To enable and promote interpretability, uncertainty quantification, unbiasedness, transparency and reproducibility into the design of learning algorithms.
Dr Fred Roosta-Khorasani
About the Theme
This theme aims to address challenges related to:
Transparency of Machine Learning (ML)
How to construct transparent ML algorithms that are not only predictive, but also provide an acceptable degree of interpretability with quantifiable uncertainty?
Low quality data
How to address issues involving large amounts of, potentially low-quality, data, whereby, in addition to algorithmic efficiency and uncertainty quantification, existing domain-specific factors need to be considered?
This research theme is led by Dr Fred Roosta-Khorasani at The University of Queensland. Theme 3 focuses on developing trustworthy machine learning models that can enhance knowledge discovery from the data. Research leaders and partners connected with this theme focus on the study of techniques that allow for uncertainty quantification of the predictive performance as well as interpretability of the deployed models. The key deliverables will involve designing, implementing, and deploying automatic model validation as well as user-friendly diagnostics tools, which ensure that the trained models learn solely from the existing evidence in the data as opposed to the potential biases and corruptions that can occur during the data collection procedures. These tools will all be enhanced by providing the abilities to incorporate human-in-the-loop strategies through collaboration with Theme 2 – Data curation at scale research and informed by Theme 5 – Agility in value creation from data research on user and organisational requirements.