Neural networks mirror how training shapes learning in mice

A mouse is rewarded for responding to timed patterns of smells. At University of Utah Health, researchers gave a neural network a computational version of that challenge and found that the order of training changed both its accuracy and its activity patterns, just as it did in living brains.

The difficult task required a specific response after two stimuli of different durations, a “go” trial, and no response when the two stimuli were the same length, a “no-go” trial. The researchers first trained some networks only on go trials, then introduced the complete task. Those networks learned the complex version more accurately than networks trained on the full challenge from the start.

The difference showed up in the errors. Networks that skipped the simpler stage often responded too early, jumping after the first long stimulus. Living animals make the same predictable mistake when training is poorly structured, giving the computational model a way to suggest how real brains may learn—and fail.

The researchers then measured neuron-level activity in mice in the medial entorhinal cortex, a brain region involved in task learning. The patterns broadly matched those adopted by the neural networks: both moved through a cycle during a trial, ending in roughly the state where they began, and both veered from that path when a response was required. “We can use these complex models to make specific predictions,” said Jack Bowler, the study’s first author.

Concretely, the immediate benefit is for research: Bowler says modeling helped the team reach useful hypotheses faster and reduce the number of animals needed for testing. The longer-term possibility is better-structured training programs, along with new clues about diseases that disrupt complex thinking. But the work remains a study of neural networks and mice; no human learning program or treatment has been tested.