Improving Deep Learning Explainability with Interpretable Outputs and Representations

Anders Christensen (Ph.D. Student)

Through recent years it has become abundantly clear how powerful machine learning is. Still, despite the empirical success of deep learning models, there is rarely access to the reasoning behind why a model yields a certain output for some input. Achieving more insight into the decision making of such models and developing techniques for interpreting them is therefore crucial. Better understanding will allow models to be used as advisory tools in a more trustworthy manner if they can be understood and verified by domain experts in e.g., healthcare. In this project we will therefore attempt to develop tools and models that allows for peeking inside the “black box” of deep learning models. One such example is extending models designed for zero-shot classification with interpretable intermediate outputs, thereby assisting in verification of generalization to unseen classes.

Primary Host:	Ole Winther (University of Copenhagen & Technical University of Denmark)
Exchange Host:	Zeynep Akata (University of Tübingen)
PhD Duration:	01 December 2021 - 30 November 2024
Exchange Duration:	01 September 2022 - 31 August 2023 - Ongoing