erms used across the research pages, defined in plain words, each with a pointer to where it does its work. This is a seed: 96 terms now, more as rooms open. Definitions are our own wording, not quotations.

A

action potential

The brief, all-or-none electrical pulse that travels along a nerve fibre, caused by voltage-dependent sodium and potassium currents. Hodgkin and Huxley described it in equations in 1952. Also called a spike. → archive: the squid

activation patching

Copying internal activations from one run of a model into another (for instance, from a clean input into a corrupted one) and measuring how the output changes, to locate the components that matter. The answer can depend on the corruption and the metric chosen. → interpretability

algorithm aversion

Avoiding an algorithm after seeing it err, even when it outperforms the human alternative. → automation bias

anthropomorphism

Attributing human mental states, such as intentions or feelings, to non-human things. → anthropomorphism

automation bias

Relying on an automated aid in place of one's own vigilance, producing omission errors (missing what it did not flag) and commission errors (following it when it is wrong). → automation bias

B

base rate

How common something is before any specific evidence is considered. Ignoring it is a classic error. → bounded rationality  Bayes

blocking

When a cue that already predicts an outcome prevents a new cue presented alongside it from being learned about. Evidence that learning depends on surprise, not just pairing. → prediction error

boids

Simulated flocking creatures from Craig Reynolds's 1987 computer-graphics model. Each one steers by three local rules: keep a little distance from close neighbours, match their heading, and move toward their average position. Nobody tells the flock its shape; the shape comes out of the rules. → trail settings
Source: Reynolds, C. W. (1987). Flocks, herds and schools: a distributed behavioral model. ACM SIGGRAPH Computer Graphics 21(4), 25–34. doi:10.1145/37402.37406. The three rules above are our summary [paraphrase, not a quotation].
[playful] On a desktop, on a later visit, leave a research page alone for a few seconds and a few faint, slow birds drift by. Type the name of this entry, anywhere on those pages, to send them home or call them back.

Brain-Score

A public benchmark that scores artificial networks by how well their internal activity predicts recorded brain responses and behaviour. Useful, and argued over: a good score is not the same as a good theory. → brains and models compared

Brier score

The mean squared difference between probability forecasts and outcomes (1 if it happened, 0 if not). Lower is better. → calibration

C

calibration

Agreement between stated confidence and actual accuracy: of all the times you say 70 percent, about 70 percent should come true. → calibration

causal abstraction

A standard for explanations of a network: a simple, human-readable causal model counts as an account of the network if intervening on the network's internal states has the same effects as intervening on the matching variables of the simple model. It asks for more than correlation between activations and concepts. → interpretability

cell assembly

Hebb's 1949 term for a group of neurons that have become wired together by firing together, so that activating part of the group tends to activate the rest. A memory, on this view, is an assembly. → Hebb  the engram

change blindness

Failing to notice a large change in a scene when the change happens during a blink, a cut or a brief blank, so that it makes no flicker of its own. → try it  signal detection

chronoscope

A clock for very short intervals. The Hipp chronoscope, read in thousandths of a second, was the standard instrument for reaction times in nineteenth-century laboratories. → the chronoscope  Wundt

Clever Hans effect

Apparent ability that comes from unintended cues given by the questioner. Named after the horse; now also used for machine-learning shortcuts. → Clever Hans  shortcut learning

clinical versus statistical prediction

The comparison of expert judgment with a formula applied to the same information. Since Meehl's 1954 review, formulas have matched or beaten experts in most published comparisons. → algorithm aversion  Meehl

complementarity

A human–machine team outperforming both the human alone and the machine alone. → complementarity

confabulation

Producing a plausible account, sincerely, without knowing that it is unsupported. Used in neurology for some patients with memory loss, and more loosely for people explaining their own behaviour, and now for language models. → hallucination  interpretability

connectome

A complete map of the neurons in a nervous system, or a piece of one, and the synapses between them. Complete for the worm C. elegans (1986) and the adult fruit fly (2024). → connectomics

consolidation

The process by which a new memory becomes stable over time. The idea goes back to Müller and Pilzecker in 1900; it is studied at the level of synapses and of whole brain systems. → timeline: memory  memory drum

corpus callosum

The thick band of nerve fibres joining the two cerebral hemispheres. Cutting it, to stop seizures spreading, produced the split-brain patients. → Sperry

criterion (response bias)

