Toward the application of XAI methods in EEG-based systems Andrea Apicella Francesco Isgr o Andrea Pollastro Roberto Prevete

2025-04-24 0 0 2.3MB 19 页 10玖币

侵权投诉

Toward the application of XAI methods in

EEG-based systems

Andrea Apicella, Francesco Isgr`o, Andrea Pollastro, Roberto Prevete

Laboratory of Augmented Reality for Health Monitoring (ARHeMLab)

Laboratory of Artiﬁcial Intelligence, Privacy & Applications (AIPA Lab)

Department of Electrical Engineering and Information Technology, University of Naples Federico II

Abstract

1An interesting case of the well-known Dataset Shift Problem is the

classiﬁcation of Electroencephalogram (EEG) signals in the context of

Brain-Computer Interface (BCI). The non-stationarity of EEG signals can

lead to poor generalisation performance in BCI classiﬁcation systems used

in diﬀerent sessions, also from the same subject. In this paper, we start

from the hypothesis that the Dataset Shift problem can be alleviated by

exploiting suitable eXplainable Artiﬁcial Intelligence (XAI) methods to

locate and transform the relevant characteristics of the input for the goal

of classiﬁcation. In particular, we focus on an experimental analysis of

explanations produced by several XAI methods on an ML system trained

on a typical EEG dataset for emotion recognition. Results show that many

relevant components found by XAI methods are shared across the sessions

and can be used to build a system able to generalise better. However,

relevant components of the input signal also appear to be highly dependent

on the input itself.

Keywords: BCI, XAI, EEG, Dataset, Shift cross-session

In this research work, we experimentally investigate the performances of

several well-known eXplainable Artiﬁcial (XAI) methods proposed in the liter-

ature in the context of Brain-Computer Interface (BCI) problems using EEG

input-based Machine Learning (ML) algorithms to evaluate the possibility of al-

leviating the Dataset Shift problem. This is not a trivial issue as, diﬀerently from

other signals, the non-stationarity of EEG signals makes them hard to analyse.

In recent years, Brain-Computer Interfaces (BCIs) have been emerging as tech-

nology allowing the human brain to communicate with external devices without

the use of peripheral nerves and muscles, enhancing the interaction capability of

the user with the environment. In particular, several proposals of BCI methods

based on Electroencephalographic (EEG) signals are receiving growing interest

by the scientiﬁc community thanks to its implication in medical purposes [5, 8],

1accepted to be presented at XAI.it 2022 - Italian Workshop on Explainable Artiﬁcial

Intelligence. Please refer to the published version at https://ceur-ws.org/Vol-3277/paper1.

pdf

arXiv:2210.06554v4 [cs.LG] 18 May 2024

other than other ﬁelds such as entertainment [15], education [4], and marketing

[16]. This is because measuring and monitoring the brain’s electrical activity can

provide important information related to the brain’s physiological, functional,

and pathological status. EEG signals are particularly suitable to this aim thanks

to their important qualities such as non-invasiveness and high temporal reso-

lution [30]. Furthermore, several solutions for comfortable and wearable EEG

acquisition devices are being proposed [10, 7], allowing an acquisition process

less inﬂuenced by noise due to the user-device interaction. Thanks to its prop-

erties, the EEG signal is one of the most promising candidates to become one

of the most used communication channels between man and machine.

Several BCI solutions adopting ML methods are proposed in the literature.

Generally, EEG data acquired from persons subjected to well-known stimuli are

used in the training stage. These data are labelled following some established

protocol, usually dependent on the task. For example, in an Emotion Recogni-

tion (ER) task, stimuli can be images or videos considered able to elicit partic-

ular emotions. Therefore the labels can be inferred by the stimuli or declared

by the subject, who will say whether or not he felt a speciﬁc emotion during

the stimulus administration. If the training stage is successful, the model can

generalise on new unlabelled data, such as new acquisition from another subject

or the same subject in another session.

However, one of the main defects of the EEG signal is that its statistical

characteristics change over time. This implies that even under the same condi-

tions and for the same task, signiﬁcantly diﬀerent signals can be acquired just

as time passes. It is important to highlight that this phenomenon can also oc-

cur using the same stimuli-reaction (e.g., same emotions with the same stimuli)

to the same subject at diﬀerent times, leading to substantially diﬀerent EEG

signals even for the same subject. This problem is even more present among

diﬀerent subjects, who, given the same stimuli and emotions, can produce very

diﬀerent acquisitions between them. For these reasons, EEG is considered a

non-stationary signal [12]. More in detail, the following cases in an EEG-based

task can arise: i) a model trained on a set of EEG data acquired from a given

subject at a speciﬁc time could not work on data acquired from the same sub-

ject at diﬀerent times (Cross-Session generalisation problem), or ii) a model

trained on data acquired from one or more subjects should not work as ex-

pected in classifying EEG signals acquired from a diﬀerent subject at diﬀerent

times (Cross-Subject generalisation problem).

