Graph sparsification on deep neural network
Yijia Qiu · Illinois Digital Environment for Access to Learning and Scholarship (University of Illinois at Urbana-Champaign) · 2018
Pruning on deep neuron networks can reduce the computation cost and memory use. Graph Sparsification is a method which considers the neuron network as a graph, where neurons are vertices and the connections between neurons are edges. We are able to generate an ultrasparse sub-graph that well preserves the structure of the original one. The remaining connections of reduced neuron networks are 30 times or fewer than before and the accuracy maintains almost the same after the re-training. Such an optimization would lead to a highly efficient implementation of the reduced deep neural networks unto hardware accelerators such as FPGAs.