27 mars 2024 presenterar doktoranden Alejandro Kuratomi sitt pågående arbete med titeln ”Orange juice – Enhancing Machine Learning Interpretability”. Seminariet genomförs på Institutionen för data- och systemvetenskap (DSV) vid Stockholms universitet.

Respondent: Alejandro Kuratomi, DSV
Opponent: Professor Toon Calders, University of Antwerp, Belgien
Huvudhandledare: Tony Lindgren, DSV
Handledare: Panagiotis Papapetrou, DSV
Närmast berörda professor: Rahim Rahmani, DSV

Kontakta Alejandro Kuratomi

Sammanfattning på engelska

In the current state of AI’s development, it is reasonable to think that AI will continue to expand and be increasingly utilized across different fields, highly impacting every aspect of humanity’s welfare and livelihood. However, not all future scenarios described by AI researchers and institutions are good, i.e., AI has both the potential to be extremely beneficial, and pose existential threats to humanity.

Even if such scenarios develop in the longer term, it is necessary to start focusing now on tools that allow humans to open the black-box of the AI algorithms, to increase their understandability and trustworthiness, and in that way try to avoid such conceivably harmful future scenarios. Additionally, the interpretability of the AI algorithms is even now a challenge to the AI development, as it is an obstacle equivalent to those previously seen which triggered past AI winters, such as hardware constraints or the over-expectation of results, in other words, research in interpretability and model understanding will help avoid a third AI winter, which, if triggered, could be devastating for the world economy.

This thesis aids in researching the field of AI interpretability, focusing mainly on the algorithms related to local explainability, which are the algorithms that attempt to extract useful explanations for the output of machine learning models on individual instances, specifically targeting areas of improvement in the counterfactual and local surrogate explanations. Since, for some of the datasets, these instances are humans, opening the black-box of the machine learning models may also reveal the existence of biases or fairness issues among different groups of people.

This thesis then focuses also on the topic of machine learning fairness assessment through the use of local explainability methods, proposing two novel elements: one single accuracy-based and counterfactual-based bias detection measure and a counterfactual generation method for groups intended for bias detection and actionability-oriented fairness. Finally, the idea behind interpretability is to be able to eventually implement such methods into real-world applications. This thesis presents an application of the conformal prediction framework to a regression problem related to autonomous vehicle localization systems that both generates the predicted positioning error value of a vehicle and its confidence interval with some level of significance.