Structured Early Training Shapes Neural Dynamics and Strategy Learning in Mice

Investigating how early experience guides complex decision-making, neuroscientists have demonstrated that carefully structured training curricula dictate whether neural circuits discover flexible cognitive strategies or fall into rigid, error-prone habits. In a study published in Nature Neuroscience, a research team led by John C. Bowler and James G. Heys examined how learning histories reshape internal dynamical landscapes within computational models and the medial entorhinal cortex of mice. Animals frequently confront unfamiliar environments where success requires adapting previously acquired knowledge, a capacity governed not merely by experience volume but fundamentally by its structural design. To understand the neural mechanisms translating training into cognitive adaptability, the researchers combined recurrent neural network simulations with in vivo electrophysiological recordings. Their findings reveal that structured behavioral shaping embeds low-dimensional geometric attractors within recurrent circuits, establishing a scaffold that separates internal representations of elapsed time from stimulus-driven action triggers. Without this foundation, artificial models and mammalian brains converge on inflexible behavioral policies that severely limit generalization.

To probe how dynamical systems master complex cognitive behaviors, the investigators employed an olfactory timing task designed to exceed simple stimulus-response associations. The paradigm, a modified temporal variant of delayed non-match to sample, required subjects to evaluate two successive presentations of an identical odorant separated by a fixed three-second inter-stimulus interval. Each stimulus lasted either two seconds, designated as a short cue, or five seconds, designated as a long cue. In the full task, trials appeared in pseudo-randomized blocks of four, comprising one match trial type where both cues were short and two non-match trial types featuring differing cue durations: short-long and long-short combinations. Long-long match trials were omitted during baseline training to maintain consistency with prior work. Correct performance demanded that mice withhold licking during the initial stimulus, throughout the delay, and during match trials, while licking during a three-second window following the second cue offset on non-match trials. A brief 250-millisecond visual flash signaled trial onset three seconds prior to odor delivery.

The computational component utilized recurrent neural networks constructed from a fully connected layer of 128 leaky recurrent units governed by hyperbolic tangent activation functions and a decay constant of 0.2. Each unit received two input signals consisting of a trial-start indicator and a timed-cue marker, both one-hot encoded and disturbed by additive Gaussian noise with zero mean and standard deviation of 0.15, alongside private internal unit noise drawn from a normal distribution with standard deviation of 0.3. The recurrent layer projected through feedforward connections to a single response unit trained to match a target output profile corresponding to the valid reward window. Networks were optimized using backpropagation through time and the Adam algorithm with learning rate 0.00001, mini-batch sizes of two, and L2 regularization applied to weights and activations. Models completed 1,000 blocks of four trials, with shaped networks receiving 300 initial blocks of non-match trials, while unshaped controls trained directly on the full task.

Direct comparisons between shaped networks and unshaped controls exposed pronounced differences in behavioral competence and robustness against sensory noise. Unshaped networks developed stereotypical errors characterized by premature responses during long cues on long-short trials and blunted response amplitudes during short-long sequences, creating overlapping output profiles that precluded defining an effective decision threshold. In contrast, networks receiving structured shaping solved the task without error, exhibiting stability when subjected to escalating input noise and cue timing jitter. A two-way mixed analysis of variance revealed a significant interaction between input noise and shaping experience (F10,242 = 2.01, P = 0.0013), alongside a main effect of shaping (F1,242 = 215.31, P = 2.67 × 10^-35). Introducing temporal jitter to cue start times and durations similarly demonstrated a significant interaction with shaping history (F8,198 = 2.08, P = 0.0392; main shaping effect F1,198 = 245.24, P = 1.71 × 10^-36). Whereas unshaped systems functioned as leaky evidence accumulators prone to early triggering, shaping instilled robust dynamical constraints.

Examining hidden layer activations through principal component analysis demonstrated that structured experience reconfigures state-space geometry. Unshaped networks produced tangled, high-dimensional activation trajectories with poorly resolved structure, whereas shaped models confined their activity to a low-dimensional manifold where the first three principal components captured significantly more total variance (independent t-test, t = 4.3, P = 2.78 × 10^-4). Trajectory tangling, a metric assessing whether divergent temporal states occupy overlapping state-space positions, was significantly higher in unshaped networks than in shaped counterparts (P = 0.02). To test whether this low-dimensional confinement supported conceptual abstraction, researchers implemented linear discriminant analysis decoders to predict elapsed trial time to the nearest second. Decoders trained on specific trial types successfully decoded time on held-out trials across differing trial types, with shaped networks significantly outperforming unshaped controls in cross-context decoding (t = 12.1, P = 1.3 × 10^-23) and within-context decoding (t = 3.9, P = 2.3 × 10^-4). This cross-context generalization demonstrated that shaped networks abstracted elapsed trial time independently of stimulus identity.

Analyzing recurrent weight matrices showed that behavioral shaping left a persistent structural imprint within circuit connectivity. When nodes were ordered by their peak activation times, both network types displayed local excitatory connections between temporally adjacent units and inhibitory projections between temporally distant units, matching classic ring attractor topologies. However, eigendecomposition of the recurrent weight matrices unmasked a definitive structural divide: shaped networks consistently possessed two prominent conjugate pairs of complex eigenvalues that were absent in unshaped architectures. A cross-validated random forest classifier trained on the eight-dimensional space of these leading eigenvalues classified network training histories perfectly, significantly exceeding 1,000 shuffled-label controls (P < 0.001). To establish whether these spectral features causally generated shaped dynamics, researchers intervened in naive networks early in training, adjusting their leading recurrent eigenvalues to match shaped networks across the first 300 blocks. These intervened models converged on shaped dynamical solutions, yielding significantly lower mean squared errors compared to unshaped controls (independent t-test, t = 2.266, P = 0.0377).

