Abstract

The use of neural networks has advanced the state of the art in many fields, including computer vision, speech processing and machine translation, and there is reason to believe that more is yet to come. To enable all this, one of the key challenges is to find ways to execute neural networks efficiently on computer hardware.

In this talk, we will discuss how neural networks map onto conventional and unconventional computer architectures, including case studies for CPUs, GPUs and dedicated accelerators. We will also look at accelerator techniques for further improving performance, as well as look at the challenge of keeping some programmability to deal with future changes in the state of the art of neural networks.