From algorithmic fairness to algorithmic robustness and back

21 June 2023

1:00pm-2:00pm

 

Summary

Machine learning models have been employed in various scenarios, including safety-critical situations like self-driving cars and disease diagnosis, as well as socially significant areas such as healthcare, education, and credit allocation. However, many machine learning algorithms require a significant amount of labelled data to train and often rely on complex relationships and distant correlations found in the training dataset. Although these models can increase overall accuracy, relying on spurious correlations for decision-making may have adverse effects on generalization, especially when those correlations are not applicable in the test distribution. For example, a model that associates cats with indoor surroundings may not accurately predict the presence of cats in outdoor settings. To address these issues, it is crucial to develop models that are both algorithmically robust and fair. Both aims seek to train a model that generates predictions that are statistically independent of domain variations, with fairness being focused on protected characteristics or subgroups like age or gender. During this talk, I will present recent research on developing robust models that achieve high worst-subgroup accuracy using statistical matching methods.

Bio

Novi Quadrianto is a machine learning scientist and a Professor of Machine Learning at the University of Sussex. He is also an Adjunct Professor (Data Science) at Monash University Indonesia, and leads a BCAM Severo Ochoa Strategic Lab on Trustworthy Machine Learning in Spain. In 2019, Novi was awarded a European Research Council ERC grant for a project on developing a machine learning framework for addressing fairness under uncertainty in a static and a dynamic setting, while also ensuring transparency in fairness (BayesianGDPR). He is part of the consortium of 21 partner organisations from 9 European countries which was recently selected to develop a new generation of human-centric AI systems and to strengthen the leadership of Europe in this area. This new EU Horizon Europe funded project TANGO will run from 2023-2027.

For more information, please check https://wearepal.ai/

Further reading

Sara Romiti, Christopher Inskip, Viktoriia Sharmanska, Novi Quadrianto: ‘RealPatch: A Statistical Matching Framework for Model Patching with Real Samples‘. ECCV (2022).

Myles Bartlett, Sara Romiti, Viktoriia Sharmanska, Novi Quadrianto: ‘Okapi: Generalising Better by Making Statistical Matches Match‘. NeurIPS (2022).

Ainhize Barrainkua, Paula Gordaliza, José Antonio Lozano, Novi Quadrianto: ‘A Survey on Preserving Fairness Guarantees in Changing Environments‘. CoRR abs/2211.07530 (2022).