In signal detection theory, the internal evidence level above which an observer says "yes". Payoffs and expectations move it; sensitivity stays put. → signal detection

D

d′ (d-prime)

Sensitivity in signal detection theory: the separation between the noise and signal-plus-noise distributions, in standard-deviation units. With equal variances, d′ = z(hit rate) − z(false-alarm rate). → signal detection

demand characteristics

Cues in an experiment that tell participants what is expected of them, so that they behave accordingly. → shortcut learning

E

ELIZA effect

Reading understanding into simple program output. Named after Weizenbaum's 1966 program. → Weizenbaum  talk to ELIZA

engram

The physical trace a memory leaves in the brain. Richard Semon's word (1904); Lashley spent decades failing to localise one. Now used for sets of cells whose activity during learning is needed, and can be enough, for recall. → the engram  Lashley

episodic memory

Tulving's term (1972) for memory of personally experienced events, located in time and place, as distinct from semantic memory for facts and meanings. → timeline: memory

expected calibration error

Group a model's predictions by stated confidence, and average the gap between confidence and accuracy across groups, weighted by group size. Simple and common, but sensitive to how the groups are chosen. → calibration

experimenter expectancy

An experimenter's hypothesis influencing results through subtle, usually unintended, differences in how participants are treated or how data are recorded. → Clever Hans

F

faithfulness (of an explanation)

Whether an explanation reflects the process that actually produced a model's output. It is different from plausibility, which is whether the explanation convinces people; an explanation can have either without the other. → interpretability  mental models

feedback (negative)

Control by comparing what is happening with what is wanted and acting to reduce the difference, as a steam governor does with engine speed. The core idea of cybernetics. → purpose  Wiener

file drawer problem

Studies with null results going unpublished, so the published record overstates effects. → replication  metascience

flashbulb memory

A vivid, confident memory of where one was on hearing shocking news. Studies that tested people soon after and again later found the confidence stays high while the details drift, as with ordinary memories. → timeline: memory

forgetting curve

The decline of retention with time since learning, steep at first and then slower. First measured by Ebbinghaus with savings, and replicated in 2015. → Ebbinghaus  forgetting

forking paths, garden of

Gelman and Loken's phrase for the many analyses a researcher could have run, which inflate false positives even when only one is run, if it was chosen after seeing the data. → try it  replication

G

goal misgeneralization

A trained system competently pursuing a goal other than the intended one when conditions change, even though its training signal was correct: it learned something that matched the goal in training and diverges outside it. → specification gaming  shortcut learning

good regulator theorem

Conant and Ashby's 1970 result, usually summarised as: every good regulator of a system must be a model of that system. What exactly the proof establishes is debated. → timeline: cybernetics

Goodhart's law

The observation, from economics, that a measure used as a target stops being a good measure. Campbell's law says much the same about social indicators. → specification gaming  replication

H

hallucination (in language models)

Generated text that is fluent but unsupported by, or contrary to, the source or the facts. Intrinsic hallucinations contradict the source given; extrinsic ones cannot be checked against it. → hallucination  calibration

heuristic

A rule of thumb that trades accuracy for speed or simplicity. Whether a given heuristic is an error or a good tool depends on the environment. → bounded rationality

homunculus (cortical)

The distorted figure drawn along the motor and sensory strips of the cortex, with large hands and lips and a small trunk, summarising Penfield's stimulation results. A summary of variable data, not a literal map. → the little man  Penfield

I

iconic memory

A brief, large-capacity visual store, lasting a fraction of a second after a display ends; named by Neisser after Sperling's experiments. It covers at least two things, visible persistence and informational persistence. → Sperling, 1960

imitation game

Turing's 1950 replacement for the question whether machines can think: can an interrogator, talking in writing, tell a machine from a person? Whether passing would show intelligence is still argued. → Turing, 1950  Turing

interpretability

Research that tries to say, in terms a person can check, what a trained model uses to reach its outputs. It covers looking inside (transparency) and explaining outputs after the fact (post-hoc explanation), which are not the same thing. → interpretability

J

just-noticeable difference (JND)

The smallest change in a stimulus that an observer can detect reliably, usually defined at a set percentage correct. → Weber  the line toy

L

long-term potentiation (LTP)

A lasting increase in the strength of a synapse after brief, intense stimulation, first reported in the rabbit hippocampus in 1973. The best-studied candidate for the cellular basis of learning; whether it is the memory or a condition for it is still argued. → the engram

