Associative learning

Associative learning is learning that one event predicts another, such as a tone predicting food, or an action predicting a reward. Most modern theories of it rest on prediction error, the difference between what was expected and what happened: learning occurs when outcomes are surprising (Holland & Schiffino, 2016). These theories split into two families. In the Rescorla–Wagner model, prediction error determines how effective the outcome is at driving learning; in the Pearce–Hall model, it determines how much attention a cue receives, and so how readily the cue is learned about (Holland & Schiffino, 2016). Evidence from behaviour and from the brain suggests that both mechanisms operate (Holland & Schiffino, 2016; Roesch et al., 2012). The main dispute about humans is whether associative learning is the automatic formation of links at all, or a form of reasoning about propositions (Mitchell et al., 2009).

History

Not covered by the research: classical and operant conditioning, and the development of the Rescorla–Wagner model (1972) and its successors.

Core concepts

Learning as part of memory

Textbooks of cognitive psychology do not give learning its own chapter; they treat it within memory (Eysenck & Keane, 2020; Sternberg, 2008; Pilegard, 2024; Gobet et al., 2011). Reviews of memory do the same: the types of long-term memory are themselves types of learning, divided into declarative (conscious, dependent on the medial temporal lobe) and nondeclarative or implicit (Reber, 2013). Implicit memory, a change in performance from earlier experience without conscious recollection of it, has priming and skill learning as its main forms (Schacter et al., 1993). Reber proposes that implicit learning is general plasticity of the cortex rather than a separate memory system (Reber, 2013).

Prediction error

In the Rescorla–Wagner model, the prediction error sets how effective the outcome, the reinforcer, is (Holland & Schiffino, 2016). A well-known consequence is blocking: if one cue already predicts an outcome, a second cue added alongside it learns little, because there is no error left to drive learning. Todd and Holmes reviewed blocking in two forms of higher-order conditioning, sensory preconditioning and second-order conditioning, where cues are linked to other cues rather than directly to rewards. They conclude that a common error term of the Rescorla–Wagner kind also governs these forms of learning, which had been an open question (Todd & Holmes, 2022).

Attention to cues

In the Pearce–Hall model, prediction error sets the associability of a cue, roughly how much attention it gets and how readily it is learned about (Holland & Schiffino, 2016). Holland and Schiffino report evidence of changes in associability induced by prediction error, which they take to show that the behavioural models must be modified to include this mechanism and that Rescorla–Wagner alone is insufficient (Holland & Schiffino, 2016).

Questions, controversies, and new developments

One mechanism or two?

Neural evidence points to both mechanisms. Signals consistent with the Rescorla–Wagner error, carried by dopamine neurons, and signals consistent with Pearce–Hall attention coexist in the brain; they are dissociable but possibly linked, so the two classes of model appear complementary rather than competing (Roesch et al., 2012).

Links or propositions?

Mitchell, De Houwer and Lovibond argue that human associative learning is not the automatic formation of links between mental representations. Instead, people form propositions, beliefs such as “the tone means food is coming”, by reasoning, in cooperation with unconscious memory retrieval and perception (Mitchell et al., 2009). This is an advocated position that challenges the link-based view on which error-driven models were built; it ties associative learning to reasoning.

Replication record

Not covered by the research: replication of blocking and other classic conditioning effects in humans.

Broader connections

Connectionist networks, like associative models, store what they learn in the strengths of connections rather than as symbols (Garson & Buckner, 2019). Attention and memory form a cycle in learning: memory establishes what is relevant, and attention decides what is encoded (Turner & Sloutsky, 2024).

Further reading

• Holland and Schiffino (2016): prediction error and the two families of models.
• Mitchell et al. (2009): the propositional account of human associative learning.
• Reber (2013): learning within the memory systems.