Toward reliable signals decoding for electroencephalogram: A benchmark study to EEGNeX

The development of brain-computer interfaces (BCI) has facilitated our study of mental representations in the brain. Neural networks (NNs) have been widely used in BCI due to their decent pattern learning capabilities; however, to our best knowledge, a comprehensive comparison between various neural network models has not been well addressed, due to the interdisciplinary difficulty and case-based study in the domain. Here, we tested the capabilities of common NN architectures in deciphering mental representations from electroencephalogram (EEG) signals, which were recorded in representative classification tasks. In this study, we: 1. Construct 20 mechanism-wise different, typical NN types and their variants on decoding various EEG datasets to show a comprehensive performance comparison regarding their EEG information representation capability. 2. Lighten an efficient pathway based on the analysis results to gradually develop general improvements and propose a novel NN architecture: EEGNeX. 3. We open-sourced all models in an out-of-the-box status, to serve as the benchmark in the BCI community. The performance benchmark contributes as an essential milestone to filling the gap between domains understanding and support for further interdisciplinary studies like analogy investigations between the brain bioelectric signal generation process and NN architecture. All benchmark models and EEGNeX source code is available at:https://github.com/chenxiachan/EEGNeX.