Bio-Inspired Wide-Field Visual Neuron Implemented with Ultra-Low Information Loss Population Coding.

Abstract

Modern neuromorphic systems face critical bottlenecks in emulating biological vision, particularly in reconciling wide-spectrum perception, distortion-free encoding, and population-level signal processing. Drawing inspiration from the stochastic-resilient population coding of macaque visual neurons, an advanced visual neuron prototype is developed that incorporates photoelectric multi-stimulation field-effect transistor and innovative parallel threshold-switch architecture. The visual neuron integrates broadband photodetection (350-1000 nm) with biomimetic spike population encoding in a monolithic architecture. The photosensitive MoSe₂/MoS₂ heterojunction region in field-effect transistor extends the visual perception field from UV to infrared wavelengths (350-700 nm to 350-1000 nm), doubling the original field. Meanwhile, under the same conditions, the photocurrent response achieves a 1.36-fold increase from 0.109 to 0.148 A (W cm^-2)^-1. The parallel threshold-switching design transforms single-unit encoding into cooperative population coding, achieving an 82.1% reduction rate in signal distortion. When deployed in a spiking neural network, this population-coding paradigm demonstrates high accuracy in pattern recognition, surpassing single neuron architectures by 12.1%, while maintaining the information processing time at the biological scale (<200 ms). By unifying van der Waals heterostructure photonics with macaque-derived neural population coding principles, this work establishes a transformative framework for bioinspired vision hardware, bridging the critical gap between neuromorphic materials and cortical processing efficiency.