Brain-inspired sparse population coding for robust object recognition under sensory uncertainty

Abstract

Compared to traditional machine learning systems, the human visual system shows very high robustness against noise, occlusion and variation of the environment. This paper presents a brain-inspired sparse population coding framework that combines a neural manifold, mixed selectivity, and FPGA-based adaptive encoding for robust object recognition under sensory uncertainty. The model uses sparse distributed representations to improve feature discrimination and reduce redundancy and computational cost. The experiments on common vision datasets (n=200 samples) confirm that the model can achieve higher accuracy, about 3.1%-4.5% higher than traditional dense coding, in the face of noise and occlusion. FPGA implementation reduces processing latency by 18%-25%, enabling efficient real-time inference; paired t-tests confirm these performance gains (p<0.05). Additionally, the model shows greater robustness to adversarial attacks and illumination variations. These results indicate that the biologically-motivated sparse encoding scheme provides an effective and reliable visual recognition system. The proposed framework is well suited for edge robots, autonomous navigation, and surveillance systems where stable, efficient perception is critical.