A new pre-conditioned STDP rule and its hardware implementation in neuromorphic crossbar array

Abstract

This paper proposes a new pre-conditioned spike-timing-dependent plasticity (STDP) learning rule as well as an efficient system-level time-domain circuit modeling and simulation method for the whole crossbar array structure. First, the new pre-conditioned STDP rule is conceived by exercising a slight perturbation to the conventional STDP curve. Software-wise, it can be described in spiking neural network (SNN) codes, while hardware-wise, it is realized by the memristor crossbar array in conjunction with specific spike signals, just as the conventional STDP. The new STDP rule enables the neural network in a neuromorphic chip to achieve better performance both in terms of software and hardware implementation compared with the conventional STDP rule. Second, a time-domain circuit modeling method is proposed for the simulation of the whole crossbar array. Numerical experiments demonstrate that the new pre-conditioned STDP is superior to the conventional STDP both in terms of accuracy and low power consumption. At the software level, a 784 × 100 spiking neural network, trained by the pre-conditioned STDP, for the MNIST handwritten digit recognition achieves an accuracy of 85.90%, which is 3.09% higher than that trained by the conventional STDP. At the hardware level, a downscaled 196 × 10 crossbar array is simulated and trained. The crossbar array trained by the new STDP consumes 21.5% less power than that trained by the conventional STDP. Moreover, it does not compromise the original robustness when subject to non-ideal characteristics such as variation of conductance and resistance of the interconnect in the crossbar array.