Cognitive Psychology: Domain Guide

The stage model is a story about bottlenecks. Sensory memory is enormous but lasts under a second; short-term memory lasts tens of seconds and holds a handful of chunks; long-term memory has no practical limit but is hard to search. Almost every classic memory finding is a consequence of one of those constraints — chunking works because the limit is in chunks rather than items, the recency effect exists because the last items are still in the short-term store, and the misinformation effect happens because retrieval reconstructs rather than replays.

Classical and operant conditioning are distinguished by what the organism controls. In classical conditioning the outcome arrives regardless of what the animal does, and the learning is about prediction: this signal means that event is coming. In operant conditioning the outcome depends on the behaviour, and the learning is about control: doing this produces that. When an item is ambiguous, ask whether the consequence would have arrived anyway. If yes, it is classical.

Perception is a hypothesis, not a recording. The visual system receives a flat, ambiguous, incomplete image and returns a stable three-dimensional scene, which it can only do by adding assumptions — about lighting, about continuity, about what objects usually look like. Illusions are not malfunctions; they are those assumptions being applied to displays built to violate them. Reading top-down processing as 'the system supplying what the input leaves out' makes the whole perception section coherent instead of a list of effects.

Memory begins with three stores. Sensory memory holds a rich but rapidly fading trace — visual traces last a fraction of a second, auditory traces a few seconds — long enough for attention to select from. Short-term memory holds roughly seven chunks for something under half a minute unless it is rehearsed, and chunking is what lets an apparently long string fit inside that limit. Long-term memory is effectively unlimited in capacity and duration, and it divides into explicit memory, retrievable with awareness and split into episodic and semantic, and implicit memory, which shows up as procedural skill, priming, and conditioned responses. Working memory is the modern refinement of the short-term store: not a passive buffer but an active workspace with separate verbal and visuospatial subsystems under an executive that allocates attention between them.

What gets remembered depends on how it was processed. Material encoded for meaning is retained better than material encoded for sound or appearance; material related to oneself better still. Practice distributed across sessions beats the same total practice massed into one, and retrieval practice — testing yourself — beats rereading. Retrieval is also cue-dependent: the encoding specificity principle says a cue helps to the degree it was present at encoding, which is why context and internal state at the time of learning can act as cues. The serial position curve puts these mechanisms in one picture: early items are recalled well because rehearsal moved them into long-term memory, and late items are recalled well because they are still in the short-term store. A filled delay before recall removes the recency effect while leaving primacy intact; speeding up presentation reduces rehearsal and therefore damages primacy while leaving recency intact.

Forgetting has distinct causes and the test wants them distinguished. Encoding failure means the information was never stored. Interference means competing memories obstruct access — proactive when older learning obstructs newer, retroactive when newer learning obstructs older. Retrieval failure means the trace exists but the cue does not reach it, the tip-of-the-tongue state being the everyday case. Ebbinghaus's forgetting curve describes the shape: loss is steepest immediately after learning and then flattens, so the first review matters far more than the fifth. Separately, memory is reconstructive, which produces errors of a specific kind — the misinformation effect, in which post-event information alters the report of the original event, and source monitoring errors, in which the content is remembered but its origin is not.

Classical conditioning is the learning of predictive relationships. An unconditioned stimulus already elicits an unconditioned response; pairing a neutral stimulus with it turns the neutral stimulus into a conditioned stimulus that elicits a conditioned response on its own. Acquisition is fastest when the conditioned stimulus reliably precedes and predicts the unconditioned stimulus — mere co-occurrence is weaker than genuine predictiveness. Presenting the conditioned stimulus alone produces extinction, and a rest interval can produce spontaneous recovery. Generalisation extends the response to similar stimuli and discrimination narrows it again. Not all pairings are equally learnable: taste aversions form after a single pairing and across delays of hours, a preparedness that fits the biology of poisoning and violates the usual timing rules.

Operant conditioning is the learning of consequences. The four cells of the reinforcement square are generated by two independent questions: does the behaviour become more or less likely, and is a stimulus added or removed? Positive reinforcement adds something desirable, negative reinforcement removes something aversive, and both increase behaviour — which is why calling negative reinforcement a punishment is the single most common error in the domain. Positive punishment adds something aversive and negative punishment removes something desirable, and both decrease behaviour. Reinforcement schedules then govern the pattern: fixed-ratio schedules produce bursts with post-reinforcement pauses, variable-ratio schedules produce the highest and steadiest response rates, fixed-interval schedules produce the scalloped pattern of responding that accelerates as the deadline nears, and variable-interval schedules produce slow steady responding. Partial reinforcement makes behaviour markedly more resistant to extinction than continuous reinforcement. Shaping reinforces successive approximations to build a behaviour that never occurs spontaneously.

A third mechanism needs no direct experience at all. In observational learning a behaviour is acquired by watching a model, and whether it is performed depends on the consequences the model was seen to receive. The sequence usually taught is attention to the model, retention of what was observed, capacity to reproduce it, and motivation to do so — which explains the standard finding that learning and performance can be separated, as they are in latent learning, where an animal exploring a maze without reward shows it learned the layout as soon as reward is introduced. The overjustification effect is the corresponding warning about reinforcement: rewarding an activity someone already enjoys can reduce the intrinsic motivation to do it once the reward stops.

Attention is the gate in front of all of this. Dichotic listening studies show that unattended input is largely lost, but not entirely — hearing one's own name in the unattended channel rules out a filter that discards everything before meaning is extracted, which is why an early complete filter was replaced by accounts in which unattended input is attenuated rather than blocked. Automatic processes run without capacity cost and are hard to suppress, which is what the Stroop effect demonstrates: reading is so automatic that it interferes with naming the ink colour. Failures of attention are equally examinable — inattentional blindness for an unexpected object in an attended scene, and change blindness for a large change across an interruption.

Perception organises what attention lets through. The Gestalt principles describe the grouping rules: figure and ground, proximity, similarity, continuity, closure, and common fate. Depth comes from binocular cues, retinal disparity and convergence, which work only within a few metres, and from monocular cues — interposition, relative size, linear perspective, texture gradient, aerial perspective, and motion parallax — which work at any distance and are all that a picture can supply. Constancies keep size, shape, and colour stable as the retinal image changes, and perceptual set describes the top-down influence of expectation on what is seen at all.

Language and higher cognition close the domain. Language is layered as phonemes, morphemes, syntax, semantics, and pragmatics, and it develops on a robust timetable: cooing in the first months, babbling around six months, first words near a year, two-word telegraphic speech in the second year. Overregularisation — a child saying 'goed' after previously saying 'went' — is prized evidence that children extract rules rather than imitate strings. The nativist argument holds that the input is too impoverished to explain the speed of acquisition without an innate predisposition; the behaviourist alternative appealed to reinforcement and imitation, and the debate between them is one of the standard set pieces of the cognitive revolution. In reasoning and decision making, algorithms guarantee a solution but cost effort while heuristics are fast and fallible: availability substitutes ease of recall for frequency, representativeness substitutes resemblance for probability and so ignores base rates, and anchoring lets an initial value contaminate an estimate. Functional fixedness and mental set are the corresponding obstacles in problem solving — seeing an object only in its usual role, and persisting with a strategy that used to work.