Friday, 6th September - 11am Michael Hahn : Seminar Title: Understanding Language Models via Theory, Interpretability, and Humans Abstract: Recent progress in LLMs has rapidly outpaced our ability to understand their inner workings. This talk describes our work aiming to understand language models from three angles. First, we develop rigorous mathematical theory describing the abilities (and limitations) of transformers in performing computations foundational to reasoning. We also examine differences and similarities with state-space models such as Mamba. Second, we propose a method for reading out information from activations inside neural networks, and apply it to mechanistically interpret transformers performing various tasks. Third, we use language models to build a cognitive model of human language comprehension. I will close with directions for future research. Short Bio: Michael Hahn is a tenure-track Professor at Saarland University, Germany. He received his PhD from Stanford University in 2022, and has also done research at MIT, Facebook AI Research, and University of Edinburgh. His research interests focus on theoretical and cognitive angles on natural language processing. Among recent honors, his work received a best paper award at ACL 2024. Website: https://www.mhahn.info Sep 06 2024 11.00 - 12.00 Friday, 6th September - 11am Michael Hahn : Seminar This event is co-organised by ILCC and by the UKRI Centre for Doctoral Training in Natural Language Processing, https://nlp-cdt.ac.uk. IF G.03 and on Teams Contact
Friday, 6th September - 11am Michael Hahn : Seminar Title: Understanding Language Models via Theory, Interpretability, and Humans Abstract: Recent progress in LLMs has rapidly outpaced our ability to understand their inner workings. This talk describes our work aiming to understand language models from three angles. First, we develop rigorous mathematical theory describing the abilities (and limitations) of transformers in performing computations foundational to reasoning. We also examine differences and similarities with state-space models such as Mamba. Second, we propose a method for reading out information from activations inside neural networks, and apply it to mechanistically interpret transformers performing various tasks. Third, we use language models to build a cognitive model of human language comprehension. I will close with directions for future research. Short Bio: Michael Hahn is a tenure-track Professor at Saarland University, Germany. He received his PhD from Stanford University in 2022, and has also done research at MIT, Facebook AI Research, and University of Edinburgh. His research interests focus on theoretical and cognitive angles on natural language processing. Among recent honors, his work received a best paper award at ACL 2024. Website: https://www.mhahn.info Sep 06 2024 11.00 - 12.00 Friday, 6th September - 11am Michael Hahn : Seminar This event is co-organised by ILCC and by the UKRI Centre for Doctoral Training in Natural Language Processing, https://nlp-cdt.ac.uk. IF G.03 and on Teams Contact
Sep 06 2024 11.00 - 12.00 Friday, 6th September - 11am Michael Hahn : Seminar This event is co-organised by ILCC and by the UKRI Centre for Doctoral Training in Natural Language Processing, https://nlp-cdt.ac.uk.