Structured experience shapes strategy learning and neural dynamics in the medial entorhinal cortex

TL;DR

Structured early training helps animals and AI learn better by shaping flexible strategies and neural patterns.

Contribution

Shows how the structure of early experience, not just its presence, affects learning flexibility and neural dynamics.

Findings

Structured early training leads to better generalization and flexible strategies in both mice and RNNs.

Unstructured training results in rigid strategies and repeated errors in learning complex tasks.

Dynamical motifs in neural networks trained with structured curricula support generalizable abstractions.

Abstract

Animals can solve new, complex tasks by reusing and adapting what they’ve learned before. This kind of flexibility depends not just on having prior experience, but on how that experience was structured in the first place. The design of early training curriculum is especially important: poorly structured experiences can hinder abstraction and limit generalization, while carefully structured training promotes more flexible and adaptive behavior. Yet, the neural mechanisms supporting this process remain unclear. To investigate how early training shapes learning we first trained recurrent neural networks (RNNs) on variants of an odor-timing task previously used to study complex timing behavior in mice. We then tested the RNN predictions on how previous experience affects generalization using behavioral and electrophysiological recordings in mice trained on the same task using staged training sequences. RNNs and mice trained without well-structured early experience developed rigid strategies and made repeated errors. In contrast, those given more balanced early training were better able to generalize and showed similar neural activity patterns that reflected the task’s underlying temporal structure. Using dynamical systems approaches, we reveal a mechanism for this effect: networks trained with appropriately structured curricula developed distinct dynamical motifs that support the correct abstractions when complexity was increased. Networks that lacked early training or received remedial curricula developed single fixed-point solutions that failed to generalize beyond the training stimuli. Together, these findings demonstrate that it is not just the presence of prior experience, but its structure, that governs how flexible and generalizable knowledge emerges in both biological systems and computational models.

Figures