Stochastic Computing with Simulated Event Camera Data

This paper presents initial results of the first time event camera data is being paired with a stochastic computing system for image processing applications in a near-sensor neural network system. Stochastic computers have great potential to reduce the size, complexity, and power required to complete common computing tasks. However, stochastic computing re-quires data to be spread across time in pseudorandom bit-strings; producing these bit-streams is costly from a power perspective. The saccades observed by the event camera can be used to generate pseudo-random bit-streams at the camera’s output. Therefore, it is natural to pair these two systems together. The system is described and performance results are shown from pairing simulated Dynamic Vision System (DVS) event camera data with a MATLAB simulation of our Field-Programmable Gate Array-based stochastic computing system.