M

Müller-Lyer illusion

A line between outward-pointing fins looks longer than an equal line between arrowheads. Published in 1889; how strong it is varies between people and groups. → try it  replication

mental model

A person's working theory of how a system behaves: what it can do, when it fails, what it will do next. Usually incomplete, and good enough when it predicts the cases that matter. → mental models  complementarity

meta-d′

A measure of metacognitive sensitivity: how well someone's confidence ratings separate their correct answers from their errors, expressed in the same units as d′ so the two can be compared. → calibration  signal detection

N

neuron doctrine

The view, argued by Cajal from the 1880s, that the nervous system is made of separate cells, neurons, which touch without fusing and pass signals in one direction. It won; it has since been qualified by electrical synapses, dendritic computation and signalling by glia. → Cajal  the nervous system

P

p-hacking

Making analysis choices after seeing the data (when to stop, what to exclude, which measure to report) until a result crosses the significance line. Often unintentional. → replication

partial report

Sperling's procedure: after a brief display, a cue says which part to report. Better performance on the cued part than whole report would predict shows more was briefly available than could be reported. → Sperling, 1960  tachistoscope

phenomenal overflow

The claim that conscious experience holds more than can be accessed and reported at once. Partial report is its main evidence; critics argue that the sense of richness is partly filled in. → Sperling, 1960

prediction error

The difference between an outcome and the outcome that was predicted. In the Rescorla–Wagner model learning is proportional to it. → prediction error

predictive processing

A family of theories in which the brain constantly predicts its own input and passes forward mainly the errors in those predictions. Broader and less settled than prediction error in learning theory. → predictive processing  prediction error

preregistration

Recording hypotheses, design and analysis plan in a time-stamped public registry before collecting data. → replication

probing classifier

A small model trained to read a property (a part of speech, say) from a network's internal activations. Success shows the information can be decoded, not that the network uses it; control tasks and causal tests are needed for that. → interpretability

psychometric function

A curve of response probability against stimulus strength, usually S-shaped. Thresholds and slopes are read from it. → the staircase

R

reaction time

The time from a signal to a response. Simple reaction time has one signal and one response; choice reaction time adds a decision and grows with the number of alternatives. → try it  timeline

readiness potential

A slow negative shift in scalp voltage before a self-initiated movement, first described in 1965. Libet compared its onset with reported intention times; an accumulator model reads it as the average of noise before a threshold crossing. → Libet, 1983

receptive field

The region of the sensory world, and the pattern within it, that makes a given neuron respond. A frog's "bug detector" and a cat's orientation-selective cell are described by their receptive fields. → archive: the cat  archive: the frog

reconsolidation

The finding that a consolidated memory, when recalled, can become unstable again and must be re-stored, so that it can be altered or weakened in that window. Strong in animal work; harder to reproduce in people. → timeline: memory

registered report

A journal article whose question and method are peer-reviewed and accepted before the data are collected, so that publication does not depend on the result. → preregistration  timeline: replication

reject option

Letting a classifier decline to decide when its confidence is too low, trading coverage for accuracy. Worked out by Chow in 1970. → where the line is

requisite variety, law of

Ashby's principle that a regulator can hold an outcome steady only if it can produce at least as much variety of response as there is variety in the disturbances it must counter. → Ashby  timeline: cybernetics

reticular theory

The older view, held by Golgi, that nerve fibres join into one continuous net through which activity spreads. Disproved as a general account by the electron microscope; partly echoed by cells joined through gap junctions. → Cajal and Golgi

reward hacking

A system finding ways to score well on its training signal that do not achieve what the signal was meant to measure. → sycophancy  shortcut learning

reward model

A model trained to predict which of two outputs a person would prefer; used as the training signal in RLHF. → sycophancy

RLHF

Reinforcement learning from human feedback: people compare model outputs, a reward model learns their preferences, and the model is tuned to score well on it. → sycophancy

ROC curve

Receiver operating characteristic: hit rate plotted against false-alarm rate as the criterion varies. The further it bows toward the top-left corner, the more sensitive the observer or classifier. → signal detection  calibration

S

saliency map

A picture that colours each part of an input by how much it is said to have mattered to a model's output. Useful, but some methods look much the same for a trained and a scrambled model, so a map has to be tested before it is trusted. → interpretability  shortcut learning

satisficing

Herbert Simon's word for choosing the first option that meets an aspiration level, instead of searching for the best. → Simon

savings (method of)

