Neuromorphic Hardware Design

Abstract

In this thesis, a novel technique to embed synaptic plasticity in neuromorphic hardware is proposed named as Frequency Dependent Synaptic Plasticity. This technique provides an alternate interpretation for plasticity which was conventionally modeled as weight value. In the proposed model of neuromorphic hardware, plasticity is implemented in frequency domain by considering a synaptic connection performing bandpass filtering operation. Currently most of the neuromorphic hardware are based on time domain based plasticity techniques. The proposed model attempts to contribute and suggest an out of the box solution for neurocomputational applications. It has been established through this thesis that the proposed model is biologically plausible in terms of implementation of different phenomena observed in a biological brain such as rate-encoding by Class I type of neuron, decoding of rate-encoded information by selective triggering at post synaptic side, resonance aware synaptic plasticity, population coding, role and interpretation of tuning curve, and frequency based neuronal communication. The proposed hardware operates in analog domain which is closer to the operational domain of brain as compared to its digital counterpart. This thesis also proposes a novel architecture to embed population coding on neuromorphic hardware. Further, it is explained in this thesis that a synaptic junction based on proposed model has a non linear transfer function which omits the requirement of hidden layer in classifying non-linear problems. This will allow applications to use least possible resources, employing input and output nodes only, as compared to a network based on linear weight values.