Search Papers | Poster Sessions | All Posters
Poster A21 in Poster Session A - Tuesday, August 6, 2024, 4:15 – 6:15 pm, Johnson Ice Rink
A Domain-general Strategy for Hidden-state Inference in Humans and Noisy Neural Networks
Junseok Lee1 (), Valentin Wyart; 1Institut National de la Santé et de la Recherche Médicale, 2École normale supérieure
Understanding the hidden (latent) states and structures that generate observations of the world is a fundamental aspect of cognition, wherein humans demonstrate exceptional proficiency in the ability to apply similar cognitive strategies across superficially dissimilar contexts sharing the same latent structure. While previous efforts to understand the computational bases of these cognitive strategies through cognitive modeling have largely focused on single contexts, here we take a novel approach which combines two tasks with the same reversal structure. Through cognitive modeling, we first show that humans use the same noisy hidden-state inference strategy across these superficially dissimilar tasks that require reversal learning. Then, using recurrent neural networks (RNNs) featuring either exact or noisy computations trained on the same tasks, we show that noisy RNNs – like humans – benefit from reusing the same latent representations for solving the two tasks. Together, our findings underscore the significance of computation noise in constraining the use of mental resources, shed- ding light on its potential functional role in cognition.
Keywords: hidden-state inference human behavior recurrent neural networks