Learning Rate Dynamics Across Training Phases in Mouse Visual Decision-Making: A Population-Level PsyTrack Analysis

Abstract

How quickly do decision variables continue to change as perceptual expertise develops? Accuracy and contrast sensitivity have been studied extensively, but the dynamics of the underlying decision weights are much less well characterized. Contrary to the expectation that learning rates should decline with expertise, we find that within-session contrast smoothness increased across training, rising by 116% from early to late phases (p<0.001, r = 0.81). We applied PsyTrack, a time-varying logistic regression model, to 54 Tier 1 mice from 9 International Brain Laboratory sites, treating the model’s smoothness hyperparameters as chunk-level proxies for effective learning rate. A four-weight model (bias, signed contrast, previous choice, and win-stay-lose-shift) was fit independently to each of three curriculum-defined training phases (Chunk 1: high contrasts only; Chunk 2: intermediate contrasts added; Chunk 3: full stimulus set including near-threshold and 0% contrasts). The contrast smoothness increase was selective: bias and history terms showed no comparable rise, arguing against a channel-general fitting artifact, though not a contrast-specific noise explanation. A mixed-effects model recovered the same population-level pattern (+108%, p<0.001). Overnight contrast jump magnitudes also increased, whereas jump direction showed no consistent per-mouse strengthening trend. Bias weights showed overnight correction toward zero during the curriculum expansion phase (per-animal p = 0.015). The ratio of previous-choice to win-stay-lose-shift smoothness separated the population into 32.1% pure perseverators, 24.5% reward-sensitive mice, and 43.4% mixed-strategy mice, indicating multiple viable routes to proficiency. The main result is not simple late-stage stabilization but continued movement in the contrast weight under the full stimulus set. The main caveat is that Chunk 3 introduces near-threshold stimuli and a possible compression-scaling confound. However, a high-contrast control restricting Chunk 3 to physical contrasts ≥25% retained roughly 90% of the original effect, indicating that the lowest-contrast trials are not the primary driver. Together, the analyses enable a phase-matched, channel-resolved comparison of within-session and overnight learning dynamics.