Democratising Compilers for Machine Learning Systems

With the recent surge in AI, specialised systems for machine learning are becoming increasingly popular and necessary. The end of Moore's law marks a time where researchers and developers cannot steadily rely on more processing power to accelerate their software systems - they have to introduce new performance engineering approaches for further optimisations. In the Machine Learning era, this manifests through the development of specialised hardware architectures (such as Google's TPU), using parallelisation and concurrency offered with many modern processors or defining new hardware-aware optimisations. There is a clear need for a compiler framework that allows the end user to leverage the hardware efficiently and intuitively. This PhD proposal aims to democratize the development of compilers for machine learning systems.