Flicker-Suppressed Neuromorphic Unit for Dynamic Vision Processing

TL;DR

This paper introduces a neuromorphic unit inspired by insect vision to process dynamic visual data efficiently.

Contribution

A novel neuromorphic unit combining homojunction and heterojunction to emulate nerve signal transmission modes and reduce flicker noise.

Findings

The device achieved an information transmission rate of 2100 bits s−1.

Action potentials and postsynaptic potential responses were successfully generated.

Trajectory recognition across four car orientations was achieved using in-sensor reservoir computing.

Abstract

Inspired by the dynamic
visual perception of flying insects, rapid
collision warning systems are crucial for advancing autonomous driving
and machine control. Although neuromorphic devices show significant
potential for replicating insect vision systems, they are hindered
by limitations in the sensing frequency, signal-to-noise ratio, and
flicker noise. Here, we use a combination of a homojunction and heterojunction
to emulate the two different transmission modes of nerve signals via
gate-voltage modulation. The structural design and heterojunction
effects enabled artificial neurons to respond to high-frequency visible-light
signals and achieve an information transmission rate of 2100 bits
s−1. By connecting the leaky integrate-and-fire
neural device in series with the synaptic device, we successfully
generated action potentials and postsynaptic potential responses,
significantly reducing cumulative threshold flicker noise. Using in-sensor
reservoir computing, we achieved trajectory recognition across four
car orientations with an optimized training process, providing valuable
insights into device design and applications in visual bionics.

Figures