A Neuro-Inspired Spike Pattern Classifier

Abstract

The emergence of post-silicon nano-devices and the coming trillion-sensor era driven by the Internet of Things have led to a search for alternative computational paradigms that can efficiently derive useful information from abundant data while enable efficient hardware implementations under significant device variations. Such computational paradigms may emerge as we explore the information processing in biological sensory systems, which achieve unparalleled performance and energy efficiency with mediocre and unreliable components. This paper presents a neuro-inspired spike pattern classifier for an artificial olfactory system, where an analog feature extraction gas sensing front end first converts sensor array information into spike patterns. The spike pattern classifier consists of a transformation of the input spikes into high-dimensional sparse vectors and a cortical memory model. The transformation is based on a random sampling scheme that can be efficiently performed with circuits exhibiting large parametric variations. An associative memory is used to perform fast and efficient storage and retrieval of the sparse vectors. The classifier achieves an output retrieval fidelity of 0.97 with a Fano factor ( $\sigma /\mu $ ) of 0.0143 when it is implemented with delay cells having a delay Fano factor of 1.6759. This is 117 times of reduction in parametric spread. It is also robust against operation failures: the output fidelity is still greater than 0.8 even when the probability of operation failure reaches 0.45.