A bio-inspired neuromorphic system for fusing visual features and autonomous learning.

Abstract

This work presents a brain-inspired neural network model for multi-modal image classification based on dual-feature fusion. Inspired by the parallel processing of color and shape in the brain's ventral visual pathway, our model comprises three core components: 1) a feature fusion module that integrates and matches visual features from different modalities; 2) a learning memory module that employs memristor-based synaptic plasticity to learn and consolidate correct feature associations; and 3) a classification module that enables rapid inference through long-term memory. The memristor-enabled circuit dynamically generates feature weights, allowing the system to discriminate between features based on pulse width or voltage amplitude. After learning, the system bypasses the fusion module for direct classification, which is a dynamic pathway switch that enhances both speed and circuit efficiency. Experimental results validate the circuit's scalability through multi-image classification of Teenage Mutant Ninja Turtles characters and its extension to real-time person recognition in inspection robots, demonstrating its potential for high-speed, low-power AI systems.