A neuromorphic event data interpretation approach with hardware reservoir.

Abstract

Event cameras have shown unprecedented success in various computer vision applications due to their unique ability to capture dynamic scenes with high temporal resolution and low latency. However, many existing approaches for event data representation are typically algorithm-based, limiting their utilization and hardware deployment. This study explores a hardware event representation approach for event data utilizing a reservoir encoder implemented with analog memristor. The inherent stochastic and non-linear characteristics of the memristors enable the effective and low-cost feature extraction of temporal information from event streams as a reservoir encoder. We propose a simplified memristor model and memristor-based reservoir circuit specifically for processing dynamic visual information and extracting feature in event data. Experimental results with four event datasets demonstrate that our approach achieves superior accuracy over other methods, highlighting the potential of memristor-based event processing system.