Event-based visual data sets for prediction tasks in spiking neural networks

Abstract

For spiking networks to perform computational tasks, benchmark data sets are required for model design, refinement and testing. Classic machine learning benchmark data sets use classification as the dominant paradigm, however the temporal characteristics of spiking neural networks mean they are likely to be more useful for problems involving sequence data. To support these paradigms, we provide data sets of 11 moving scenes, each with multiple variations, recorded from a dynamic vision sensor (DVS128), comprising high dimensional (16k pixels) and low latency (15 microsecond) events. We also present a novel long range prediction task based on the DVS128 data, and introduce a pilot study of a spiking neural network learning to predict thousands of events into the future. © 2014 Springer International Publishing Switzerland.