Neuromorphic Processing at the Sensor Edge: Engineering Tiny Brain

Brain-inspired, neuromorphic spiking neural network emulators form distributed, parallel, and event-driven system offering signal transmission energy-efficiency, and intelligence or inference capabilities to resource-constrained devices. Computational elements of the reconfigurable neuromorphic networks, typically, only partially include dendritic, and subsequently, synaptic properties. However, increased experimental evidence indicates the existence of a large diversity of dendritic channels, which modify synaptic response by amplification, regulation, the dendritic structure scaling, etc. We abstract the fundamental dendritic functions by extracting the underlying dynamics governed by bio-chemical processes; this increase in dimensionality allows more states and transitions (and time constants), offering more flexibility in the implementation of plastic and metaplastic interactions, i.e. providing mechanism to realize and maintain robust neural computation, in addition to enhancing specifics of the sensory signal processing, e.g. accentuate changes in stimulus parameters, prevent spiking frequency saturation, tune frequency responses to specific stimulus features. In this tutorial, we address several important questions such as i) how do we leverage all the advantages of dendritic bio-chemical signal processing at different levels of time-granularity or hierarchy, and ii) how the main principles of a tiny-brain translate to competitive (sub-pJ level synaptic operation) neuromorphic hardware platforms, and software/hardware co-alignment in terms of distributed processing, technological variability, and neural network definitions and mapping. Enabling such paradigms and concepts open in-roads towards several high-potential use-cases geared specifically towards processing incoming data in an always-on, event-driven fashion facilitating truly smart-sensing and energy-optimized sensor systems. Speaker: Amir Zjajo , Innatera Nanosystems