Friday, 10th November - 11am Nico Daheim : Seminar

Title: Adapting Language Models by Model Merging

Abstract:

Adapting models to new data and the needs of their users is a fundamental challenge in NLP.

Recently, combining the parameters of multiple models, often called model merging, has emerged as a technique used to tackle many such adaptation problems.

In this talk, we show how model merging can be used to not only combine the knowledge of multiple models but also to enforce specific, fine-grained behaviours. This is particularly attractive for post-hoc “editing” of models, for example, to remove hallucinations from neural language generators.

We show that the inaccuracies of such model merging are connected to a mismatch in the gradients and propose a new method to reduce this mismatch.

Our method reveals implicit assumptions in existing techniques, which can be derived as special cases. Finally, we show relationships to influence functions and Bayesian approaches, which can be used for further improvements.

Bio:

Nico Daheim is a second year ELLIS PhD student supervised by Prof. Iryna Gurevych at TU Darmstadt and Prof. Mrinmaya Sachan at ETH Zurich. His research interests are controlling and adapting language models, natural language generation, and dialogue systems. Previously, he obtained an M.Sc. in Data Science at RWTH Aachen University under the supervision of Prof. Hermann Ney.