New Deep Learning Model for Emotion Recognition Based on EEG Data

2019-12-25

The latest advances in machine learning have enabled the development of technologies for detecting and recognizing human emotions. Some of these technologies work by analyzing electroencephalogram (EEG) signals, which essentially record the brain's electrical activity collected from the scalp.


Over the past decade or so, most EEG-based emotion classification methods have employed traditional machine learning (ML) techniques, such as support vector machine (SVM) models, because these models require fewer training samples but still lack large-scale EEG datasets. However, recently, researchers have compiled and released new datasets containing EEG brain recordings.These datasets have opened up exciting new possibilities for EEG-based emotion recognition, as they can be used to train deep learning models that achieve better performance than traditional machine learning techniques. Unfortunately, these datasets have low EEG signal resolution, which may make training deep learning models quite challenging.


To improve the resolution of available EEG data, researchers first generated so-called "topology-preserving differential entropy features" using electrode coordinates during data collection. They then developed a convolutional neural network (CNN) and trained it on the updated data to estimate three general emotion categories (positive, neutral, and negative).


The researchers stated, "Previous methods often overlook the topological information of EEG features, but our approach enhances the EEG representation by learning generated high-resolution EEG images. This approach reorganizes the EEG feature through the proposed CNN, enabling better representation of clustering."


The researchers trained and evaluated their method on the SEED dataset, which contains 62-channel EEG signals. They found that their method achieved a classification accuracy of 90.41%, outperforming other EEG-based emotion recognition machine learning techniques.


The researchers added, "If EEG signals are recorded from different emotional segments, the original DE function will fail to cluster. We will also apply our method to tasks estimating driver vigilance to demonstrate its ready-to-use nature."In the future, the proposed method by the researchers may provide insights for the development of new EEG-based emotion recognition tools, as it introduces a feasible solution to address the issue of low EEG data resolution. The same approach can also be applied to other deep learning models for analyzing EEG data, even those used for classifying human emotions.


For computer vision tasks, large-scale datasets have been used with deep learning models for image classification, achieving tremendous success, some of which have surpassed human performance. Additionally, complex data preprocessing is no longer required. In future work, we hope to use generative adversarial networks (GANs) to generate large-scale EEG datasets.

RECENT AI NEWS

RECENT AI TOOLS