This type of problem can be treated as an instance of the Dataset Shift

problem [20]. In a nutshell, Dataset Shift arises when the distribution of the

training data diﬀers from the data distribution used outside of the training stage

(that is, running or evaluation stages); therefore the standard ML assumption

[20] to have the same data distribution for both training and test set does not

hold. Consequently, standard ML approaches can produce ML systems which

exhibit poor generalisation performances.

On another side, a sub-ﬁeld of Artiﬁcial Intelligence, eXplainable Artiﬁcial

Intelligence (XAI), wants to explain the behaviour of AI systems, such as ML

ones. In general, an explanation gives information on why an ML model returns

prediction

ML system

examines

XAI

method

EEG

relevance

Feature

extractor

features

feature evaluation

Figure 1: A general functional scheme of a Machine Learning (ML) architecture

based on XAI methods to select and transform relevant input features with the

aim of improving the performance of ML systems in the context of the dataset-

shift problem.

an output given a speciﬁc input. In particular, several XAI methods applied to

Deep Neural Networks are giving promising results [1, 17, 22, 2].

In the XAI context, several explanations are built by inspecting the model’s

inner mechanism to understand the input-output relationships, assigning a rel-

evance score to each input component. However, building an explanation is

particularly challenging if the model to inspect is a DNN; this is mainly for two

reasons: i) DNNs oﬀer excellent performances in several tasks, but at the price

of high inner complexity of the models, leading toward low interpretability, ii)

to help the ML user to understand the system behaviours, typical explanations

have to be humanly understandable.

The general idea of this work is that outputs’ explanations of a trained

ML model on given inputs can help the setup of new models able to over-

come/mitigate the dataset shift problem, in general, and to generalise across

subjects/sessions in case of EEG signals, in particular.

More speciﬁcally, in this work, we focus on how several well-known XAI

methods proposed in literature behave in explaining decisions made by an ML

system based on EEG input features (Fig. 1). Notice that several current XAI

methods are usually tested on datasets, such as image and text recognition

datasets [22, 6], where the domain shift problem is slight or not present. There-

fore, this work is a ﬁrst step toward a long term goal consisting in exploiting

explanations made by XAI methods to locate and transform the main charac-

teristics of the input for each given output, and to build ML systems able to

generalise toward diﬀerent data coming from diﬀerent probability distributions

(in this context, sessions and subjects). To this end, in this paper, we evaluate

and analyse the explanations produced by a set of well-known XAI methods on

an ML system trained on data taken from SEED [34], a public EEG dataset

for an emotion classiﬁcation task. The results obtained show, on oneside, that

only some well-known XAI methods produce reliable explanations in the EEG

domain in the analysed task. On another side, it is shown that the relevant

components found in the training data can only be partially used on data ac-

quired outside of the training stage. Notably, many relevant components found

in the training data are still relevant across the sessions.

The paper is organised as follows: In Section 1, a brief description of the

related works is reported. In Section 2 the proposed evaluation framework is

presented. In Section 3 the obtained results are discussed. Finally, in Section 4

is devoted to ﬁnal remarks and future developments.

1 Related works

In general, Modern ML approaches, as Deep learning, are characterised by a

lack of transparency of their internal mechanisms, making it not easy for the

AI scientist to understand the real reasons behind the inner behaviours. In

this case, the relationships of the classiﬁed emotion with the EEG input are of-

ten challenging to understand. In the EEG-based applications, works based on

simple features selection strategies to choose the best EEG features are widely

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

TowardtheapplicationofXAImethodsinEEG-basedsystemsAndreaApicella,FrancescoIsgr`o,AndreaPollastro,RobertoPreveteLaboratoryofAugmentedRealityforHealthMonitoring(ARHeMLab)LaboratoryofArtificialIntelligence,Privacy&Applications(AIPALab)DepartmentofElectricalEngineeringandInformationTechnology,University...

展开>> 收起<<

Toward the application of XAI methods in EEG-based systems Andrea Apicella Francesco Isgr o Andrea Pollastro Roberto Prevete.pdf

共19页,预览4页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Toward the application of XAI methods in EEG-based systems Andrea Apicella Francesco Isgr o Andrea Pollastro Roberto Prevete

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: