Written exam (oral exam in case of a small number of participants)

As machine learning technology gets applied to actual products and solutions, new challenges have emerged. Models unexpectedly fail to generalise well to small changes in the distribution; some models are found to utilise sensitive features that could treat certain demographic user groups unfairly; models tend to be confident on novel types of data; models cannot communicate the rationale behind their decisions effectively with the end users like medical staff to maximise the human-machine synergies. Collectively, we face a trustworthiness issue with the current machine learning technology. A large fraction of the machine learning research nowadays is dedicated to expanding the frontier of Trustworthy Machine Learning (TML). The course covers a theoretical and technical background for key topics in TML. We conduct a critical review of important classical and contemporary research papers on related topics and provide hands-on practicals to implement TML techniques.

Students will be able to critically read, assess, and discuss research work in Trustworthy Machine Learning (TML). They will gain the technical background to implement basic TML techniques in a deep learning framework. They will be ready to conduct their own research in TML and make contributions to the research community.

Prerequisites: Students should be familiar with Python and PyTorch coding. They should have basic knowledge of machine learning concepts: supervised learning, function fitting, generalisation gap, overfitting, and regularisation. Furthermore they should have basic knowledge of deep learning (CNNs and Transformers, their components, and empirical techniques for training and evaluating them), for instance by having passed ML-4103 (or equivalent). Basic maths skills (multivariate calculus, linear algebra, probability, statistics, and optimisation) are also required.