The program will create novel computational algorithms to accommodate the demands of modern data science applications across many and diverse disciplines. The designed algorithms will be robust and capable of exploiting rapidly evolving parallel computing architectures. For more than a decade, the speed of CPUs has not changed significantly. CPU architectures are now designed to perform multiple tasks in parallel. Key challenges that the algorithms aim to address include:
- Parameter estimation and model selection for computational expensive and implicit models.
- Optimisation of noisy and multi-modal objective functions arising in statistics and machine learning.
- Processing challenging data: big, tall, wide, streaming and so on
- Robustness to model misspecification (“all models are wrong”) and targeted towards analysis objectives (e.g. prediction)
- Optimal data collection strategies to maximise information and reduce bias
The program has obvious links to the other programs. It will support innovation in data acquisition, process data-focused models and facilitate decision making under uncertainty. There are many disciplines these algorithms can be applied to across the data science network at QUT such as health, biology, ecology, robotics and the environment.