A Quantitative Exploration of Collaborative Pruning and Approximation Computing Towards Energy Efficient Neural Networks

摘要

Editor's note: This work has the goal of minimizing digital neural network computation energy consumption with little loss in accuracy. The authors describe a Dynamic Network Surgery based approach to network pruning, after which weights are incrementally selected for approximate multiplication. Considering which network components are necessary and determining the needed level of accuracy for them enables greater energy savings than solving either problem independently. - Robert P. Dick, University of Michigan