Giugliano, Salvatore (2022) Machine Learning and XAI methods for improving EEG-based BCI classification systems. [Tesi di dottorato]

Preview

Text Giugliano_Salvatore_XXXV.pdf Download (11MB) | Preview

Collection description

The use of Machine Learning (ML) techniques for EEG signal classification is gaining increasing attention in Brain-Computer interfaces (BCI) applications thanks to promising performances reported by many ML systems, from one side, and the non-invasiveness and high time resolution of the EEG acquisitions from the other one. However, several EEG-based BCI applications suffer the main drawbacks of the EEG signals, such as their non-stationarity, which makes the employing systems particularly sensitive to changes in users or time acquisitions. Performance with different acquisition times or subjects remains low in several applications. Therefore, such systems can be unreliable, particularly when used in safety-critical domains. From the ML point of view, the non-stationarity of EEG signals can be viewed as an instance of the well-known Dataset Shift problem, where, differently from the ML standard hypothesis, training and test data can belong to different probability distributions, leading ML systems toward poor generalisation performances. The research work of this PhD thesis was conducted with the long-term goal of exploiting the knowledge from eXplainable Artificial Intelligence (XAI) domain to develop EEG-based classification systems which overcome the performance returned by the current ones. XAI methods try to explain the behaviour of AI systems, such as ML ones, by providing explanations about the response of an AI system, given a specific input, in terms of relevant input features. More specifically, the contribution of this PhD thesis is threefold: firstly, a study on BCI systems that relied on EEG signals is made, leading to two different proposals for two different tasks: EEG-based emotion recognition and SSVEP classification. These proposals explore advanced ML techniques such as convolutional neural networks and domain adaptation methods on well-known EEG datasets. Secondly, a study on modern XAI methods is made, converging toward a new method to build explanations in an image classification task. Finally, on the basis of the results obtained in the previous investigations, an experimental analysis of explanations produced by several XAI methods on an ML system trained on EEG data for emotion recognition is made. Preliminary results suggest the plausibility to develop ML methods for BCI systems able to leverage on XAI methods to generalise across different subjects and different times without further efforts.

Downloads

Actions (login required)

View Item