Decoding Velocity from Spikes Using a New Architecture of Recurrent Neural Network

Abstract

Motor brain-machine interfaces (BMIs) usually collect neuronal activity from the dorsal premotor cortex (PMd) and primary motor cortex (M1). As the pattern of interconnections between PMd and M1 is complex, current BMI decoders directly decode PMd and M1 neural signals without considering the interconnections between these cortical areas. In this paper, a new architecture of recurrent neural network (Double Recurrent Neural Network, DRNN) was proposed, which took the interconnection information between PMd and M1 into account. To evaluate the performance of DRNN, we recorded the spike data and the position of a robotic arm when a rhesus monkey performed one-dimensional robotic arm reach task. When DRNN decoder offline decoded the velocity of the robotic arm, it showed high decoding accuracy of 0.92 ± 0.04 (correlation coefficient) and strong robustness to noise and recording conditions change. The DRNN outperformed basic RNN and state-of-the-art velocity Kalman filter (VKF) in both decoding accuracy and robustness. It suggested that our proposed DRNN could be a promising BMI decoding algorithm.