 8 The model demonstrates competitive performance with state-of-the-art methods. p. 6 The proposed method maintains comparability with domain adaptation/generalization-based approaches. p. 5 The proposed framework demonstrates efficient adaptation capability through its metalearned parameter initialization.

--- Page 1 ---

. This article introduces the few-shot adapter with a crossview fusion method called FACE for cross-subject EEG emotion recognition. Cross-subject EEG emotion recognition is challenged by significant inter-subject variability and intricately entangled intra-subject variability. Existing works have primarily addressed these challenges through domain adaptation or generalization strategies. However, they typically require extensive target subject data or demonstrate limited generalization performance to unseen subjects. p. 0 Recent few-shot learning paradigms attempt to address these limitations but often encounter catastrophic overfitting during subject-specific adaptation with limited samples. p. 0 This article introduces the few-shot adapter with a crossview fusion method called FACE for cross-subject EEG emotion recognition.

--- Page 2 ---

. The ability of models trained on one group to generalize effectively to unseen subjects is the ability of models trained on one group to generalize effectively to unseen subjects. p. 0 Existing studies are mainly dedicated to addressing the inter-subject variance through domain-adaptation and domaingeneralization strategies. p. 0 Domain-adaptation methods learn subject-specific representations by aligning the marginal distributions of source and target subjects. p. 0 Although these methods have achieved notable results, they require large amounts of unlabeled target subject data during training. p. 0 In contrast, domain-generalization methods attempt to learn subject-invariant representations from source data. p. 0 However, they often struggle to generalize to unseen target domains due to their dependence on source data diversity. p. 0

--- Page 3 ---

. Inter-subject variability in EEG emotion recognition mainly arises from individual differences in brain structure and cognitive processes. p. 2 Recently, Domain Adaptation (DA) and Domain Generalization (DG) techniques have emerged as promising solutions to mitigate the impact of inter-subject variance. p. 2 DA methods aim to bridge the distributional gaps between source and target subjects to improve the performance of the target subject. p. 2 DG methods attempt to learn generalized models that perform well across multiple domains without relying on explicit adaptation to specific target domains. p. 2

--- Page 4 ---

. In the cross-subject EEG emotion recognition task, each subject can be considered a distinct domain due to substantial inter-subject variance. p. 2 The core challenge arises from domain shifts between source and target subjects, characterized by P (X Ss)! p. 2= P (X St), where Ss ∈ Ssource. p. 2 The few shot crosssubject emotion recognition extends this task by imposing a strict few-shot constraint: For target subject St, the model must perform few-steps adaptation utilizing only a small support set Dsup t = (xi, yi) k×c i=1. p. 2

--- Page 5 ---

. To obtain stronger feature representation and mitigate the effects of intra-subject variability, this article proposes the cross-view fusion module which effectively integrates both global and spatial-local patterns. p. 3 p. 3 The aligned representation Zfs is then concatenated with Zg to form the joint representation Zc = [Zfs; Zg]. p. 4 To reduce the effect of inter-subject variance in view fusion, we reconceptualize the multi-head self-attention parameters as subject-agnostic meta parameters optimized via modelagnostic meta-learning. p. 4

--- Page 6 ---

. The overall architecture of the proposed FACE framework is illustrated in Fig. p. 2 2. The CVF module effectively combines global brain connectivity and local patterns, utilizing meta-learned cross-view fusion. p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 The FSA module incorporates the adapter structure and meta-learning paradigm to enable rapid subject-specific emotional feature adjustment for unseen subjects.

--- Page 7 ---

. To obtain stronger feature representation and mitigate the effects of intra-subject variability, this article proposes the cross-view fusion module which effectively integrates both global and spatial-local patterns. The unified representation Zu is obtained through channelwise concatenation: Zu = [Zfs; Zg]. p. 4

--- Page 8 ---

. The model is optimized by minimizing the cross-entropy loss as follow: Lcls = − 1/N X N i=1 Xc j=1 yi,j log (ˆyi,j). p. 3 The proposed framework yields a subject-agnostic parameter initialization Θ′ all through Ne episodes of bi-level optimization during the meta-training phase. p. 4

--- Page 9 ---

. Across different shot settings, incorporating the CVF or FSA module alone generally improves performance compared to the baseline. p. 7 The optimal results are achieved when both CVF and FSA modules are employed. The consistency of multiple asterisks across all datasets confirms that both CVF and FSA components contribute substantially to performance improvement. p. 7 The optimal results are achieved when both CVF and FSA modules are employed.

