Belief-Guided Meta-Reinforcement Learning via a Biologically Inspired Framework for Attention and Decision Making

(Co-)Principal Investigators

Samuel F. Feng (Sorbonne University Abu Dhabi, United Arab Emirates)

Jacqueline Gottlieb (Columbia University, New York, United States of America)

Co-Investigator

Massimo Silvetti (Institute of Cognitive Sciences and Technologies (ISTC), Rome, Italy)

Project Overview

This joint research project between Sorbonne University Abu Dhabi and Columbia University explores how humans and intelligent systems make decisions when information is incomplete. By combining insights from neuroscience, decision theory, and artificial intelligence, the project develops a new learning framework that allows agents to reason about uncertainty, actively seek information, and adapt their decisions accordingly.

The project extends the Reinforcement Meta-Learner (RML) framework by integrating Rao’s Recurrent Inference Framework (RIF), transforming the model’s volatility estimate into a full Bayesian belief over hidden task states. Beliefs are updated with new observations and prior predictions, allowing actions to be evaluated based on both expected reward and remaining uncertainty. The model also introduces an explicit information-sampling action when uncertainty is high, unifying exploration and decision-making. Over 18 months, the framework will be validated through behavioral simulations and neural data analysis, bridging neuroscience, decision theory, and artificial intelligence.

Project Co-Leadership

The project is jointly led by Dr. Samuel F. Feng (SUAD / SAFIR Institute) and Dr. Jacqueline Gottlieb (Columbia University / Zuckerman Institute), with Dr. Massimo Silvetti (ISTC-CNR, Italy) as a co-investigator. Dr. Feng leads the theoretical development, model design, and coordination of milestones, meeting regularly with the postdoctoral researcher and collaborators. Dr. Gottlieb co-supervises the postdoctoral researcher and provides access to behavioral and neural datasets, computational infrastructure, and laboratory resources. Dr. Silvetti contributes technical expertise from the original RML framework and supports manuscript preparation. Regular coordination meetings ensure smooth collaboration and shared responsibility for all deliverables.

Expected outcomes

The project will deliver a fully implemented, open-source computational framework, at least two high-impact journal publications, one indexed conference presentation, and the training of a postdoctoral researcher jointly supervised by SUAD and Columbia University. A final report to SAFIR will summarize outcomes, collaboration, and future directions.