Efficient and hardware-friendly methods to implement competitive learning for spiking neural networks

摘要

Spiking neural network (SNN) trained by spike-timing-dependent plasticity (STDP) is a promising computing paradigm for energy-efficient artificial intelligence systems. During the learning procedure of SNN trained by STDP, another two bio-inspired mechanisms of lateral inhibition and homeostasis are usually implemented to achieve competitive learning. However, the previous methods to implement lateral inhibition and homeostasis are not designed with hardware in mind, resulting in solutions that are not efficient for deployment on neuromorphic hardware. For example, the existing lateral inhibition methods induce a great number of connections that are proportional to the square of the number of learning neurons. The classical homeostasis methods depend on the fine-tuned membrane threshold with no hardware solution provided. In this paper, we propose two hardware-friendly and scalable methods to achieve lateral inhibition and homeostasis. Using only one inhibitory neuron for one learning layer, our proposed lateral inhibition method can reduce inhibitory connection number from N-2 to 2N and hardware overhead by sharing refractory control circuits. Utilizing the adaptive resistance of memristor, we propose a novel homeostasis method through adapting the leaky current of spiking neurons. In addition, the learning efficiency of different homeostasis methods are studied for the first time by simulating on the cognitive task of digital recognition of MNIST dataset. Simulation results show that our proposed homeostasis method can improve the learning efficiency by 30-50% while maintaining the state-of-the-art performance.