Rupsa Saha has followed the PhD Programme at Faculty of Engineering and Science, specialisation in ICT, Scientific field of Artificial Intelligence.
- Trial lecture starts at 10:15
- Public defence starts at 12:15
Title of trial lecture: "Explainability in LLMs: Methods, Metrics, Challenges"
Disputation chair: Morgan Konnestad
Assessment committee:
- Professor Christina Lioma, University of Copenhagen
- Associate Professor Karl Øyvind Mikalsen, UiT Arctic University of Norway
- Professor Andreas Prinz, UiA
Supervisors in the doctoral work:
- Main supervisor: Professor Ole-Christoffer Granmo, UiA
- Co-supervisor: Professor Morten Goodwin, UiA
Summary of thesis
Towards a Relational Tsetlin Machine in Natural Language Processing
Artificial Intelligence (AI) is becoming a crucial part of our daily lives, influencing everything from personal assistants to societal systems. However, as AI’s role grows, so does the need for these systems to explain their decisions. Explainable AI (XAI) comes in aims to make AI’s decision-making processes transparent, helping to build user trust and prevent misuse or bias.
One of the key areas where humans interact with AI is through natural language. Understanding language requires logical structures, and integrating logic programming into natural language processing (NLP) can enhance AI’s ability to process language. Language is complex and multifaceted, so AI systems must consider various aspects of each text. By using Relational Modeling, AI can break down complex information into simpler parts, making it easier to understand and process.
Tsetlin Machines (TMs) are a type of AI that uses learning automata to make interpretable decisions. They work by learning patterns, or ”clauses, ”from the data they analyze. In a classification task, multiple clauses vote to determine the class of a sample. TMs have shown success in various tasks, including image classification and NLP. This thesis explores their potential as a tool for XAI in language analysis.
In recent research, it has been demonstrated that TMs can provide both global and local explanations for their decisions. Globally, the clauses learned by a TM collectively describe the task at hand. Locally, some of the clauses are used for a single test sample, and thus explains the decision for that specific sample. This dual-level explanation helps users understand how the model shifts its focus between different aspects of the text.
Moreover, TMs can build logical structures for reasoning based on relationships in natural language text. While standard TMs use specific words from the vocabulary, a modified version called the Relational Tsetlin Machine (RTM) uses variables to create generic logical clauses. This allows RTMs to capture the logical interactions in text more effectively.
Another significant contribution of this research is a framework that helps TMs (or RTMs) handle data changes. AI applications often require large amounts of data, which can vary in characteristics depending on the source or time. This is particularly true for natural language data, which changes over time and across different contexts. The TM-based system can identify and isolate samples that do not conform to the majority, mitigating their impact on model performance.
In summary, integrating XAI with TMs using a Relational approach offers a promising process to making AI systems more transparent and trustworthy, especially in the realm of natural language processing. This research paves the way for AI models that not only perform well but also explain their decisions in a human-understandable way.
What to do as an online audience member
The disputation is open to the public. To follow the trial lecture and the public defence online, register on Zoom.
We ask online audience members to join no earlier than 10 minutes in advance. After these times, you can leave and rejoin the meeting at any time.
Opponent ex auditorio:
Deadline for the public to pose questions is during the break between the two opponents. Questions ex auditorio can be submitted to Kristine Reinfjord.
Contact person
- kristine.reinfjord@uia.no
- Phone
- +47 37 23 30 14