--- Page 10 ---

. FACE achieves an impressive accuracy of 96.72% on SEED and 95.95% on SEED-IV under the 10-shot setting. The results consistently outperform all competing approaches in the 5-shot setting. p. 6 Under the 1-shot setting, the proposed method already achieves competitive performance. p. 7 In the 3-shot setting, the proposed approach surpasses all other approaches. p. 7 The model is optimized by minimizing the smooth cross-entropy loss as follow: Lcls = − 1/N X N i=1 Xc j=1 yi,j log (ˆyi,j). p. 3 p. 3

--- Page 11 ---

. The proposed framework demonstrated remarkable capability in capturing subtle emotion-discriminative patterns while mitigating overfitting risks under data scarcity constraints. p. 0 Specifically, FACE integrated a cross-view fusion module to merge heterogeneous EEG views into a unified representation and a few-shot adapter module to facilitate rapid, subject-specific emotional feature adjustment. p. 1 The model is optimized by minimizing the smooth cross-entropy loss as follow: Lcls = − 1/N X N i=1 Xc j=1 yi,j log (ˆyi,j).

--- Page 12 ---

. The complete training procedure is formalized in Algorithm 1. p. 4 When encountering novel target subjects, only k-shot labeled EEG samples per class need to be collected as the support set Dsup t to perform rapid parameter adaptation via inner-loop optimization. p. 5 The proposed framework demonstrates efficient adaptation capability through its metalearned parameter initialization Θ′ all. p. 5 When k = 3, the proposed framework achieves performance comparable to state-of-the-art methods. p. 5 The results consistently highlight the superiority and strong generalization of the proposed method in few-shot learning scenarios.

--- Page 13 ---

. This article presents FACE, a novel method to enhance cross-subject EEG emotion recognition through subject-specific adaptation via joint feature fusion and adjustment. p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 The CVF module effectively combines global brain connectivity and local patterns, utilizing meta-learned cross-view fusion. The FSA module incorporates the adapter structure and meta-learning paradigm to enable rapid subject-specific emotional feature adjustment for unseen subjects.

--- Page 14 ---

. This article presents FACE, a novel method to enhance cross-subject EEG emotion recognition through subject-specific adaptation via joint feature fusion and adjustment. The CVF module effectively combines global brain connectivity and local patterns, utilizing meta-learned cross-view fusion. The FSA module incorporates the adapter structure and meta-learning paradigm to enable rapid subject-specific emotional feature adjustment for unseen subjects.

--- Page 15 ---

. This article presents FACE, a novel method to enhance cross-subject EEG emotion recognition through subject-specific adaptation via joint feature fusion and adjustment. The CVF module effectively combines global brain connectivity and local patterns, utilizing meta-learned cross-view fusion. The FSA module incorporates the adapter structure and meta-learning paradigm to enable rapid subject-specific emotional feature adjustment for unseen subjects.

--- Page 16 ---

. This article presents FACE, a novel method to enhance cross-subject EEG emotion recognition through subject-specific adaptation via joint feature fusion and adjustment. The CVF module effectively combines global brain connectivity and local patterns, utilizing meta-learned cross-view fusion. The FSA module incorporates the adapter structure and meta-learning paradigm to enable rapid subject-specific emotional feature adjustment for unseen subjects.

--- Page 17 ---

. This article presents FACE, a novel method to enhance cross-subject EEG emotion recognition through subject-specific adaptation via joint feature fusion and adjustment. The CVF module effectively combines global brain connectivity and local patterns, utilizing meta-learned cross-view fusion. The FSA module incorporates the adapter structure and meta-learning paradigm to enable rapid subject-specific emotional feature adjustment for unseen subjects.

--- Page 18 ---

. This article presents FACE, a novel method to enhance cross-subject EEG emotion recognition through subject-specific adaptation via joint feature fusion and adjustment. The CVF module effectively combines global brain connectivity and local patterns, utilizing meta-learned cross-view fusion. The FSA module incorporates the adapter structure and meta-learning paradigm to enable rapid subject-specific emotional feature adjustment for unseen subjects.

--- Page 19 ---

. This article presents FACE, a novel method to enhance cross-subject EEG emotion recognition through subject-specific adaptation via joint feature fusion and adjustment. The CVF module effectively combines global brain connectivity and local patterns, utilizing meta-learned cross-view fusion. The FSA module incorporates the adapter structure and meta-learning paradigm to enable rapid subject-specific emotional feature adjustment for unseen subjects.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Cross-subject EEG emotion recognition can be performed on unseen subjects using only a small support set of labeled data by employing a few-shot adapter module with meta-learning to enable rapid subject-specific emotional feature adjustment.

  2. The system can dynamically generate a unified representation from heterogeneous views by using the crossview fusion module, which projects spatial features into the semantic space of graph features and captures crossview dependencies through a multihead selfattention mechanism.

  3. The AI system can preserve individualized emotional features and statistically regularize inter-subject variability by enhancing conventional residual adapters with batch normalization layers and optimizing adapter parameters jointly with multihead selfattention parameters via bi-level optimization.

Related papers