## Definition **Learning as predictive-error minimisation** is the computational framing of human learning in which the brain is modelled as a hierarchical generative system that continuously predicts its incoming sensory and cognitive inputs. When a prediction fails — i.e., when reality diverges from the brain's internal model — the resulting prediction error propagates upward through the hierarchy and is used as the update signal to revise the model. Learning, in this view, is the iterative process of reducing prediction errors by adjusting the internal model. ## Core Mechanics The brain does not passively record experience; it actively generates hypotheses about the world and tests them. At every level of the cortical hierarchy: 1. **Top-down predictions** are sent from higher areas to lower areas, representing the brain's current best guess. 2. **Bottom-up error signals** travel from lower to higher areas, encoding only the discrepancy between prediction and input. 3. **Model update** occurs when errors exceed a precision-weighted threshold, modifying the generative model to reduce future errors. This architecture — sometimes called the predictive coding or predictive processing framework (Rao & Ballard, 1999; Friston, 2005) — unifies perception, action, and learning under a single computational principle. ## Why Errors Are the Engine of Learning A corollary is that *confirming* experiences carry little update information — the brain already predicted them. Surprising, challenging, or incorrect outcomes generate large prediction errors and therefore drive stronger learning. This explains several empirically robust phenomena: - **Desirable difficulties.** Tasks that are effortful (spaced retrieval, interleaved practice, generation of answers before study) produce more prediction errors per unit time than easy tasks, yielding steeper learning curves. - **The testing effect.** Testing forces the learner to generate predictions (retrieved answers) that can be compared against feedback, generating errors even when the learner believes they know the material. - **Conceptual change difficulty.** Entrenched prior knowledge suppresses prediction errors from incompatible new information, making the model resistant to revision. Overcoming it requires explicitly surfacing the contradiction (see source text discussion of seasonal-change misconceptions). ## Relation to the Four Pillars In the Four Pillars framework (see [[The Four Pillars of Learning]]), the pillar of *error feedback* is the direct practical expression of this principle: the system learns primarily from mismatches, not from confirmations. Attention governs prediction precision (how strongly signals are weighted); active engagement generates the predictions whose errors drive updating; consolidation stabilises the revised model. ## Mathematical Sketch Under a Bayesian formulation, the brain maintains a posterior $P(\text{model} \mid \text{data})$ and updates it via Bayes' theorem: $P(M \mid D) \propto P(D \mid M) \cdot P(M)$ The prediction error is the log-likelihood ratio $\log P(D \mid M_{\text{new}}) / P(D \mid M_{\text{old}})$. Minimising free energy (the variational bound on surprise) drives the system toward models with lower prediction error. This formulation connects learning science to Bayesian brain theories (Friston's active inference framework). ## Related - [[The Four Pillars of Learning]] - [[Brain Plasticity]] - [[Active Engagement and the Testing Effect]] - [[Consolidation and Sleep]] ## Sources - [[How We Learn (Dehaene 2020)]]