Gated Recurrent Neural Networks for EMG-Based Hand Gesture Classification. A Comparative Study

Electromyographic activities (EMG) generated during contraction of upper limb muscles can be mapped to distinct hand gestures and movements, posing them as a promising modality for prosthetic and cybernetic applications. This paper presents a comparative analysis between different recurrent neural network (RNN) configurations for EMG-based hand gesture classification. In particular, RNNs with recurrent units of long short-term memory (LSTM) and gated recurrent unit (GRU) are evaluated. Furthermore, the effects of an attention mechanism and varying learning rates are evaluated. Results show a classifier 1) with a bidirectional recurrent layer composed of LSTM units, 2) that applies the attention mechanism, and 3) trained with step-wise learning rate outperforms all other tested RNN classifiers.