Trustworthy Transfer Learning And Its Applications In Brain-Computer Interfaces

Transfer learning is a significant branch of machine learning that aims to improve the model’s ability to generalize on the test set with different distributions.A brain-computer interface(BCI)provides a direct pathway for interaction between the brain and external devices,making it possible to study,enhance,or restore human cognitive and sensory-motor functions.However,current BCI applications face challenges of individual variability and privacy leakage.Transfer learning can help address the issue of individual variability in BCIs,but it also raises two concerns that affect its trustworthiness: negative transfer and privacy leakage.Negative transfer occurs when using data or knowledge from the source domain deteriorates the performance of the target domain model? privacy leakage happens when the source domain data is accessed or transmitted directly during the transfer learning process,which exposes the subject’s private information.This dissertation aims to address the issues of untrustworthiness associated with negative transfer and privacy leakage in transfer learning,and conducts research on trustworthy transfer learning theory and algorithms for brain-computer interfaces.The main contributions are summarized as follows:To address the lack of quantitative standards of negative transfer and privacy leakage theories in transfer learning,the relevant theories were clarified in terms of definitions,detection standards,etc.For the theory of negative transfer,the causes of negative transfer,the definitions of negative transfer in unsupervised and semi-supervised scenarios,the detection standards of negative transfer under different target data,and the strategies for constructing negative transfer tasks were given.For the theory of privacy leakage,theoretical definitions and detection standards of privacy leakage under various forms of source knowledge were proposed.Experiments on motor imagery and emotion recognition paradigms verified that the proposed theories could be utilized to check the trustworthiness of existing transfer learning algorithms and guide the design of trustworthy transfer learning algorithms.To address the issue of negative transfer in brain-computer interface-based transfer learning algorithms,this dissertation delved profoundly and comprehensively into data quality,distribution divergence,and algorithm design to cover the causes of negative transfer,and proposed a variety of negative transfer mitigation algorithms.At the data quality level,a source transferability estimation strategy was proposed,which was based on the idea of meta-learning.It constructed multiple meta-features in the training stage and improved the quality of source domain data through source weighting or source selection.At the distribution divergence level,a discriminative joint probability distribution divergence metric was proposed,which was based on the joint probability distribution decomposition theory.It considered the transferability and discriminability when calculating the distribution divergence to improve the trustworthiness.At the algorithm design level,a manifold embedded knowledge transfer algorithm was proposed,which was based on the idea of staged transfer and introduced several strategies,such as centroid alignment,knowledge transfer,and source domain selection,to alleviate negative transfer.Experiments on multiple brain-computer interface paradigms,such as motor imagery,emotion recognition,and event-related potentials,demonstrated that the proposed algorithms outperformed existing ones and effectively mitigated negative transfer.To address the issue of privacy leakage in brain-computer interface-based transfer learning algorithms,privacy-preserving transfer learning in centralized and distributed braincomputer interface data scenarios were investigated.In the centralized scenario,a lightweight source-free transfer algorithm was proposed,which constructed a virtual intermediate domain based on the source domain model interface.It could transfer knowledge while protecting the privacy of the source subjects.In the distributed scenario,a multi-source decentralized transfer algorithm was proposed,which was based on strategies such as parameter transfer and knowledge distillation.It could process both white-box and black-box source domain model inputs.Experiments on motor imagery and emotion recognition paradigms demonstrated that the proposed algorithms outperformed existing ones and effectively protected the source subject’s privacy information,such as identity and physiological signals.In conclusion,this dissertation has explored the theory and algorithm of negative transfer mitigation and privacy protection in transfer learning.Its aim was to enhance the generalization and privacy of transfer learning algorithms and brain-computer interface systems in practical applications.Experiments on various brain-computer interface paradigms have demonstrated that the proposed approaches could achieve high transfer accuracy and enhance the trustworthiness of the transfer learning process,which would be beneficial to develop accurate and privacy-preserving brain-computer interface systems.