Error-driven learning is the principle that learning occurs in proportion to the discrepancy between what was expected and what actually happened — the prediction error. When outcomes match predictions, there is nothing new to learn. When outcomes violate predictions, the learning system adjusts its representations to reduce the error. This principle, first formalized in the Rescorla-Wagner model for classical conditioning, has become a unifying concept spanning associative learning, motor adaptation, language acquisition, and reinforcement learning.
From Pavlov to Backpropagation
The history of error-driven learning spans multiple disciplines. In animal learning, the Rescorla-Wagner model (1972) formalized prediction error as the driver of conditioning. In engineering, the least mean squares (LMS) algorithm (Widrow and Hoff, 1960) used error signals to adjust filter weights. In neural networks, the back-propagation algorithm distributes error signals backward through layers to adjust connection weights. In reinforcement learning, temporal difference algorithms use prediction errors to update value estimates. Despite different formalisms, all share the core principle: learning from errors.
Neural Implementation
The discovery that dopamine neurons encode reward prediction errors (Schultz, Dayan, and Montague, 1997) provided a biological implementation of error-driven learning. When rewards exceed expectations, dopamine neurons fire above baseline; when rewards meet expectations, they fire at baseline; when expected rewards are omitted, they pause. This signal, broadcast widely through the brain, is thought to drive learning in the basal ganglia (for action selection) and cortex (for value representation and decision-making).
δ = prediction error
r = immediate reward
V(s') = estimated value of next state
V(s) = estimated value of current state
γ = discount factor
A direct implication of error-driven learning is that surprising events produce the most learning. This connects to attention: unexpected events capture attention (orienting response), which may serve to maximize the processing of surprising, informative stimuli. The Pearce-Hall model of conditioning explicitly links attention to prediction error: stimuli that are poor predictors (high prediction error) receive more attention and more processing on subsequent trials. This attention-learning interaction creates an adaptive system that focuses resources where they are most needed.