Abstract:
Accurate classification of visual objects from Single-Trial EEG signals is a challenging task due to the low signal-to-noise ratio (SNR) associated with the brain signals. Recently, machine learning frameworks based on deep neural networks have shown great potential. Network architectures have grown increasingly complex with sophisticated modules to achieve the state-of-the-art performance. Unfortunately, finding the optimal network configuration is a tedious trial-and-error process. In this work, we propose to use a wider version of the simple 1D - CNN architecture with residual connections for EEG based visual object recognition. Experimental results establish that this fairly simple architecture outperforms existing techniques across five different classification tasks. Comprehensive ablation studies analyze the sensitivity of the model to varying parameters, especially the width of the network. We further showcase the features extracted for different classes using t-SNE plots, and demonstrate the superior discriminating quality of suitable network configuration through representational dissimilarity analysis.
