Psychological Discrimination

Abstract

Psychological discrimination is a form of perception: the capacity to tell two stimuli apart, quantified as the smallest difference that can be reliably detected. Classical psychophysics expressed this resolving power as the difference threshold, or just-noticeable difference, and Weber found it to be a near-constant fraction of the baseline stimulus — the regularity Fechner built into the first measurement scale for sensation. Signal-detection theory later separated true sensitivity from the observer's decision criterion, replacing a fixed threshold with a continuous measure, d′, read from hit and false-alarm rates. Modern work fits the psychometric function with maximum-likelihood methods, places trials efficiently with adaptive procedures, and ties the behavioural threshold to the accumulation of noisy neural evidence. Three interactive demonstrations let the reader vary a Weber fraction, trace a psychometric function, and separate sensitivity from criterion on a signal-detection axis.

Keywords: discrimination, difference threshold, weber's law, signal detection, psychometric function

What Psychological Discrimination Is

Psychological discrimination is the ability to detect a difference between two stimuli, or between a stimulus and its absence. It is the resolving power of a perceptual system: not whether a signal can be detected at all, but how finely two signals that are nearly alike can be told apart. A listener asked which of two tones is higher, an observer asked whether a grating has tilted, a taster asked which cup is sweeter — each performs a discrimination, and the experimenter's question is always the same: how large must the physical difference be before the observer responds correctly more often than chance allows.

That question has a quantitative answer, and supplying it was the founding achievement of psychophysics. The difference threshold, or just-noticeable difference (JND), is the smallest change in a stimulus that produces a detectable change in sensation. It is not a fixed quantity but a statistical one: because both the stimulus and the observer are noisy, the same small difference is sometimes caught and sometimes missed, so the threshold is defined as the difference detected on some agreed proportion of trials. Measuring it reduces a question about private experience to a count of correct responses, and it is this reduction — sensation rendered as a number — that made the psychology of perception an experimental science.

Discrimination in this perceptual sense should not be confused with its social sense. Here the term is the one MeSH files under Discrimination, Psychological: the differential perception and response to the properties of stimuli, the basic capacity that lets an organism treat two things as different. It is a component of nearly every perceptual task, and the methods built to measure it — the threshold, the psychometric function, the signal-detection measures — are among the most widely reused tools in all of experimental psychology (Gescheider, 1997).

Types of Psychological Discrimination

MeSH places psychological discrimination under perception and records one narrower descriptor beneath it. The classification is an indexing scheme for the biomedical literature, not a theory of perception, so the single subtype below is better read as the one branch of discrimination research large enough to have earned its own index term than as an exhaustive taxonomy of the ways organisms tell stimuli apart.

Weber's Law and the Difference Threshold

The first regularity discovered about discrimination is also its most durable. Early in the nineteenth century Ernst Heinrich Weber observed that the just-noticeable difference is not a constant amount but a constant proportion of the standard against which it is judged. A weight of 100 grams must be increased by about 2 grams to feel heavier; a weight of 200 grams must be increased by about 4. The ratio of the JND to the baseline stimulus — the Weber fraction — stays roughly fixed as the baseline grows. Formally, ΔI / I = k, where I is the stimulus intensity, ΔI the just-noticeable increment, and k the Weber fraction characteristic of the continuum.

Gustav Theodor Fechner saw in Weber's fraction the lever he needed to measure sensation itself. In Elemente der Psychophysik (Fechner, 1860/1966) he reasoned that if every JND is a subjectively equal step, then cumulating JNDs builds a scale of sensation magnitude, and because each step corresponds to a constant proportional change in the stimulus, the resulting scale is logarithmic: sensation grows as the logarithm of intensity (Fechner's law). The derivation is now known to be an idealisation — Weber's law itself fails at very low and very high intensities, and the logarithmic law was later challenged by power-law accounts — but the strategy of chaining discrimination thresholds into a measurement scale founded the entire field.

Weber's law is an approximation with instructive exceptions. In intensity discrimination the Weber fraction does not stay perfectly constant but shrinks slowly as level rises, a departure Jesteadt, Wier, and Green (1977) documented across audible frequencies and named the “near miss to Weber's law,” capturing it with a single power function. And the resolving power a continuum affords is bound up with how perceived magnitude grows along it: Teghtsoonian (1971) showed that the Weber fraction, the exponent of Stevens's power law, and the dynamic range of a sense are mutually constrained, so discrimination and the scaling of sensation are two views of one underlying quantity.

