Deep Rewiring: Training very sparse deep networks
This addresses the need for efficient deep learning implementations on hardware with connectivity limits, though it is incremental as it builds on existing pruning methods.
The paper tackles the problem of training neural networks with strict connectivity constraints for efficiency on neuromorphic and generic hardware, presenting DEEP R, an algorithm that directly trains sparse networks by automatically rewiring connections during supervised training while maintaining a fixed total number of connections, resulting in only minor performance loss on standard benchmarks.
Neuromorphic hardware tends to pose limits on the connectivity of deep networks that one can run on them. But also generic hardware and software implementations of deep learning run more efficiently for sparse networks. Several methods exist for pruning connections of a neural network after it was trained without connectivity constraints. We present an algorithm, DEEP R, that enables us to train directly a sparsely connected neural network. DEEP R automatically rewires the network during supervised training so that connections are there where they are most needed for the task, while its total number is all the time strictly bounded. We demonstrate that DEEP R can be used to train very sparse feedforward and recurrent neural networks on standard benchmark tasks with just a minor loss in performance. DEEP R is based on a rigorous theoretical foundation that views rewiring as stochastic sampling of network configurations from a posterior.