Ensemble learning for spiking neural networks

This paper demonstrates how to combine predictions of spiking networks such that the combined prediction is on average more accurate than an individual’s. Even though spiking neural networks have the same capacity to learn as traditional neural networks [5], they have not challenged the performances obtained by the latter on the same problems [6]. One well established approach of improving the accuracy of machine learning models is ensemble learning. We identify and investigate two levels at which an ensemble can be constructed: inside a spiking neural network (also known as a population) or between a group of spiking neural networks. We present two approaches of combining spiking predictions such that the Ambiguity decomposition [4] exists. We provide empirical support for the theoretical gurantee of these combiners, that on average the ensemble predictions will be more accurate than the single model predictions. As with traditional ensembles, this guarantee is in part due to diversity, thus we conclude our study by showing how diversity affected the trained ensembles and investigating the efficiency of three implicit diversity methods on the ensemble diversity.