Accelerating Deep Learning via Adaptive Deep Reuse
Event Type
Machine Learning Day
AI/Machine Learning/Deep Learning
TimeWednesday, June 19th2:45pm - 3:15pm CEST
LocationPanorama 3
DescriptionOne of the key road blocks in practical adoptions of Deep Neural Networks (DNN) is the long training process of DNN models. This talk presents adaptive deep reuse, a novel optimization to the-state-of-the-art training methods of DNN. Adaptive deep reuse offers a creative way to efficiently and effectively identify unnecessary computations in DNN training on the fly. By avoiding these computations, it cuts the training time of DNN by 69% without sacrificing accuracy. The method is fully automatic and ready to adopt, requiring neither manual code changes nor extra computing resource. It offers a promising way to substantially reduce both the time and the cost in the development of AI products.