Spiking Domain Feature Extraction with Temporal Dynamic Learning

Honghao Zheng and Yang (Cindy) Yi
Virginia Tech


Spiking neural network (SNN) plays an essential role due to its likelihood of the signal processing mechanism in biological neural systems. Several algorithms have been investigated, such as surrogate gradients and spike-timing-dependent plasticity (STDP), to train the SNN. In STDP, synaptic weights are modified according to the relative time difference between pre and post-synaptic spikes. Since the traditional pair STDP model only takes account of a pair of pre and post-synaptic spikes, a triplet STDP model was proposed. This model can better take account of a series of spikes and thus more closely mimic the activity in biological neural systems. A circuit that can switch between different STDP rules and adjust weight-changing amplitude and decaying rate was introduced to improve the reconfigurability of the STDP circuit. To apply the advantages of triplet STDP to various tasks, a mixed-signal triplet reconfigurable STDP CMOS circuit and its hardware prototype are proposed in this paper. The performance analysis of multiplexing encoding and STDP training is carried out with a hardware testbench and Pytorch SNN. This triplet STDP design achieves 3.28% and 3.63% higher accuracy than the pair STDP learning rule through datasets such as MNIST and CIFAR-10. Our design shows one of the best reconfigurability while keeping a relatively low energy per spike operation (SOP) through the performance comparison with the state of the arts.