Hardware realization of novel pulsed neural networks based on delta-sigma modulation with GHA learning rule

In this paper, the pulsed neural network architecture based on delta-sigma modulation (DSM-PNN) has been proposed. As the DSM-PNN transfers information pulse-encoded by delta-sigma modulation, between every pair of neurons, they are connected with only 1-bit. Therefore the circuit scale becomes small and is effective for hardware implementation. In addition, the noise-shaping effect, which is a feature of delta-sigma modulation, enables the DSM-PNN to transmit the signal faithfully and operate multi-input summation and weight multiplication precisely. The proposed network was evaluated with the generalized Hebbian algorithm (GHA), which is the learning rule of principal component analysis (PCA), and implemented in FPGA. The experimental results show that the proposed system has the same accuracy as those with floating-point units.