Enabling future progress in machine-learning

Amazing progress in machine-learning, largely based on deep neural networks, has started to make applications once considered impossible, such as real-time translation or self-driving cars, a reality. However, even if, on some restricted problems, machine-learning is getting close to human-level performance, we are still far from the capabilities of the human brain. Machine-learning researchers themselves acknowledge that the progress observed in the past 10 years has been largely due to rapid increase in computing performance, allowing to tackle larger neural networks and larger training sets. So the computer systems and circuits communities can play a very significant role in enabling future progress. While GPUs have been a major driver of this recent progress, both the slowing rate of improvement of standard CMOS technology and the need for even faster progress suggest to at least explore alternative approaches. In this talk, we will discuss lessons learned from research on architectures for machine-learning, and that some of the hurdles ahead largely lie at the circuit level, but can possibly be overcome in the near future.