Quantized STDP-based online-learning spiking neural network

In this work, we report a spike-timing-dependent plasticity (STDP)-based weight-quantized/binarized online-learning spiking neural network (SNN). The SNN uses bio-plausible integrate-and-fire (IF) neuron and conductance-based synapse as the basic building blocks and realizes online learning by STDP and winner-take-all (WTA) mechanism. Weight quantization/binarization is introduced into the online-learning SNN to reduce storage requirements and improve computing efficiency. After the training process with STDP and weight quantization on the MNIST training set, the quantized SNN with 4-bit weight achieves a recognition accuracy of 93.8% on the MNIST test set, showing little loss compared with the accuracy of the non-quantized 32-bit SNN (94.1%). The accuracy of the binarized SNN slightly decreases to 92.9%, which is cost-effective considering the reduction in the weight storage space by ~ 32 times, and the product of input and weight in the binarized SNN can be realized by computationally cheap 1-bit “AND” operation. The proposed weight quantization/binarization online-learning scheme can largely save hardware costs. The area of the quantized (8-bit and 4-bit) and binarized (1-bit) SNN-based hardware is evaluated to be 448,524, 179,263, and 162,129 μm2, respectively, which is much smaller than their non-quantized 32-bit competitor (area of ~ 5.862 × 108 μm2). The hardware resource evaluation also provides a guide to make a trade-off between computational cost and performance. Moreover, the quantized/binarized STDP training method can be further extended to train various types of SNNs. In this regard, a hybrid STDP SNN and a hybrid STDP convolutional SNN, which are trained by combining unsupervised quantized/binarized STDP and supervised backpropagation (BP) training methods, achieve high accuracy in facial expression recognition scenarios.