Keywords:	deep learning, machine learning, explainable AI, model compression
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

Model compression techniques (pruning, quantization, knowledge distillation, low-rank adaptation…) are commonly employed in Deep Learning to reduce the memory footprint of a neural network and allow it to run faster or in environments where memory constraint are present (e.g., microcontrollers). While these techniques are developed to increase loss as little as possible w.r.t. a reference bigger model, the resulting compressed model can behave much worse than their reference on other facets, like fairness. Thus, the idea to evaluate the performance of these models w.r.t. XAI tools to answer questions like: are the compressed models learning semantically meaningful features? Is the output of these tools faithful to the predictive dynamics of the models before/after compression? and others…

This project has a high degree of customization depending upon factors such as: modality of data (images, natural language, audio…), compression techniques (pruning, quantization, knowledge distillation, low-rank adaptation…), and architecture of models (convolutional neural networks, multilayer perceptrons, transformers…); thus, the difficulty of the thesis can be tuned depending on student’s skills and previous knowledge, ambition, and expectations.

Required work

• identify the model compression technique(s) to use
• identify the XAI tools and facets to investiagate
• literature review on the chosen topics (short live presentation required to assess understanding)
• identify datasets depending on the chosen data modality
• experiment & analyze

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

Relevant literature

• Survey on model compression: Marinó, Giosué Cataldo, et al. “Deep Neural Networks Compression: A Comparative Survey and Choice Recommendations.” Neurocomputing, vol. 520, Feb. 2023, pp. 152–70. ScienceDirect, https://doi.org/10.1016/j.neucom.2022.11.072.
• 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.
• Pruning and XAI: Weber, David, et al. Less Is More: The Influence of Pruning on the Explainability of CNNs. arXiv. 2023. https://doi.org/10.48550/arXiv.2302.08878.

Previous theses @ RUG

There have been previous thesis works centered on this topic in the past:

• Sander Hofstra, “Concern over contrast: the influence of pruning on the interpretability of GPT2.” 2023. (interpretability for text data using random pruning on GPT2)
• Thijs Lukkien, “The effect of unstructured pruning on the explanatory quality of heatmaps; a quantitative analysis”. 2023. (overlap between saliency maps and ground truth on convolutional neural networks trained on MNIST)