Source-free Subject Adaptation for EEG-based Visual Recognition (2301.08448v1)

Abstract: This paper focuses on subject adaptation for EEG-based visual recognition. It aims at building a visual stimuli recognition system customized for the target subject whose EEG samples are limited, by transferring knowledge from abundant data of source subjects. Existing approaches consider the scenario that samples of source subjects are accessible during training. However, it is often infeasible and problematic to access personal biological data like EEG signals due to privacy issues. In this paper, we introduce a novel and practical problem setup, namely source-free subject adaptation, where the source subject data are unavailable and only the pre-trained model parameters are provided for subject adaptation. To tackle this challenging problem, we propose classifier-based data generation to simulate EEG samples from source subjects using classifier responses. Using the generated samples and target subject data, we perform subject-independent feature learning to exploit the common knowledge shared across different subjects. Notably, our framework is generalizable and can adopt any subject-independent learning method. In the experiments on the EEG-ImageNet40 benchmark, our model brings consistent improvements regardless of the choice of subject-independent learning. Also, our method shows promising performance, recording top-1 test accuracy of 74.6% under the 5-shot setting even without relying on source data. Our code can be found at https://github.com/DeepBCI/Deep-BCI/tree/master/1_Intelligent_BCI/Source_Free_Subject_Adaptation_for_EEG.

• The paper introduces a source-free adaptation method that eliminates access to raw EEG source data by leveraging pre-trained model parameters.
• It employs a classifier-based generative approach and subject-independent feature learning to simulate source data and adapt to new subjects.
• Experimental results on EEG-ImageNet40 yield a top-1 accuracy of 74.6% under 5-shot conditions, demonstrating its practical value.

Source-free Subject Adaptation for EEG-based Visual Recognition

This paper introduces a novel approach to EEG-based visual recognition that addresses the limitations faced in traditional subject adaptation frameworks. The core innovation lies in the proposed "source-free subject adaptation" methodology, which eliminates the need for access to raw source subject data during model training. This is particularly crucial given the privacy concerns and logistical constraints associated with transmitting personal EEG data.

Problem Context and Motivation

The research explores brain-computer interfaces (BCI), focusing on EEG-based visual stimuli recognition systems. These systems must adapt to new users (target subjects) without extensive training data. Traditional methods rely on abundant source subject data, which is often impractical to obtain due to privacy regulations. Therefore, this paper proposes an alternative where only pre-trained model parameters from source subjects are utilized.

Methodology

The new problem setup, "source-free subject adaptation," relies on the following key strategies:

1. Pre-Trained Model Utilization: The proposed framework uses only the pre-trained model weights from source subjects for adaptation. Raw source subject data remains inaccessible.
2. Classifier-Based Data Generation: The paper introduces a generative approach that synthesizes EEG samples mimicking source subject data using classifier responses. This approach allows the framework to simulate the source data distribution without actual data access.
3. Subject-Independent Feature Learning: With synthesized data, the framework performs feature learning that leverages commonalities across different subjects, facilitating adaptation to the target subject's limited data.

Experimental Validation

The model is evaluated using the EEG-ImageNet40 benchmark. Results indicate that the proposed method achieves a top-1 test accuracy of 74.6% under the 5-shot setting even without reliance on source data. Notably, the adaptability and generalizability of the framework allow it to integrate various subject-independent learning methods, consistently improving performance irrespective of the chosen approach.

Implications and Future Directions

The results underscore the potential of source-free subject adaptation to overcome privacy and data availability constraints in EEG-based recognition systems. By focusing on classifier responses and generative modeling, the authors illustrate a viable path for deploying BCI technologies in environments where data accessibility is a critical concern.

Moving forward, the framework could be expanded to explore more sophisticated generative models, potentially incorporating architectures like GANs or VAEs to enhance the diversity and realism of synthesized EEG data. Furthermore, integration with other physiological signals might offer a more comprehensive understanding of brain activity patterns, enhancing recognition accuracy and robustness.

In conclusion, this research makes a significant contribution to the field of BCI by proposing a practical, privacy-conscious approach to EEG-based subject adaptation, laying the groundwork for future advancements in adaptive learning models within constrained environments.