Making computer language models understandable and controlable
Researcher James Henderson (Idiap Research Institute) received an ERC Advanced Grant to work on large language models (LLMs), deep learning computational models that can perform language-related tasks.
This project is a continuation of his work within the NCCR Evolving Language which improves our understanding of what makes LLMs so successful for human languages. It will leverage this theoretical understanding to improve the interpretability (i.e. understanding why certain predictions are made) and controllability (i.e. influencing the behaviors) of LLMs.
“In this new project, we’ll address the interpretability and controllability of LLMs by augmenting them with a memory,” explains the researcher. “We’ll then focus on making the beliefs and knowledge in this memory easier to understand and modify.”
