Description

Electroencephalogram (EEG) characteristics that underlie emotional processes in the brain include high temporal resolution and asymmetries in spatial activations. We suggest TSception, a multi-scale convolutional neural network that can identify emotions from EEG, in order to understand the temporal dynamics and spatial asymmetries of EEG towards accurate and widespread emotion detection. Dynamic temporal, asymmetric spatial, and high-level fusion layers make up TSception, which concurrently learns discriminative representations in the time and channel dimensions. The dynamic temporal layer, which learns the dynamic temporal and frequency representations of the EEG, is made up of multi-scale 1D convolutional kernels whose lengths are linked to the EEG sampling rate. The asymmetric spatial layer learns the discriminative global and hemispheric representations by making use of the asymmetric EEG patterns for emotion. Using a high-level fusion layer, the learnt spatial representations will be combined. Using the ensemble algorithm, we incorporate emotion recognition using EEG in this method.