Cohere For AI - Guest Speaker: Lucas Lehnert, Meta AI FAIR

Abstract:While Transformers have enabled tremendous progress in various application settings, such architectures still lag behind traditional symbolic planners for solving complex decision making tasks. In this work, we demonstrate how to train Transformers to solve complex planning tasks and present Searchformer, a Transformer model that optimally solves previously unseen Sokoban puzzles 93.7% of the time, while using up to 26.8% fewer search steps than standard $A^$ search. Searchformer is an encoder-decoder Transformer model trained to predict the search dynamics of $A^$. This model is then fine-tuned via expert iterations to perform fewer search steps than $A^$ search while still generating an optimal plan. In our training method, $A^*$'s search dynamics are expressed as a token sequence outlining when task states are added and removed into the search tree during symbolic planning. In our ablation studies on maze navigation, we find that Searchformer significantly outperforms baselines that predict the optimal plan directly with a 5-10$\times$ smaller model size and a 10$\times$ smaller training dataset. We also demonstrate how Searchformer scales to larger and more complex decision making tasks like Sokoban with improved percentage of solved tasks and shortened search dynamics.

About the speaker: Lucas Lehnert is a postdoctoral researcher as part of the FAIR team at Meta. Lucas' research focuses on the underlying computational principles of intelligent decision making and how intelligent agents can generalize and re-use previously learned knowledge. Before joining Meta, he worked at the Mila Quebec AI Institute and the Université de Montréal. In 2021 Lucas completed his PhD at Brown University in Computer Science and before that his MSc and BSc in Computer Science at McGill University. His interdisciplinary work was recognized by an NIH training grant in interactionist cognitive neuroscience and a best student workshop paper award. Dr. Lucas Lehnert will be joining Computer Science department at the University of Saskatchewan as an Assistant Professor in July 2024.

Add event to calendar

Apple Google Office 365 Outlook Outlook.com Yahoo