From Thresholds to Signal Detection

The classical threshold carried a hidden assumption: that there exists a sensory cut-point below which nothing is perceived and above which it is. By the mid-twentieth century that assumption was failing under its own data, because an observer's apparent threshold moved whenever the payoffs or the probabilities changed, without any change in the stimulus. Swets (1961) put the question in its title — Is there a sensory threshold? — and answered that once the observer's willingness to say “yes” is separated from sensory sensitivity, no fixed threshold is evident; what looks like a threshold is partly a decision.

Signal-detection theory, imported into psychophysics by Green and Swets (1966), made that separation exact. It models each presentation as a sample from one of two overlapping probability distributions — noise alone, or signal-plus-noise — along a single internal decision axis. The observer sets a criterion on that axis and responds “yes” whenever the sample exceeds it. Two quantities then fall out of the hit and false-alarm rates: d′, the distance between the two distributions in standard-deviation units, which measures genuine sensitivity, and the criterion, which measures the observer's response bias. A discrimination that classical methods would summarise with one shifting threshold is thereby split into a sensory part that the stimulus controls and a decision part that payoffs and expectations control.

The practical payoff is a recipe for computing sensitivity from response counts in almost any discrimination design — yes-no, forced-choice, same-different, rating — set out in full by Macmillan and Creelman (2005). Because d′ is, in principle, independent of criterion, it lets an experimenter compare sensitivity across observers and conditions who differ in caution, which a raw percent-correct cannot do. Detection theory did not discard the threshold so much as explain it: the threshold is the stimulus level that yields a criterion level of d′, now a derived landmark on a continuous sensitivity scale rather than a sensory floor.

The Psychometric Function

Whatever the measure, discrimination performance is summarised by the psychometric function: the curve relating the probability of a correct (or “yes”) response to the physical magnitude of the stimulus difference. It rises from chance at zero difference to near-certainty at large differences, typically as a sigmoid when the difference is plotted on a logarithmic axis. Two numbers capture it: the threshold, the stimulus value at some criterion performance level (say 75% correct), which locates the curve along the stimulus axis, and the slope, which indexes how abruptly performance improves as the difference grows.

Estimating those two numbers from a finite set of noisy trials is a statistical problem with real traps, and Wichmann and Hill (2001) gave the modern maximum-likelihood recipe for it. Their central refinement was to fit a lapse rate alongside threshold and slope — a small probability that the observer responds wrongly for reasons unrelated to the stimulus, such as a blink or a lapse of attention — because a handful of stray errors at easy stimulus levels can otherwise bias the threshold and badly distort the slope. Fitting the lapse rate, rather than assuming perfect performance at the top of the curve, protects the discrimination estimate from exactly the kind of error it is most prone to.

The fit itself can fail silently. Prins (2019) showed that a maximum-likelihood fit of the psychometric function can converge on a local maximum of the likelihood, or on a boundary where no proper maximum exists, returning a plausible-looking threshold and slope that are in fact artefacts of the fitting procedure rather than estimates of the observer's sensitivity. The lesson is methodological but consequential: a reported threshold is only as trustworthy as the fit that produced it, and diagnosing false convergence is now part of responsible psychophysical practice.

Measuring Discrimination Efficiently

Measuring a threshold by presenting a fixed grid of stimulus levels wastes most trials, because levels far from the threshold carry almost no information about where it lies. Adaptive procedures solve this by letting each trial depend on the responses so far, concentrating presentations near the threshold being estimated. Leek (2001) surveyed the family — transformed up-down staircases that track a fixed point on the psychometric function, parameter-estimation routines such as PEST, and maximum-likelihood trackers — each a rule for choosing the next stimulus from the running history of hits and misses.

The Bayesian members of the family are now dominant. Watson and Pelli (1983) introduced QUEST, which maintains a probability distribution over the threshold's location and places each trial at its current best estimate, exploiting the fact that the psychometric function has a roughly invariant shape on log intensity so that only its position need be found. Watson (2017) generalised the idea to QUEST+, which estimates several parameters at once — threshold, slope, and lapse together — and extends to multidimensional stimulus and response spaces, letting a single adaptive framework handle experiments that older one-parameter staircases could not. The contemporary practical literature, Kingdom and Prins (2016) foremost among it, packages these methods into toolboxes that make efficient, well-fitted threshold estimation the routine starting point of a discrimination experiment rather than an expert's art.

