Going Small: Using the Insect Brain as a Model System for Edge Processing Applications

Abstract

In this work I explore bio-inspired architectures for adaptive and smart sensing incorporating two key aspects present on the insect brain that are not found in more traditional neural network approaches: modulated, hierarchical processing and modulated learning. Our architecture incorporates two central ideas: 1) a state-dependent processing of inputs that can be triggered internally or externally, and 2) state-dependent online learning capabilities, in this specific case allowing the system to change the valence associated to different types of input. These ideas are explored through a hybrid design in which information is processed through a spiking neural network, while a recurrent non-spiking component provides the modulatory feedback to the system. The proposed approach exemplifies how neuromorphic computing approaches naturally integrate sensing and processing within a single functional unit. The proposed architecture can be implemented using conventional VLSI processing, though the integration of novel materials can help simplify its implementation.