Auto-associative network implementation in the neuromorphic chip „Spikey“

Abstract

Background and Aim: Simulations of computational models of brain activity are computationally expensive and require high resources of computational power and energy. Neuromorphic computing systems are designed to accelerate neuronal network emulations and use analogue circuits to mimic neurobiological architectures. One of such systems, created by the Electronic Vision(s) Group in Kirchhoff Institute for Physics, Heidelberg University, Germany, is the “Spikey” chip. This neuromorphic system comprises 384 neurons and 98304 synapses and emulates neuronal activity 10000-fold faster than biological real-time. The aim of the study is to estimate the simulation efficiency and learning accuracy of the auto-associative memory model, implemented on the Spikey neuromorphic system. Materials and Methods: „Spikey“ chip is programmable with the PyNN API in Python programming language. Auto-associative neural network was programmed using PyNN API. Network consists of leaky integrate-and-fire neurons connected all-to-all by excitatory synapses and a group of all-to-all inhibitory connections controlling a spiking rate. Symmetric and asymmetric STDP learning rules were used for excitatory-excitatory synaptic connections, while excitatory-inhibitory and inhibitory-excitatory synaptic connections had constant weight values assigned at the start of the experiment. Spike bursts with the varying spike counts were used as the inputs. Results: Pattern recall accuracy was highest with the symmetric STDP learning rule and a larger number of spikes in the spike bursts. There is a strong interference with the network activity from the hardware inherent noise and spike delays, reducing overall recall accuracy and learning capacity.[...].