Neural Evidence Accumulation

Discrimination is, finally, a decision, and the last two decades have located its machinery in the brain. Gold and Shadlen (2007) synthesised a large body of single-neuron and behavioural work into a now-standard account: in a discrimination task the brain accumulates noisy sensory evidence over time toward a decision bound, and the rate of that accumulation is set by the quality of the sensory signal. The framework ties three levels together. The firing of sensory neurons supplies the momentary evidence; association-cortex neurons integrate it, their activity climbing toward a threshold; and the psychophysical difference threshold measured at the behavioural level reflects how much evidence must accumulate before the bound is reached. The same signal-detection quantities that describe the observer — sensitivity and criterion — reappear as properties of the accumulation, the criterion now a bound on integrated neural activity. Discrimination thereby becomes a bridge concept, the one place where a psychophysical measurement, a statistical decision model, and a recording from a single neuron describe the same event.

Figure

Interactive Demonstrations

Weber’s law: the just-noticeable difference grows with the baseline

With a constant Weber fraction the JND is a straight line through the origin: at k = 0.10 a 100-unit standard needs a 10-unit increment, a 400-unit standard needs 40. Turning on the near-miss bends the line gently downward at high levels, the small systematic departure Jesteadt and colleagues measured for tone intensity.

The first demonstration makes Weber's law tangible. Set a baseline stimulus and a Weber fraction, and the demo reports the just-noticeable difference ΔI = k · I and shows how the increment required for detection grows in step with the baseline while the ratio stays fixed. A toggle introduces the “near miss” of Jesteadt, Wier, and Green (1977), letting the fraction shrink slowly with level so the reader can see how a real sensory continuum departs from the idealised constant.

The psychometric function: reading a threshold off the curve

Slope sets how sharply performance climbs from chance to ceiling; the lapse rate pulls the upper asymptote below 1 so that stray errors on easy trials are not misread as a shallower slope. Raising the criterion level moves the red read-off point rightward — a stricter definition of “threshold” demands a stronger stimulus.

The second demonstration implements Figure 1 as a manipulable curve. Move the threshold and slope and watch the psychometric function slide and steepen; set a criterion performance level and read the threshold off the stimulus axis, exactly as a fitted function yields it (Wichmann and Hill, 2001). A lapse-rate control pulls the top of the curve below perfect performance, illustrating why fitting the lapse protects the threshold estimate.

Signal detection: sensitivity (d′) versus decision criterion

At d′ = 1.98 with a slightly liberal criterion (c = −0.49) the observer hits about 93% of signals at the cost of a 31% false-alarm rate — the worked example in the text. Sliding the criterion traces the whole ROC without changing d′: bias moves the red point along a fixed curve, while raising d′ bows the curve further from the diagonal.

The third demonstration separates sensitivity from criterion on the signal-detection axis. Drag the two distributions apart to raise d′, slide the criterion to trade hits against false alarms, and watch the point trace out a receiver-operating-characteristic curve. It shows directly why two observers with identical sensitivity can produce different percent-correct scores when their criteria differ (Green and Swets, 1966; Macmillan and Creelman, 2005).

Measures of Discrimination

Worked Example

Consider an intensity-discrimination task with a Weber fraction of k = 0.10. At a baseline of I = 100 units, the just-noticeable difference is ΔI = k · I = 0.10 × 100 = 10 units; at I = 400 units it is ΔI = 0.10 × 400 = 40 units. The increment quadruples with the baseline, yet the ratio — the resolving power — is unchanged, which is the content of Weber's law.

Now take the same discrimination through signal-detection theory. Suppose that over many trials the observer scores a hit rate of H = 0.93 on signal-present trials and a false-alarm rate of F = 0.31 on signal-absent trials. Sensitivity is d′ = z(H) − z(F), where z is the inverse normal. Here z(0.93) ≈ 1.48 and z(0.31) ≈ −0.50, so d′ = 1.48 − (−0.50) = 1.98 — the two distributions are just under two standard deviations apart. The criterion is c = −½[z(H) + z(F)] = −½[1.48 + (−0.50)] = −0.49, a negative value marking a liberal bias: the observer leans toward saying “yes.”