Dynamical systems analysis of fixed and slow points within state space clarified the exact mechanism generated by structured pretraining. By identifying locations where network velocity dropped below 10^-6, the researchers revealed that shaped models operated via an orthogonal mechanism termed the Detector + Timer strategy. During intervals when sensory cues were absent, network states followed a stable limit cycle spanning the first two principal components, reliably encoding trial phase across the start cue, odor intervals, and the inter-stimulus delay. When an odor stimulus appeared, dynamics switched toward an isolated fixed point offset along the third principal component, allowing cue duration to be tracked orthogonally to trial phase. Peak output responses emerged only when activity returned from this offset fixed point toward the limit cycle within the defined response phase. In unshaped networks, dynamics collapsed along a single line attractor where prolonged cues pushed state vectors directly toward the response subspace via leaky integration, causing premature responses during long initial cues.

Investigating whether any partial training curriculum could instill functional abstractions, the researchers evaluated shaping protocols featuring only single non-match trial types. Training networks exclusively on short-long trials yielded no significant performance benefit over unshaped models (one-way ANOVA across four groups, F3,41 = 7.91, P = 2.82 × 10^-4; Tukey’s post hoc comparison, P = 0.0354 versus shaped models, while unshaped versus short-long shaping was not significant). Networks trained only on short-long sequences adopted a flawed heuristic of responding after any long cue, producing substantial errors during full task testing. Conversely, training networks exclusively on long-short trials produced bimodal outcomes: certain models successfully formed circular state-space manifolds and generalized, whereas others failed entirely and mimicked unshaped dynamics. The authors concluded that experiencing diverse temporal contexts during initial shaping is essential, as exposure to multiple non-matching sequences forces recurrent circuits to decouple elapsed duration from specific stimulus sequences, preventing convergence on suboptimal behavioral strategies.

To validate model predictions in living biological systems, the researchers trained adult C57BL/6 mice on the cue timing paradigm while chronically recording single-unit activity from the medial entorhinal cortex using Neuropixels 2.0 probes. Mice trained under a short-long only curriculum perseverated on a lick-after-long heuristic during eight days of full task testing, achieving approximately 75% accuracy by succeeding on short-long and match trials while failing on long-short trials (Welch’s t-test on day eight performance, t = 4.2, P = 0.001 compared to fully shaped mice). Electrophysiological recordings revealed trial-time-locked receptive fields across both groups. However, cross-context time decoding using linear decoders was severely impaired in mice trained under the single-cue protocol compared to animals receiving comprehensive shaping (independent t-test, t = 3.2, P = 8.72 × 10^-3). Furthermore, decoding errors were significantly modulated by trial outcome and training history (two-way mixed-effect ANOVA, shaping experience F1:700 = 8.360, P = 0.004; outcome F1:700 = 54.591, P = 4.24 × 10^-13), while neural trajectories in fully shaped mice occupied a lower-dimensional state space (t = 4.9, P = 2.73 × 10^-3).

The predictive validity of the framework was further demonstrated when models and trained mice were evaluated on novel probe configurations comprising 20% of presentation blocks. Testing included long-long pairings of five seconds each, medium-medium sequences of 3.5 seconds each, and single extra-long ten-second exposures. Because the learned attractor landscape channels activity into the response-potent space whenever sensory stimuli pull dynamics toward the Cue On fixed point during the appropriate temporal phase, the model predicted that subjects would initiate licking on these untrained sequences. Experimental animals robustly confirmed this prediction, exhibiting response probabilities significantly higher than match baseline rates across all probe conditions (two-factor ANOVA trial type effect, F5,54 = 72.5, P = 1.03 × 10^-22; Tukey’s post hoc comparisons versus match trials, P < 1.3 × 10^-14 for all probes). Although simplified networks omitted biophysical spike generation and multi-task forgetting mechanisms, the alignment between computational dynamics and entorhinal circuit activity illustrates how simulated systems can guide cognitive curriculum design in biological experiments.

Subject of Research: Neural dynamics and curriculum learning in the medial entorhinal cortex

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

Article References: Bowler, J. C., Azhar, D. B., Jensen, C. M., Lee, H.-W., & Heys, J. G. (2026). Structured experience shapes strategy learning and neural dynamics in the medial entorhinal cortex. Nature Neuroscience, 29(10), 2552-2565. https://doi.org/10.1038/s41593-026-02409-7

Image Credits: AI Generated

DOI: 10.1038/s41593-026-02409-7

Keywords: Medial Entorhinal Cortex, Recurrent Neural Networks, Curriculum Learning, Dynamical Systems, Attractor Dynamics, Temporal Non-Match to Sample, Neuropixels Recordings, Neural State Space, Time Decoding, Principal Component Analysis, Eigenspectrum Analysis, Behavioral Shaping, Cognitive Flexibility, Mouse Electrophysiology, Trajectory Tangling, Linear Discriminant Analysis