Structured early training gave mice and recurrent models flexible timing strategies in odor task
Researchers led by John C. Bowler and James G. Heys compared shaped and unshaped training in recurrent models and mice; shaped models solved an odor timing task without error.

In a study published in Nature Neuroscience, a team led by John C. Bowler and James G. Heys examined how the structure of early training shapes later learning. They combined recurrent circuit models with recordings from the medial entorhinal cortex of mice nature.com
Nature Neurosciencenature.comOpen source ↗. According to the authors, without structured early experience both the models and the mice developed rigid, error-prone strategies, whereas structured training produced neural activity that reflected the temporal structure of the task.
Odor timing task and model design
The task was a temporal variant of delayed non-match to sample. Subjects evaluated two successive presentations of the same odor separated by a fixed three-second interval. Each odor lasted either two seconds (short) or five seconds (long). Trials came in pseudo-randomized blocks of four: one match trial (short-short) and two non-match types, short-long and long-short. Long-long trials were omitted. Mice had to withhold licking during the first stimulus, the delay and match trials, and lick during a three-second window after the second cue ended on non-match trials. A 250-millisecond visual flash signaled trial onset three seconds before the odor.
The model was a fully connected layer of 128 leaky recurrent units with one response unit. Models completed 1,000 blocks of four trials. Shaped networks received 300 initial blocks of non-match trials, while unshaped controls trained directly on the full task.
Differences in dynamics and connectivity
Unshaped networks made stereotyped errors: premature responses during long cues on long-short trials and weak responses on short-long trials, so output profiles overlapped and no effective decision threshold could be set. Shaped networks solved the task without error and stayed stable under growing input noise and cue timing jitter; the main effect of shaping was significant (F1,242 = 215.31, P = 2.67 × 10^-35) nature.com
Nature Neurosciencenature.comOpen source ↗.
In shaped networks the first three principal components captured more variance (t = 4.3, P = 2.78 × 10^-4), and trajectory tangling was lower than in unshaped ones (P = 0.02). Decoders of elapsed trial time worked better in shaped networks, including across trial types (t = 12.1, P = 1.3 × 10^-23).
Eigendecomposition of the recurrent weights showed two conjugate pairs of complex eigenvalues in shaped networks that were absent in unshaped ones. A classifier using these eigenvalues identified training histories perfectly. When the leading eigenvalues of naive networks were adjusted to match shaped ones during the first 300 blocks, the models reached lower mean squared errors than unshaped controls (t = 2.266, P = 0.0377).