The example makes the central lesson concrete. A second observer who was more cautious might report H = 0.69 and F = 0.07, giving z(0.69) ≈ 0.50 and z(0.07) ≈ −1.48, so d′ = 0.50 − (−1.48) = 1.98 — identical sensitivity — but a criterion of c = −½[0.50 + (−1.48)] = +0.49, a conservative bias. Their raw hit rates differ sharply, yet their discrimination is the same; only by separating sensitivity from criterion does that equality become visible. The SignalDetectionDemo recomputes these same values as the distributions and criterion are moved, so the prose and the demonstration agree.

Discussion

Discrimination is the thread that runs through the whole history of perceptual measurement. It began as a single number, the just-noticeable difference, and a single law, Weber's, from which Fechner spun the first scale of sensation. It was reconceived in the twentieth century as a decision problem, with signal-detection theory separating what the senses deliver from what the observer chooses to say, and the difference threshold demoted from a sensory floor to a landmark on a continuous sensitivity scale. It is now measured with maximum-likelihood fits and Bayesian adaptive methods that extract a well-characterised psychometric function from a few hundred efficient trials, and it is increasingly understood mechanistically, as the accumulation of noisy evidence by populations of neurons toward a bound.

What unifies these layers is that each is a way of asking how reliably a system can tell two states of the world apart, and each gives a commensurable answer. A Weber fraction, a d′, a psychometric slope, and an evidence-accumulation rate are not rival measures but views of one quantity at different grain. That coherence is why discrimination remains a load-bearing concept across sensory psychology: a result expressed in any one of these currencies can be converted, at least in principle, into the others, and an improvement in resolving power — whether from practice, attention, or a change of task — can be chased from behaviour down to its neural cause.

Current Directions

The most active current front is perceptual learning: the finding that discrimination thresholds fall with practice, often dramatically, and the effort to say exactly what changes in the observer when they do. Dosher and Lu (2017) reviewed the computational models that attribute the improvement not to a sharpening of the sensory signal itself but to reweighting — the decision stage learning to read the informative sensory channels more heavily and to discount the noisy ones — and showed how observer models that separate external-noise filtering from internal-noise reduction can decide between competing accounts of a given training effect. Lu and Dosher (2022) surveyed the field's open questions: why learning is often specific to the trained stimulus, retina location, or task and when it instead transfers; how training schedules and stimulus variability govern that specificity; and how the same principles can be turned to clinical use in visual rehabilitation, where structured discrimination training recovers function in amblyopia and low vision. The through-line to the rest of the article is direct: perceptual learning is discrimination measured twice, before and after practice, with the tools of threshold estimation and signal detection now used to localise where in the processing chain the resolving power improved.

Common Misconceptions

Glossary

Key Researchers

Barbara Anne Dosher

. Distinguished Professor of Cognitive Sciences at the University of California, Irvine and a member of the National Academy of Sciences, whose work with Zhong-Lin Lu built the observer models that partition discrimination improvement into template retuning and internal-noise reduction.

Gustav Theodor Fechner

(1801–1887). Founder of psychophysics; in Elemente der Psychophysik (1860) he turned Weber's fraction into the first measurement scale for sensation, deriving the logarithmic law from the cumulation of just-noticeable differences.

David M. Green

(1932–2022). Co-author with John A. Swets of the foundational text that imported signal-detection theory into psychophysics, giving sensory psychology the separation of sensitivity from decision criterion.

Zhong-Lin Lu

. Professor of Neural Science and Psychology at New York University and NYU Shanghai, known for the perceptual-template model of external-noise discrimination and computational accounts of perceptual learning.

Michael N. Shadlen

. Professor of Neuroscience at Columbia University and a Howard Hughes Medical Institute investigator, whose work established the accumulation of sensory evidence to a bound as the neural basis of perceptual discrimination.

Ernst Heinrich Weber

(1795–1878). Discovered that the just-noticeable difference is a constant proportion of the baseline stimulus — Weber's law — the empirical foundation of difference-threshold measurement.

Frequently Asked Questions

What is psychological discrimination?

In its perceptual sense it is the capacity to tell two stimuli apart — to detect a difference between them. It is measured as the difference threshold, or just-noticeable difference: the smallest physical difference an observer can reliably detect. This is the sense MeSH files under Discrimination, Psychological, distinct from the social meaning of the word.

What is a just-noticeable difference?

The just-noticeable difference (JND), or difference threshold, is the smallest change in a stimulus that produces a detectable change in sensation. Because perception is noisy, it is defined statistically — as the difference detected on some agreed proportion of trials, often 75% correct — rather than as a sharp all-or-none boundary.

