Study Of The Multi-categorization Capability Of EEG Signals

The recognition of object categories is easily accomplished in everyday life,yet its neural basis is still not fully understood.Humans have demonstrated a superior performance in interpreting visual scenes that is beyond the reach of machines.Although the rediscovery of convolutional neural networks in recent years has led to significant improvements in the performance of automatic visual classification,their ability to generalise has not reached the level of humans.In the last few years,advances in artificial intelligence and neuroscience have made it possible for brain activity to interact with computers and other devices.In particular,advances in various signal processing methods such as EEG,combined with AIdriven algorithms,have enabled us to delve into the world of brain-computer interfaces.EEG signals in brain-computer interfaces are often used in the study of classification tasks,and the multi-classification capabilities of EEG hold a wide range of promise for practical applications.Indeed,a number of reliable methods exist for solving the problem of decoding visual object-related EEG data,but most of these methods are designed primarily for binary classification(e.g.the presence or absence of a certain object class).It is not always this simple classification task of presence and absence in practical application scenarios.However,research on multi-category EEG decoding is very limited.Therefore,this paper addresses a series of studies on the EEG signal multiclassification problem.In this paper,there are four main steps in the EEG signal multi-classification system,namely EEG signal acquisition,EEG signal pre-processing,EEG signal feature extraction,and EEG signal classification.In terms of EEG signal acquisition,the EEG equipment used in this study is the Acti CHamp EEG system provided by Brain products,Germany.In terms of EEG signal pre-processing,standard filtering,re-referencing and principal component analysis operations were used in this paper.For EEG signal feature extraction,spatial filtering was used.For EEG signal classification,five common classifier systems were selected,including two machine learning methods and three deep learning methods.In this paper,based on the existing data of the eight classifications of rhesus monkeys,human EEG signals were collected with the same experimental paradigm to explore the difference between human EEG signal classification ability and that of rhesus monkeys.In this paper,firstly,EEG signals in the V4 and IT regions of rhesus monkeys were studied,and an average classification accuracy of 52.87% was obtained in the V4 region of rhesus monkeys and84.19% in the IT region.Secondly,a study of eight classifications of human EEG signals was conducted,ensuring the same experimental paradigm as in rhesus monkeys.However,the results of the experiments showed that humans are far less capable of EEG classification than rhesus monkeys.For humans,this experimental paradigm has too short a dwell time per picture to produce valid categorisation information.In addition,although the paradigm for this experiment was identical to the best effort control experiment,the acquisition equipment was fundamentally different.The rhesus monkeys acquired EEG signals through an array of electrodes implanted in the scalp,whereas the human EEG signals were acquired through a non-invasive electrode device,which was more stable and had better signal quality than the non-invasive human EEG signal acquisition.At present,there are still controversies about the multi-classification ability of EEG signals at home and abroad.To address the problem of quadruple classification of EEG signals based on visual perception,three comparison experiments were done in this paper based on different stimulus image datasets in order to investigate the effects of different datasets on the classification results.To investigate the effect of different time periods on the classification results,the data were segmented at 200 ms intervals for classification studies and detailed classification results are given.To demonstrate the deep model training process,visualisations of the model training process are given for all three experiments.To quantify this issue of whether the corresponding representations of the brain are the same across subjects,the paper uses a representation dissimilarity matrix and a multidimensional scaling method to analyse the representation differences between subjects.The results show that the early time after stimulus images start to be presented is the main source of categorization information generation,that colour increases EEG categorization information and improves categorization accuracy,and that the context facilitates subjects' processing of implicit categorization tasks.There have been many studies on EEG signal classification based on visual perception,but fewer studies based on visual cognition.To explore the EEG classification ability under cognitive tasks,two additional comparison experiments were conducted in this paper based on the presence or absence of background information.The results of the experiments show that background information facilitates subjects to produce category-specific brain responses.In addition to this,the background increases EEG categorisation information and improves categorisation accuracy,but the enhancement is limited.