Probabilistic neural network based motor cortical decoding method and hardware implementation

Probabilistic neural network (PNN), a kind of radial basis networks, is usually used for classification problems. It has the advantages of much faster training process and more accurate results using the minimum Bayesian risk criterion compared with other neural networks. In this paper, we use this neural network in brain-machine interface for decoding neural ensemble activity. Rats were trained to perform a lever pressing task for water rewards and implanted a chronic 16-channel microelectrode array in the primary motor cortex. The multi-channel neural activity and pressure signal of the lever was recorded simultaneously. In the PNN architecture, input vector was modified and contained not only current neuronal activity but also previously estimated pressure value which was taken as one input. This modified PNN was implemented in Matlab and obtained a good performance. Furthermore, field-programmable gate array (FPGA) based hardware implementation of PNN were developed and tested with real data. Because of the parallel computation ability and optimized architecture of FPGA, the results are as accurate as the realization of Matlab-based but the running speed is much faster. This indicates that the performance of current FPGA is competent for portable BMI applications.