Paul Brasnett, Principal Research Engineer at Imagination Technologies, presents the "Training CNNs for Efficient Inference" tutorial at the May 2017 Embedded Vision Summit.
Key challenges to the successful deployment of CNNs in embedded markets are in addressing the compute, bandwidth and power requirements. Typically, for mobile devices, the problem lies in the inference, since the training is currently handled offline. One approach to reducing the inference cost is to take a trained network and use a tool to map it to a lower cost representation by, for example, reducing the precision of the weights. Better inference performance can be obtained if the cost reduction is integrated into the network training process. In this talk, Brasnett explores some of the techniques and processes that can be used during training to optimize the CNN inference performance, along with a case study to illustrate the advantages of such an approach.