Power efficient computing for statistical machine learning

Overview

The carbon footprint of computing globally is estimated to be comparable with that of the aviation industry. With the advent of generative artificial intelligence, there is a growing awareness of this environmental impact both in terms of the carbon footprint and other environmental impacts including e-waste and water consumption, predominantly through the use of power-hungry graphics processing units (GPUs).

These are particularly relevant issues to many fields that rely on computationally intensive simulations for data analysis or calibration of statistical machine learning models.  For example, simulation-based statistical inference (SBI) methods are foundational to many scientific disciplines, including fields essential to addressing the climate crisis.

There have been recent advances in low-power, high performance computing, however, there are challenges to overcome for them to become a mainstream technology. It is therefore critical to develop new mathematical and statistical methods tailored to these state-of-the-art breakthroughs in energy efficient computing.

Research activities

In this research project, you will develop novel numerical methods for stochastic simulation and inference. An implementation target for these methods will be modern low-power, high performance reconfigurable computing devices, such as the state-of-the-art in field programmable gate arrays (FPGAs). These devices have the potential to outperform GPU computing at a fraction of power-consumption.  In addition, you will develop user-friendly software packages in the most popular programming environments to  promote adoption of low-power computing by practitioners.

Outcomes

In completion of this project  you will:

• develop new stochastic simulation and statistical inference algorithms  suitable for reconfigurable computing deployment
• evaluate the performance, accuracy, and power efficiency of these methods implemented in  real hardware
• apply  techniques to problems in biology, ecology, or epidemiology as exemplars
• present results through publications in high quality academic conference proceedings and  journal articles.

Skills and experience

The following skills will be necessary:

• understanding of probability and stochastic processes
• programming abilities (preferred languages are MATLAB, R, Python, or Julia)
• understanding of statistical inference (classical or Bayesian)
• knowledge of differential equations is desirable
• experience with heterogeneous computing is desirable (e.g., CUDA, OpenCL or OpenACC).

Scholarships

You may be eligible to apply for a research scholarship.

Explore our research scholarships

Keywords

• simulation-base inference
• green computing
• stochastic simulation

Contact

Contact the supervisor for more information.