Mice given balanced early practice adapted better, while poorly structured training led to rigid strategies and repeated errors.
Researchers trained recurrent neural networks and mice on versions of an odor-timing task, using different sequences of early training. Mice and networks that lacked well-structured early experience tended to develop fixed strategies and repeat errors when faced with more complex versions of the task.
Mice given more balanced early training generalized better to new task demands. Their brain activity showed patterns related to the task’s underlying timing structure, similar to patterns in the networks. Dynamical-systems analyses indicated that the networks with structured training developed multiple activity patterns that supported the needed abstractions, whereas other networks settled on single fixed-point solutions that did not generalize beyond the training stimuli.
How early training changed learning
Early experience affected how well mice and recurrent neural networks could adapt to new versions of an odor-timing task. Training without a well-structured sequence, or with later remedial training, was associated with rigid strategies and repeated errors. More balanced early training was associated with better generalization in mice and with neural activity patterns that reflected the task’s temporal structure.
In the network analyses, appropriately structured training produced distinct patterns of ongoing activity that supported the correct abstractions as task complexity increased. Networks without suitable early training developed single fixed-point solutions and failed to generalize beyond the stimuli used during training.
Why training structure matters
The findings suggest that flexible learning depends not only on whether an animal or model has prior experience, but also on how that experience is arranged. This provides a possible explanation for why early training can either support abstraction—recognizing the underlying pattern—or leave learners tied to the specific situations they have already seen.
The similar results in mice and neural networks also point to a shared relationship between training structure, task representations and the ability to adapt when new demands are added. The study therefore links the design of learning experiences to both behavior and neural dynamics.
Evidence and open questions
The evidence combines recurrent-neural-network experiments with behavioral and electrophysiological recordings from mice trained on the same odor-timing task using staged training sequences. Dynamical-systems analyses were used to examine how activity patterns changed as the task became more complex.
The abstract does not report the number of mice or networks, the size of the effects or how well the findings extend beyond this task and training design. The network analyses identify a proposed mechanism for the behavioral pattern, but they do not by themselves show that the same mechanism fully explains learning in the mouse brain.
// Source
Nature Neuroscience · 2026 · DOI: 10.1038/s41593-026-02409-7
Authors: John C. Bowler, Dua Azhar, Cambria M. Jensen, Hyun‐Woo Lee, James G. Heys
Institutions: University of Utah
