On New York University Week: We’re always making decisions in life, so how can we improve our decision-making?

Marcelo Mattar, assistant professor of psychology and neural science, uses tiny AI models to find out.

Dr Marcelo Mattar is an Assistant Professor of Psychology and Neural Science at New York University. He holds a PhD in Psychology and a Master’s in Statistics from the University of Pennsylvania, and a Bachelor’s in Electronics Engineering from the Aeronautics Institute of Technology in Brazil. Prior to his current role, Dr. Mattar was a postdoctoral researcher at Princeton University under the mentorship of Nathaniel Daw and at the University of Cambridge’s Department of Engineering, working with Máté Lengyel. Dr. Mattar’s research integrates neuroscience, psychology, and computational modeling to explore how the brain supports flexible decision-making. Focusing on the interplay of learning, memory, and planning, his lab uses computational modeling, behavioral experiments, and neural recordings to understand how the brain builds internal models from experience to simulate the future and guide choices.

AI and Decision Making

Every day, we make countless decisions, learning from our successes and failures. For decades, scientists like myself have tried to map the hidden rules of decision-making using computational models. But these models present a dilemma: cognitive science models are too simple to capture behavioral complexity, while AI models predict behavior accurately but are too complex to interpret. In other words, we end up having to choose between a simplistic account we understand and an accurate account we don’t.

In my lab at NYU, we wondered if we could combine the best of both worlds. My colleagues and I developed a new approach using what we call “tiny” recurrent neural networks — a simplified AI model that functions like a miniature artificial brain making decisions as a human would.

We trained these networks to imitate choices of humans and animals in reward-learning tasks, where subjects figured out which option pays out more often.  We found that the tiny networks did a better job at imitating human decisions than the simplistic models typically used in cognitive science.

But the real breakthrough came when we looked inside these tiny networks. Because they are so small, they aren’t black boxes like typical AI systems. We can analyze their mechanics and discover the exact strategies they use to make human-like decisions. It’s like having a microscope for cognition.

We uncovered entirely new cognitive strategies that don’t appear in any neuroscience textbook.  For instance, we found that subjects don’t learn at a constant rate; their learning speed changes depending on their confidence.  We also found that, when people received no reward, they treated that action as having similar value to alternatives, not zero value.

This matters beyond academic curiosity. In computational psychiatry, understanding individual differences in decision-making could transform how we diagnose and treat conditions like addiction or anxiety. Our tiny networks can capture these individual signatures with unprecedented accuracy.

The post Marcelo Mattar, New York University – AI and Decision Making appeared first on The Academic Minute.