Spurious learning in networks with Spike Driven Synaptic Plasticity

Abstract

Biologically plausible learning is of interest not only to neuroscientists but also to the neuromorphic community due to local learning rules, and the promise of efficient hardware implementations. Spike timing dependent plasticity is the conventional biologically plausible learning rule that has been used to train SNNs. However, due to its advantages, spike driven synaptic plasticity (SDSP) rule has begun to be used in some implementations. Unlike STDP, however, under some parameter settings, we note that learning is not proportional to the incoming input received by an output neuron. This phenomenon leads to false positive learning, or otherwise spurious learning. In this paper, we first present a simple 5-class network that gets 99.97% accuracy on simple patterns. Then we present a detailed examination of spurious learning in SDSP networks.