Machine learning models are being deployed in critical domains such as healthcare, education and fintech. The current approach to deploying machine learning models is based on considering a data-centric approach where the models are evaluated using performance measures on a test set. However, the high performance of the model on test data is not indicative of its reliability,
An important aspect of reliability is in the understanding of what exactly a machine learning model encodes, and to verify if it learns the domain knowledge as expected. It is therefore essential to detect the possible misconceptions of the model proactively. It is also important to incorporate human domain knowledge to refine the model and eliminate model errors due to data or model parameters.
In this project, we aim to develop methods and technique to identify the sources of errors:
- data-specific errors
- model or conceptual errors
- model usage errors.
You'll develop techniques to debug model errors and incorporate human inputs to correct errors.
The research activities can be scoped to cater for different types of research student projects, and include:
- designing and implementing techniques for identifying different types of model errors
- designing and implementing methods and techniques that incorporate human inputs to refine and eliminate model errors
- developing and testing an open-source framework for realising the proposed techniques.
As a result of this research, you'll create:
- methods to identify and categorise errors in machine learning models such as model errors, data-distribution errors, parameter setting errors, and so on
- new and improved methods and techniques for taking additional inputs from humans to refine the models.
Skills and experience
You should have:
- knowledge of machine learning
- programming skills (Python)
- problem-solving and critical thinking skills
- reasonable writing skills
You may be eligible to apply for a research scholarship.
Contact the supervisor for more information.