EEG-based Emotion Classificationusing Deep Belief Networks

Key Lab. of Shanghai Education Commission forIntelligent Interaction and Cognitive Engineering

Among the approaches to emotion recognition, methods based on electroencephalogram (EEG) sig-nals are more reliable because of its high accuracy and objective evaluation compared to other external appearance such as facial expression and gesture. However due to the low signal to noise ratio (SNR), it is often very hard to analyze EEG signals ‘by hand’even for neurophysiologists. Recent developing deep learning in machine learning community allows au-tomated feature extraction and feature selection and eliminates the limitation of hand-crafted feature.

gamma frequency bands do exist. Second, we show that differential entropy (DE) features extracted from EEG data possess accurate and stable information for emotion classification. Finally, the paper com-pares the results between deep modals and shallow models like KNN, SVM and GELM. Moreover, DBN-HMM performs well when compared with the state-of-the-art classification methods.

Deep Belief Networks (DBN) is a probabilistic gen-erative model with deep architecture, which charac-terizes the input data distribution using hidden vari-ables. A DBN is constructed by stacking a predefined number of restricted Boltzmann machines (RBMs) on top of each other where the output from a lower-lev-el RBM is the input to a higher-level RBM.

The graphical depiction of DBN

Wei-Long Zheng, Jia-Yi Zhu, Yong Peng, and Bao-Liang Lu

FeatureExtraction

Raw Data

SVM

FeatureExtraction

Raw Data

DBN

HMM

FeatureExtraction

Raw Data

DBN

(b) (d) (e)

FeatureExtraction

Raw Data

GELM

(c)

FeatureExtraction

Raw Data

KNN

(a)

Overview of the five setups for EEG-based emotion classification used in this work

Session K-2 Session K-1 Session K Session K+1 Session K+2

Hint of start

Movie clip Rest

10 sec 4 min 20 sec

Positive Emotion

Positive Emotion

Positive Emotion

Negative Emotion

Negative Emotion

Protocol of the EEG experiment

...

...

...

500 1000 1500 2000 2500

50

100

150

200

250

300

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

Time

Del

taTh

eta

Alp

haB

eta

Gam

ma

Freq

uenc

y ba

nds

There exist specific neural patterns in beta and gamma bands for positive and negative emotions

Delta Theta Alpha Beta Gamma Total0

10

20

30

40

50

60

70

80

90

100

Frequency Bands

Acc

urac

y (%

)

GELMSVMDBNDBN−HMM

Introduction

Method

Conclusion In this paper, we introduce recent advanced deep learning models to EEG-based emotion classifica-tion. The main contributions of this paper are as fol-lows: First, we find that neural signatures associated with positive and negative emotions in beta and

IEEE International Conference on Multimedia & Expo (ICME), 2014