Abstract
This paper builds upon our previously reported growth transform based optimization framework to present a novel spiking neuron model and demonstrate its application for spike-based auditory signal processing. Unlike conventional neuromorphic approaches, the proposed Growth Transform (GT) neuron model is tightly coupled to a system objective function, which results in network dynamics that are always stable and interpretable; and the process of spike generation and population dynamics is the result of minimizing an energy functional. We then extend the model to include axonal propagation delays in a manner that the optimized solution of the system or network objective function remains unaffected. The paper characterizes the model for different types of stimuli, and explores how changing different aspects of the cost function can reproduce known single neuron dynamics. We then investigate the properties of a coupled GT neural network that can generate spike-encoded auditory features corresponding to the output of a gammatone filterbank. We show that the discriminatory information is not only encoded in the traditional spike-rates and interspike-interval statistics, but is also encoded in the subthreshold response of GT neurons for inputs that are not strong enough to elicit spikes. We also demonstrate that while different forms of coupling between the neurons could produce compact and energy-efficient representation of the auditory features, the classification performance for a speaker recognition task remains invariant across different types of coupling. Thus, we believe that the proposed GT neuron model provides a flexible neuromorphic framework to systematically design large-scale spiking neural networks with stable and interpretable dynamics.