Neural Network Topologies for Sparse Training
This addresses the scalability issue in deep learning for researchers and practitioners, but it is incremental as it builds on existing sparse training methods like X-Nets.
The paper tackles the problem of hardware limitations for storing and training large deep neural networks by proposing an algorithm that deterministically generates diverse sparse topologies, which can train to the same precision as dense networks at lower cost.
The sizes of deep neural networks (DNNs) are rapidly outgrowing the capacity of hardware to store and train them. Research over the past few decades has explored the prospect of sparsifying DNNs before, during, and after training by pruning edges from the underlying topology. The resulting neural network is known as a sparse neural network. More recent work has demonstrated the remarkable result that certain sparse DNNs can train to the same precision as dense DNNs at lower runtime and storage cost. An intriguing class of these sparse DNNs is the X-Nets, which are initialized and trained upon a sparse topology with neither reference to a parent dense DNN nor subsequent pruning. We present an algorithm that deterministically generates sparse DNN topologies that, as a whole, are much more diverse than X-Net topologies, while preserving X-Nets' desired characteristics.