Information-preserving Transforms: Two Graph Metrics for Simulated Spiking Neural Networks

Abstract

We are interested in self-organization and adaptation in intelligent systems that are robustly coupled with the real world. Such systems have a variety of sensory inputs that provide access to the richness, complexity, and noise of real-world signals. Specifically, the systems we design and implement are ab initio simulated spiking neural networks (SSNNs) with cellular resolution and complex network topologies that evolve according to spike-timing dependent plasticity (STDP). We desire to understand how external signals (like speech, vision, etc.) are encoded in the dynamics of such SSNNs. In particular, we are interested in identifying and confirming the extent to which various population-level measurements (or transforms) are information-preserving. Such transforms could be used as an unambiguous way of identifying the nature of the input signals, when given only access to the SSNN dynamics. Our primary objective in this paper is to empirically examine the extent to which a couple of graph metrics provide an information-preserving transform between the input signals and the output signals. In particular, we focus on the standard deviation of the time-varying distributions for local influence (weighted out-degree) and local impressionability (weighted in-degree), which provide insight into information encoding at the population-level in the dynamics of SSNNs. We report the encouraging results of an experiment carried out in the Language Acquisition and Robotics Group.