Zero-Weight aware LSTM Architecture for Edge-Level EEG Classification

Abstract

In the electroencephalograph (EEG) brain-computer interface field, the classification of EEG signals with neural networks is an emerging research topic. However, previous works mainly focused on choosing appropriate networks and implementing them in a software fashion. For edge-level EEG classification, this paper proposes a Zero-Weight (ZW) aware long short-term memory (LSTM) based EEG classifier implemented on field-programmable gate array (FPGA). ZW-aware LSTM network optimizes the matrix-vector multiplication (MxV) not only using the sparse weight of the LSTM layer itself but also considering the sparse weights of the following layers. Public BCI competition data are used for the evaluation of the proposed ZW-aware LSTM network EEG classifier. Our hardware-implemented EEG classifier shows ×9.53 speedup compared to the software classifier implemented on CPU @3.40GHz and achieves accuracy up to 81%.