HCNN：用于EEG情绪识别的层次卷积神经网络

187 浏览量更新于2024-08-29 收藏 1.39MB PDF 举报

"Hierarchical Convolutional Neural Networks for EEG-Based Emotion Recognition" 这篇研究论文探讨了使用层次卷积神经网络（Hierarchical Convolutional Neural Networks，HCNN）在基于脑电图（Electroencephalogram，EEG）的情绪识别中的应用。传统的机器学习方法在处理EEG情绪识别时往往存在过拟合问题。作者通过组织不同通道的差异熵特征作为二维映射来训练HCNN，这种做法保留了电极空间拓扑的信息。在实验中，他们将HCNN与堆叠自动编码器（Stacked Autoencoder，SAE）、支持向量机（Support Vector Machine，SVM）和K近邻算法（K-Nearest Neighbors，KNN）作为对比方法。结果显示，HCNN获得了最高的准确性，而SAE的表现略逊一筹，但都明显优于传统的浅层模型，如SVM和KNN。这表明深度学习模型在处理复杂、高维度的EEG数据时具有显著优势。此外，论文指出高频波段的贝塔（Beta）和伽马（Gamma）波段是进行情绪识别最合适的频段，这为EEG情绪识别提供了新的视角。为了深入了解HCNN的特征变换过程，作者还可视化了HCNN的隐藏层，揭示了特征沿着层次结构如何逐步转换和抽象，这有助于理解模型如何从原始EEG信号中提取情绪相关的特征。通过这种层次化的方法，HCNN能够有效地捕获EEG信号中的时空模式，从而提高情绪识别的精确度。这项工作为基于EEG的情绪识别技术提供了新的方向，并强调了深度学习模型在生物信号处理中的潜力，尤其是在神经科学和情感计算领域。

there was still a lot of diversity among different people and

circumstances [40]. Transfer learning attempts to solve this

problem. In emotion recognition, transfer learning makes use

of the knowledge of the source domains (previous subjects) to

help improve the classification accuracy in the target domain

(the present subject) [41]. A recent study [20] offers us a

comprehensive survey on the transfer learning in BCIs. In

the traditional models, the most popular algorithms are based

on the common spatial pattern (CSP) [42–44], which aims at

finding an invariant subspace to project all subjects’ data to. In

this way, the shared part of different subjects’ data is pre-

served, and the difference is compensated. To our knowledge,

seldom works have been conducted concerning the transfer

methods in aBCIs. Lu et al. personalized the emotion classifier

trained on the previous subjects to the present person with

transductive parameter transfer (TPT) [39]. However, the

transfer methods are combined with shallow models, e.g.,

SVM. In our work, we will discuss the most popular transfer

method in DL, i.e., the fine-tune technique [45].

In this paper, we adopt DE to characterize EEG signals as it

has been proven to be suitable for emotion decoding in previ-

ous studies [10, 25, 26]. In order to maintain the information

contained in the positional relatio nship between the elec-

trodes, we organize data from different channels as 2-D sparse

maps to train HCNN classifiers on Delta, Theta, Alpha, Beta,

and Gamma bands. We implement the recognition models

with SVM and KNN and compare them with DL. We also

compare the performance of two deep models (HCNN and

SAE) in emotion decoding. We analyze the classification ac-

curacy across frequency bands to locate the critical bands. To

evaluate the transfer capability of HCNN parameters among

different peoples with fine-tune technique, we define three

training strategies. In the end, we visualize the evolution pro-

cess of representations along the hierarchical structures in 2-D

maps with PCA technique and discuss the information-

processing mechanism inside HCNN.

The rest of this paper is organized as follows: BMethods^

introduces the core methods we use. BExperiments^ describes

the processes in detail. BResults^ is the evaluation of the ex-

perimentation. BDiscussion^ is the analysis of the results and

BConclusion^ is the completion section of the study.

Methods

In this section, we first propose the general framework of

HCNN-based emotion recognition system, and then introduce

each module in sequence.

The General Framework for Emotion Recognition

The HCNN-based emotion recognition system is shown in

Fig. 1. The typical aBCI is a closed-loop system [46], where

feedbacks are involved. Emotion recognition is the key module

in aBCI. For illustration, we draw feedback in the figure. The

visual, together with auditory stimuli, is used to evoke the sub-

ject’s emotion, and EEG signals are recorded on 62 electrodes.

After some standard preprocessing procedures, we compute DE

features on all channels at certain time interval and organize

them as 2-D maps. After the sparse operation, we use HCNN

to process the input maps and classify emotion states.

Preprocessing

The raw EEG data are downsampled to 200-Hz sampling rate.

The signals are visually checked and the recordings seriously

contaminated by EMG and EOG are removed manually. With

the help of EOG recordings, the blink artifacts are located and

removed. In order to filter the noise and remove the artifacts,

the EEG data are processed with a bandpass filter between 0.3

and 50 Hz.

Short-Time Fourier Transform and Differential

Entropy

Fourier transform (FT) is often used to analyze the frequency

configuration of a time-domain signal, and it is widely used in

EEG decomposition. However, the FT operation assumes that

brain wave activities are stationary, which is a false hypothesis

apparently. Therefore, the time series should be cut into small

time segments, and within each segment, the brain electric

activities are approximately considered as stationary. The idea

is called the short-time Fourier transform (STFT). STFT de-

composes a function of time (in our case, EEG signal) into the

frequencies that make it up at fixed time intervals. The calcu-

lation formulation of STFT is as follows:

X τ; ωðÞ¼∫

∞

−∞

xtðÞω t−τðÞe

−jωt

dt; ð1Þ

where x(t) is the original signal and ω(t) is the window func-

tion. The Hanning window (shown in (2), a discrete version) is

a linear combination of modulated rectangular windows, and

it usually emerges in applications that require low aliasing and

less spectrum leakage.

ω nðÞ¼

1−cos

2πn

N−1



; ð2Þ

where n is the window length and N is the sampling number.

DE [25] is referred to as the extension of the Shannon entropy

hxðÞ¼−∑

i¼1

ðÞlog px

ðÞðÞ: ð3Þ

DE is the continuous version of Shannon entropy and the

original computing formulation could be written as the following:

ðÞ

¼ −∫ fx

ðÞ

log fx

ðÞðÞ

dx: ð4Þ

Cogn Comput

剩余12页未读，继续阅读

weixin_38693476

粉丝: 1
资源: 949

HCNN：用于EEG情绪识别的层次卷积神经网络

2019-MINCUT POOLING IN GRAPH NEURAL NETWORKS-网文-rrrr1

A Survey of the Recent Architectures of Deep Convolutional Neural Networks.pdf

Multimodal MRI-based classification of Alzheimer's disease using hierarchical convolutional neural networks下载

Hierarchical-Recurrent-Neural-Networks-for-Speech-Bandwidth-Extension:论文编号

Hierarchical-Double-Attention-Neural-Networks-for-Sentiment-Classification:分层双注意力神经网络在情感分类中的应用

A hierarchical ontology context model for work-based learning

2019-IJCAI-自动所-Hierarchical Graph Convolutional Networks for Sem

Hierarchical Deep Click Feature Prediction for Fine-grained Image Recognition

Hierarchical Neural Networks for Image Interpretation

Hierarchical Bilinear Pooling for Fine-Grained Visual Recognition解读

最新资源