Hardware Aware Training for AI Systems

Project description

Neural network based AI systems have made remarkable progress over the last decade, with applications in computer vision, language processing and robotics. As performance has increased, so has the data required to train these systems, along with the size of the underlying models. This has resulted in growing energy demands, both during training and in deployment.

Solutions to the energy challenges of neural network systems are being developed in the form of new hardware, including advances in traditional massively parallel systems such as GPUs, but also with novel approaches such as memristor based systems and neuromorphic hardware. These novel systems provide orders of magnitude reductions in power, offering game-changing performance both in data centres and mobile edge devices. However, these systems often present challenges when implementing conventional neural network training methods, as operations represented in software are not accurately or reliably implemented in hardware. This makes exploiting the energy potential of these systems difficult.

The goal of this project is to explore hardware-aware training methods for AI systems. Developing algorithms and representations which take into account the underlying characteristics of the hardware during training, to deliver low-power hardware implementations which are accurate, reliable and robust.