PhD student position at the Department of Computing Science at Umeå University.
We are seeking one PhD student to work in formal language theory, focusing on applying theoretical computer science to the analysis of the transformer neural network architecture.
This project analyses the limits of representability and learnability in transformers, the popular neural network architecture used by many state-of-the-art systems, for example, large language models like ChatGPT and the AI Sweden/RISE/WARA M&L GTP-SW3 model. Transformers are known to be limited, allowing them to be formally analyzed in a principled way. This project leverages recent results, which characterize transformers in terms of Boolean circuits and/or logical descriptions, and aims to apply algebraic decompositions to understand and modify transformer models.
This project is fully funded by WASP. The project will be led by Dr. Martin Berglund as part of the Foundations of Language Processing Group. As a PhD student you would work as part of this group, which has a broad cross-disciplinary focus including formal and natural language processing, related machine learning techniques, and a broad mix of theoretical computer science.