Successful PhD defense by Srishti Gautam

We congratulate Srishti Gautam for successfully defending her PhD thesis and achieving the PhD degree in Science at UiT The Arctic University of Norway on March 15th 2024.

By: Petter Bjørklund, Communications advisor, Visual Intelligence.

Gautam is a researcher at UiT Machine Learning group and the research centre. Her thesis, "Towards Interpretable, Trustworthy and Reliable AI", focuses on enhancing the interpretability of deep learning through the development of self-explainable models.

The title of Gautam's trial lecture was "Vision-language models: applications, limitations and future directions".

Supervisors:

• Associate professor Michael Kampffmeyer, UiT Machine Learning Group, Department of Physics and Technology, UiT (main supervisor)
• Postdoctoral researcher Ahcene Boubekki, Physikalisch-Technische Bundesanstalt (co-supervisor).
• Professor Robert Jenssen, UiT Machine Learning, Department of Physics and Technology, UiT (co-supervisor).

Evaluation committee:

• Professor Georgios Leontidis, Chair of Machine Learning, lnterdisciplinary Director of Data and Al, Turing Academic Liaison, Vice-Principals' Office, University of Aberdeen, UK (1. Opponent)
• Professor Lilja Øvrelid, Language Technology Group, Section for Machine Learning, Department of lnformatics, University of Oslo, Oslo, Norway (2. Opponent)
• Associate Professor Elisabeth Wetzer, Machine Learning Group, Department of Physics and Technology, UiT (intern member and leader of the committee)

Interview with Srishti Gautam

Could you provide a short summary of your thesis?

In the rapidly evolving field of artificial intelligence (AI), the development of deep learning models has marked a significant milestone, enabling breakthroughs across various applications. However, these advancements have also surfaced critical challenges, notably the models' vulnerability to inheriting biases from their training data. This issue can further be compounded by these large models’ inherent lack of transparency in their decision making. Such issues not only undermine thetrust in these technologies but also pose a barrier to their widespread adoption.

Recognizing the importance of these concerns, my thesis focuses on enhancing the interpretability of deep learning through the development of self-explainable models. These models aim to shift the paradigm towards more transparent AI systems by integrating explanations directly into their architecture, thereby offering direct insights into their decision-making processes. Further, we address the inadvertent learning of biases in deep learning by putting these self-explainable models into action.

Why have you focused on this particular topic? What is the importance of researching this topic?

I have chosen to focus on this topic because of the critical role AI plays in our lives today and its potential for even greater impact in the future. As AI technologies become increasingly integrated into various sectors—ranging from healthcare and education to finance and security—the need for these systems to be transparent, fair, and trustworthy becomes paramount.

Bias in AI can lead to unfair outcomes, such as discrimination against certain groups, while opacity in AI decision-making processes can prevent users from understanding and trusting the results. By developing self-explainable models that are inherently transparent and designed to detect biases, this research aims to foster a generation of AI systems that are not only high-performing but also equitable and comprehensible. This is crucial for ensuring that AI technologies benefit society as a whole, facilitating their widespread adoption in a responsible and ethical manner.

What methods have you used in your thesis?

In my thesis, I employed a multi-faceted approach to develop and enhance AI models, including:

1. Enhancement and Development of Self-Explainable Models: I designed novel self-explainable models that integrate explanations directly into their architecture. This approach allows the models to provide insights into their decision-making processes inherently, making them more transparent and understandable. Further, I introduced a novel algorithm aimed at improving the explanation quality of existing state-of-the-art self-explainable models. This algorithm enhances the clarity and relevance of the explanations provided by the models, making it easier for users to understand the rationale behind AI decisions.

‍

2. Counteracting Data Artifacts: An important aspect of my research involved identifying and mitigating the learning of artifacts—spurious correlations that models might pick up from the training data. By focusing on this, the methodology helps in reducing the inadvertent perpetuation of biases, ensuring that the models make decisions based on relevant features rather than biased or irrelevant correlations.

3. Fairness in Large Language Models: Given the increasing use of large language models in various applications, my thesis also extends to exploring fairness within these models. This involves analyzing and demonstrating how such models can reinforce social biases, specifically against gender and race, and exploring strategies to mitigate these biases, thereby promoting fairness.

‍

What significance may your results have on society/the general public?

The results of my research have the potential to significantly impact various sectors of society and the general public. By enhancing AI transparency and fairness,consumers and end-users gain access to more reliable and understandable AI-driven services. Marginalized and underrepresented groups stand to benefit from efforts to reduce biases in AI, aiming to ensure equitable applications and prevent the perpetuation of social inequalities. For example, in healthcare, transparent and unbiased AI can lead to more accurate diagnoses and treatments, supporting fair medical decisions for patients. Businesses and organizations using self-explainable AI models can build customer trust and ensure regulatory compliance, fostering ethical practices. Additionally, the development of transparent and fair AI models can provide valuable insights for policy makers and regulators, informing better guidelines and regulations for AI use.

Summary of the thesis

The field of artificial intelligence recently witnessed remarkable growth, leading to the development of complex deep learning models that perform exceptionally across various domains. However, these developments bring forth critical issues. Deep learning models are vulnerable to inheriting and potentially exacerbating biases present in their training data. Moreover, the complexity of these models leads to a lack of transparency, which can allow biases to go undetected. This can lead to ultimately hindering the adoption of these models due to a lack of trust. It is therefore crucial to foster the creation of artificial intelligence systems that are inherently transparent, trustworthy, and fair. This thesis contributes to this line of research by exploring the interpretability of deep learning through self-explainable models. These models represent a shift towards more transparent systems, offering explanations that are integral to the model’s architecture, yielding insights into their decision-making processes. Consequently, this inherent transparency enhances our understanding, thereby providing a mechanism to address the inadvertent learning of biases.

‍

‍