Keywords:	deep learning, machine learning, explainable AI
Prerequisites:	neural networks, knowledge of deep learning (also self-taught is good)
Difficulty:	variable (depends on the chosen topics)
Group work (only for B.Sc.):	possible

Abstract

With eXplainable AI (XAI) being a hot topic in recent years, a lot of research has been published presenting techniques for input attribution, i.e., identifying which parts of a given input are more important/relevant for the prediction of a model, with specific regard to neural networks, due to their inherent black-box nature. Often, though, these tools are mere approximations which can fail to identify where the “attention” of a model lies. With so many tools having been published recently, little interest has been given at identifying quantitatively pros and cons of these techniques. The goal of this project is hence to identify a homogeneous set of XAI tools and evaluate them on one or more benchmark datasets, with the goal of assessing their quality on one or more facets.

Required work

• identify (& implement) the XAI tools
• identify the facets of explainability to investiagate
• identify datasets (bonus for different modalities—images, audio, text…)
• experiment & analyze

Python + PyTorch is the preferred language/library, but I’m open to considering also other options

Relevant literature

• Openly-available course on XAI: CHI2023
• Assessing quality of XAI explanations: Nauta, Meike, et al. From Anecdotal Evidence to Quantitative Evaluation Methods: A Systematic Review on Evaluating Explainable AI. arXiv:2201.08164, arXiv, 31 May 2022. arXiv.org, https://doi.org/10.48550/arXiv.2201.08164.
• Example of when XAI fails to produce reliable outputs: Arrighi, Leonardo, et al. A Survey on Limitations and Opportunities in XAI. 2023.