A measure of memory: how much less work it takes to relearn something than it took to learn it first. It detects memory that recall alone would miss. → Ebbinghaus  forgetting

schema

In Bartlett's sense, an active organisation of past experience that shapes what is noticed and how it is recalled. Revived in cognitive psychology for knowledge structures such as what an office usually contains. → Bartlett, 1932  Bartlett

selection task

Wason's four-card problem: which cards must be turned over to test an if-then rule? Most people miss the card that could break it. → try it  bounded rationality

semantic entropy

A measure of a language model's uncertainty: sample several answers, group them by meaning rather than wording, and compute the entropy over the groups. High values flag answers likely to be confabulated. → hallucination

serial reproduction

One person's recall of a story or picture becomes the next person's material, as in the children's game of telephone. Bartlett used it to study how material changes as it passes between people. → Bartlett, 1932

shortcut learning

Learning a decision rule that works on the training and test data for the wrong reason, and fails when conditions change. → shortcut learning

sparse autoencoder

A network trained to rebuild a model's activations from a larger set of features, only a few of which are active at once. Used to pull apart features stored in superposition; whether the features it finds are the model's own units is debated. → interpretability

specification gaming

Behaviour that satisfies the literal statement of an objective without achieving what its designers intended, like a racing program that scores points by circling instead of finishing. → specification gaming  purpose

staircase

An adaptive procedure: make the stimulus weaker after correct responses and stronger after errors, so testing concentrates near the threshold. Different up/down rules converge on different points of the psychometric function. → run one  signal detection
In the tour, the dark room keeps a staircase log of its own: the same up-and-down walk toward a faint thing, written up the way a lab would. That log is part of the tour's story, not a real study.

statistical power

The probability that a study detects an effect of a given size if it is real. Low power misses real effects and exaggerates the ones it finds. → replication

Stroop effect

Slower naming of a word's ink colour when the word itself names a different colour. Reading happens whether you want it or not. → try it  timeline

superposition (in neural networks)

The finding, in small models, that a network can represent more features than it has dimensions by giving them overlapping directions, as long as each feature is rarely active. One consequence: a single unit may stand for several unrelated things. → interpretability

sycophancy

A model shifting its answers toward a user's apparent views or wishes at the expense of accuracy. → sycophancy  the polite machine

synapse

The junction where one neuron passes a signal to another. Most are chemical: a transmitter crosses a narrow gap. Some are electrical: the cells are joined by channels and current flows straight through. Sherrington named it in 1897. → Sherrington  archive: 1906

T

tachistoscope

An instrument that shows a picture or word for a brief, controlled time. The name covers a family of devices, from falling screens to shutters to computer displays. → the tachistoscope  spot the change

teleology

Explanation by purpose or goal. Rosenblueth, Wiener and Bigelow proposed in 1943 to define purposeful behaviour as behaviour controlled by negative feedback; Taylor objected that the definition was both too broad and too narrow. → the 1943 paper  purpose

temporal-difference (TD) learning

Learning from the difference between successive predictions (plus any reward received in between), rather than waiting for the final outcome. → prediction error

threshold (limen)

The stimulus intensity at which something is detected, in the classical view a fixed line. In practice detection rises smoothly, so a threshold is defined as the intensity giving some chosen proportion of correct or "yes" responses. → signal detection  where the line is

two-alternative forced choice (2AFC)

A design in which every trial has a signal in one of two intervals or places and the observer must choose. Because they must answer, response bias matters much less than in yes/no designs. → the staircase toy uses it

two-point threshold

The smallest distance between two points touching the skin at which they feel like two, not one. Small on the fingertips and lips, large on the back. Weber measured it with a pair of compass points. → Weber's compass  Weber

Type M and Type S errors

Errors of magnitude and of sign: a statistically significant estimate that greatly exaggerates the true effect, or points the wrong way. Both are common when power is low. → try it

U

ultrastability

Ashby's term for a system with a second feedback loop that, when its essential variables go out of bounds, changes the system's own parameters until it is stable again. Built in metal as the homeostat. → the homeostat  Ashby

W

Weber fraction

The JND divided by the starting intensity. Weber found it roughly constant over a middle range for many senses: a heavier weight needs a proportionally larger change to be noticed. → Weber  Fechner  find yours

research areas  |  people  |  annotated bibliographies