Influence-Driven Data and Input Attribution for Explainability in Machine Learning and Large Language Models

#Computing #Data
Share

On Tuesday 14th July, Ikhtiyor Nematov will give a lecture on the following subject: "Influence-Driven Data and Input Attribution for Explainability in Machine Learning and Large Language Models."


  Date and Time

  Location

  Hosts

  Registration


  • Add_To_Calendar_icon Add Event to Calendar

Loading virtual attendance info...

  • ULB Solbosch Campus,
  • Brussels, Unknown
  • Belgium
  • Room Number: R42.4.110
  • Contact Event Host

  • Please send a mail to tsagi@cs.aau.dk if you want to join the defense online.

  • Co-sponsored by Sean Bin Yang

  Speakers

Ikhtiyor Nematov of Aalborg University

Topic:

Influence-Driven Data and Input Attribution for Explainability in Machine Learning and Large Language Models

This thesis explores how data attribution can make machine learning models and Large Language Models (LLMs) more explainable —highlighting how transparency is not only a technical challenge, but also a question of understanding how data influences complex AI decisions.

These systems are widely used in critical contexts and may produce outputs that are difficult to interpret. While existing explanation methods can identify influential data points, they often struggle with outliers, user relevance, and modern generative models, creating a need for more robust and practical approaches.

Using influence-based techniques, the research develops new methods and frameworks for identifying how training data and retrieved documents shape predictions and generated responses. A key contribution is extending these methods to generative and Retrieval-Augmented Generation (RAG) systems, enabling meaningful, user-centered explanations despite complex interactions and noisy signals.

Biography:

Ikhtiyor Nematov is a PhD researcher affiliated with Université Libre de Bruxelles and Aalborg University. His research interests include Machine Learning, Explainable Artificial Intelligence, and Large Language Models. His work focuses on explainability in AI systems, particularly example-based explanations, intent-aware explanations, source attribution in retrieval-augmented generation, and input attribution methods for machine learning and large language models.

He has published research in venues related to machine learning, knowledge discovery, data engineering, and AI ethics. His recent publications include studies on source attribution in retrieval-augmented generation, antithetical and diverse example-based explanations, susceptibility of explainability methods to class outliers, and influence-driven data attribution. According to the profile shown, his work has received 15 citations, with an h-index of 3.

Ikhtiyor’s research contributes to making AI systems more transparent, interpretable, and trustworthy, especially as large language models become increasingly important in real-world applications.

Email:

Address:Selma Lagerløfs Vej 300, , Aalborg East, Denmark, 9220