Image Recognition Using an Event Camera and a Stochastic Neural Network

This study presents an innovative approach to image recognition by combining an Event Camera (EC) with a Stochas-tic Computing (SC) neural network. The novel aspect of this work lies in the fusion of EC, known for its temporal resolution and motion blur reduction, with SC, which offers efficient arithmetic using simple logic processes. This pairing offers the potential for power-efficient neural network systems. The EC data, which captures pixel-level intensity changes asynchronously, drives the SC neural network. A sample-and-hold system and checkerboard filter are incorporated to overcome data sparsity issues and enhance event locality. The experimental results show significant improvements, adding 20% to performance when event frequency is increased, suggesting the promise of this approach in achieving more efficient image recognition systems.