0.2 Structured experience creates reusable structure

A recent study by Bowler and collaborators provides an unusually direct experimental bridge between this developmental picture, artificial networks, and animal learning [ Bowler et al. , 2026 ] . The study asks not merely whether prior training helps, but whether the organization of that training determines which later strategy becomes learnable. Recurrent neural networks and mice learned a temporal delayed non-match-to-sample task. Two odor cues could be short or long; the learner had to respond after non-matching pairs and withhold a response for a matching pair.

The critical manipulation was the curriculum. In the fully structured shaping condition, learners first experienced both orders of the non-matching primitive: short–long and long–short. An impoverished curriculum presented only short–long before the full task. Both curricula offered experience, reward, and a genuine subtask. They did not, however, make the same latent organization identifiable. Exposure to both orders forced a distinction between detecting cue structure and tracking elapsed time. Exposure to only one order admitted a simpler shortcut—roughly, respond after a long cue—that earned substantial reward but failed on the reversed composition.

The consequences persisted. Shaped RNN activity occupied fewer dimensions, followed less tangled trajectories, and exhibited dynamical motifs that separated cue detection from timing; they generalized trial time more effectively across contexts and were more robust to noise and temporal jitter. Networks without the appropriate shaping often converged to rigid error patterns. The animal comparison showed the same qualitative dependence: mice receiving the standard two-order shaping curriculum outperformed mice receiving short–long-only shaping after eight days of the full task, and medial entorhinal cortex recordings showed stronger cross-context temporal structure. The principal behavioral comparison used 16 fully shaped and 4 single-order-shaped mice, so the study is evidence for the mechanism rather than a final measurement of its generality.

The ORACLE interpretation is sharper than the generic claim that curriculum matters. A curriculum is part of the presentation of the world. Different presentations can expose the same observations and rewards while inducing different observational quotients. The single-order presentation failed to separate a task-generating temporal construction from a reward-compatible shortcut. The two-order presentation supplied a counterposed probe: any hypothesis had to account for the same temporal relation under both orders. That pressure favored a reusable decomposition whose components could be recombined when the matching case was introduced.

This result also gives empirical content to persistence. Early structure did not merely accelerate optimization toward the same endpoint. It altered the dynamical organization and constrained which later strategy was reached. Conversely, poorly structured experience created a locally adequate strategy that both mice and RNNs found difficult to revise. In Piagetian terms, a curriculum can prepare schemas for conservative accommodation, or it can make later accommodation expensive by allowing brittle assimilation to harden. Nothing in the experiment proves that the brain represents a category. It does show that the ordering and compositional coverage of experience can determine whether transferable structure emerges at all—exactly the phenomenon a categorical theory of lifelong learning must explain.