Insect-Inspired Elementary Motion Detection Embracing Resistive Memory and Spiking Neural Networks

Computation of the direction of motion and the detection of collisions are important features of autonomous robotic systems for course steering and avoidance manoeuvres. Current approaches typically rely on computing these features in software using algorithms implemented on a microprocessor. However, the power consumption, computational latency and form factor limit their applicability. In this work we take inspiration from motion detection studied in the Drosophila visual system to implement an alternative. The nervous system of the Drosophila contains 150000 neurons [1] and computes information in a parallel fashion. We propose a topology comprising a dynamic vision sensor (DVS) which provides input to spiking neural networks (SNN). The network is realised through interconnecting leaky-integrate and fire (LIF) complementary metal oxide semiconductor (CMOS) neurons with hafnium dioxide (HfO2) based resistive random access memories (RRAM) acting as the synaptic connections between them. A genetic algorithm (GA) is used to optimize the parameters of the network, within an experimentally determined range of RRAM conductance values, and through simulation it is demonstrated that the system can compute the direction of motion of a grating. Finally, we demonstrate that by modulating RRAM conductances and adjusting network component time constants the range of grating velocities to which it is most sensitive can be adapted. It is also shown that this allows for the system to reduce power consumption when sensitive to lower velocity stimulus. This mimics the behavior observed in Drosophila whereby the neuromodulator octopamine adjusts the response of the motion detection system when the insect is resting or flying.