Chaotic thinking: New neural network model recognizes human emotion from EEG data

Intuitive to anyone using a virtual assistant on their phone or computer, today’s interactions between humans and machines tend to occur without any exchange of emotions. One approach using electroencephalography (EEG), a common way to record electrophysiological function in the brain, has the potential to provide an accurate way to identify human emotion but is fraught with seemingly chaotic brain signal and low signal-to-noise ratios. New work leverages the power of chaos theory to parse through EEG signals and identify human emotions.

Yang et al. demonstrated a new channel-frequency convolutional neural network that, when combined with recurrence quantification analysis of nonlinear time series, can identify emotions of those monitored with EEGs in a lab environment more than 92.24 percent of the time. The technique feeds EEG data comprising of 30 different channels analyzed by recurrence quantification analysis into the convolutional neural network to classify patterns of signals as happiness, sadness and fear.

Several oscillations in EEG activity have characteristic frequency ranges and spatial distributions. Entropy measurements from one such oscillation occurring from 31-50 hertz, the gamma wave, were found to correspond to emotion identification most prominently among the five bands the group investigated.

The authors compared their approach to other techniques that combine power spectral density or differential entropy with support vector machines to classify emotion, which yielded accuracies of 72.59 and 82.75 percent, respectively. The group’s convolutional neural network also boasted a kappa value of 0.884, measureing how little of the classification was random.

Authors Yuxuan Yang and Zhong-Ke Gao said that the group hopes that their findings open the door for refining other forms of human-machine interactions, such as detecting fatigue.

Source: “A recurrence quantification analysis-based channel-frequency convolutional neural network for emotion recognition from EEG,” by Yu-Xuan Yang, Zhong-Ke Gao, Xin-Min Wang, Yan-Li Li, Jing-Wei Han, Norbert Marwan, and Juergen Kurths, Chaos (2018). The article can be accessed at https://doi.org/10.1063/1.5023857 .