"Deep learning has revolutionized the fields of computer vision, natural language understanding, speech recognition, information retrieval and more. However, the exploding costs of training neural networks and using them at inference time are becoming unsustainable. In this talk, we discuss the challenges of designing compute-efficient deep learning models and then present the latest techniques for making deep learning models smaller, faster, and better. We also discuss the recent advances in hardware acceleration for deep learning and present the challenges of designing compute-efficient deep learning systems. Finally, we conclude with a discussion of the open research directions in compute-efficient deep learning.

"