How Humans Learn Like Bayesians
A concise, practitioner-friendly synthesis showing how everyday cognition mirrors Bayesian updating. We also summarize a recent Metropolis–Hastings (MH) forecast from Dispensight’s pipeline and connect it to human learning dynamics.
TL;DR
- Humans carry priors (beliefs) that are updated by evidence in proportion to its diagnosticity and reliability.
- Attention, memory, and learning-rate act like implicit hyperparameters that tune update size (akin to step size / posterior curvature).
- Noise and uncertainty are not bugs—our brains keep distributions, not single numbers, enabling robust decisions under ambiguity.
- Dispensight’s MH regression shows the same principles at work: strong prior structure, likelihood from data, and uncertainty carried into the forecast.
Human Learning → Bayesian Building Blocks
1) Priors = Expectations
We arrive with expectations shaped by experience. In the brain, these act like priors that compress the hypothesis space before evidence arrives.
- Habit & schema ≈ structured priors
- Anchoring ≈ over-confident priors
2) Likelihood = Evidence Use
Humans weigh evidence by how surprising it would be if our beliefs were wrong. Diagnostic cues get larger updates; noisy cues get damped.
- Source reliability sets noise level (σ²)
- Signal-to-noise → update magnitude
3) Posterior = Updated Belief
Beliefs become posteriors, ready to act on—yet still uncertain. We carry intervals, not certainties, and revise with new data.
4) Learning-Rate & Plasticity
Neuromodulators (e.g., dopamine, norepinephrine) modulate plasticity like an adaptive learning-rate: volatile environments → bigger updates; stable ones → smaller.
5) Regularization = Biases That Help
Heuristics look like constraints that prevent overfit: simplicity bias, recency weighting, and sparsity all map to regularization.
6) Exploration ↔ Exploitation
Humans balance trying new actions vs. exploiting known ones—exactly the posterior-guided trade-off in bandits and decision theory.
MH Forecast Snapshot (Dispensight)
Bayesian regression with MH sampling on a transformed target (avg_retention_rate). Forecast horizon: 72 steps. Below is a human-learning reading of the coefficients.
- Strong state-dependence: lagged retention (transformed) is the dominant positive driver → persistence of prior state.
- Volume vs. quality tension: transactions (log1p_tx) tilt negative on retention, while top-performer sales lean positive → focused service quality helps keep people.
- Throughput cost: avg_sales_per_hour_log shows a negative effect on retention (speed can erode experience).
The sampler explores posterior uncertainty; coefficient intervals propagate into credible forecast bands.
Human-learning parallel: yesterday’s state (habit/expectation) strongly anchors today (large prior), while high-pace operations can trade off against perceived quality—reducing the probability of returning. Expert attention (top performer) counteracts this by raising perceived value.
Practical Takeaways
- Shape priors deliberately: set default scripts for staff (greeting, cross-sell rules) to bias outcomes toward retention.
- Control throughput: avoid over-optimizing speed; protect quality touch-points driving return probability.
- Exploit uncertainty: A/B exploratory tweaks (hours, bundles) and update beliefs with posterior intervals, not point hopes.
- Make uncertainty visible: show bands in dashboards so decisions respect risk, not just mean predictions.
Glossary
- Prior: belief before seeing the day’s data.
- Likelihood: how probable the data are under a hypothesis.
- Posterior: updated belief after combining prior + likelihood.
- MH (Metropolis–Hastings): Markov-chain sampler used to explore the posterior.
References & Assets
- MH forecast PDF: conjugate-mh.pdf