What is Weber's law?

Weber's law states that the just-noticeable difference is a constant proportion of the baseline stimulus: ΔI / I = k. A heavier standard weight needs a proportionally larger increment to feel heavier. The ratio k, the Weber fraction, characterises the resolving power of a sensory continuum and stays roughly constant across its working range.

How does signal-detection theory change the picture?

It separates two things the classical threshold confounded: sensitivity, the true capacity to tell signal from noise, and the criterion, the observer's willingness to respond “yes.” From hit and false-alarm rates it computes d′ for sensitivity and a separate criterion measure for bias, so that observers who differ only in caution are no longer mistaken for differing in discrimination.

What is a psychometric function?

It is the curve relating the probability of a correct response to the size of the stimulus difference, rising from chance to near-certainty as the difference grows. It is summarised by a threshold (its location on the stimulus axis) and a slope (its steepness), both estimated by fitting the curve to the data.

Why use adaptive procedures to measure discrimination?

Because trials far from the threshold carry little information about where it lies. Adaptive methods such as staircases and the Bayesian QUEST and QUEST+ procedures choose each trial from the running history of responses, concentrating presentations near the threshold and so estimating it accurately from far fewer trials than a fixed grid would need.

Is the difference threshold a real sensory limit or a decision?

Both, and signal-detection theory is what tells them apart. The threshold has a genuine sensory component — the separation of the underlying distributions — but its apparent location also depends on where the observer sets a decision criterion, which payoffs and expectations can move without any change in the senses.

Can discrimination improve with practice?

Yes. Perceptual learning lowers discrimination thresholds with training, sometimes markedly. Computational work attributes the gain largely to the decision stage learning to weight the informative sensory channels more heavily, and the same principles are being applied clinically in visual rehabilitation for conditions such as amblyopia.

References

Dosher, B., & Lu, Z.-L. (2017). Visual perceptual learning and models. Annual Review of Vision Science, 3, 343–363. https://doi.org/10.1146/annurev-vision-102016-061249

Fechner, G. T. (1966). Elements of psychophysics (H. E. Adler, Trans.; D. H. Howes & E. G. Boring, Eds.). Holt, Rinehart and Winston. (Original work published 1860). 📖

Gescheider, G. A. (1997). Psychophysics: The fundamentals (3rd ed.). Lawrence Erlbaum Associates. 📖

Gold, J. I., & Shadlen, M. N. (2007). The neural basis of decision making. Annual Review of Neuroscience, 30, 535–574. https://doi.org/10.1146/annurev.neuro.29.051605.113038

Green, D. M., & Swets, J. A. (1966). Signal detection theory and psychophysics. Wiley. 📖

Jesteadt, W., Wier, C. C., & Green, D. M. (1977). Intensity discrimination as a function of frequency and sensation level. Journal of the Acoustical Society of America, 61(1), 169–177. https://doi.org/10.1121/1.381278

Kingdom, F. A. A., & Prins, N. (2016). Psychophysics: A practical introduction (2nd ed.). Academic Press. 📖

Leek, M. R. (2001). Adaptive procedures in psychophysical research. Perception & Psychophysics, 63(8), 1279–1292. https://doi.org/10.3758/BF03194543

Lu, Z.-L., & Dosher, B. A. (2022). Current directions in visual perceptual learning. Nature Reviews Psychology, 1(11), 654–668. https://doi.org/10.1038/s44159-022-00107-2

Macmillan, N. A., & Creelman, C. D. (2005). Detection theory: A user's guide (2nd ed.). Lawrence Erlbaum Associates. 📖

Prins, N. (2019). Too much model, too little data: How a maximum-likelihood fit of a psychometric function may fail, and how to detect and avoid this. Attention, Perception, & Psychophysics, 81(5), 1725–1739. https://doi.org/10.3758/s13414-019-01706-7

Swets, J. A. (1961). Is there a sensory threshold? Science, 134(3473), 168–177. https://doi.org/10.1126/science.134.3473.168

Teghtsoonian, R. (1971). On the exponents in Stevens' law and the constant in Ekman's law. Psychological Review, 78(1), 71–80. https://doi.org/10.1037/h0030300

Watson, A. B. (2017). QUEST+: A general multidimensional Bayesian adaptive psychometric method. Journal of Vision, 17(3):10, 1–27. https://doi.org/10.1167/17.3.10

Watson, A. B., & Pelli, D. G. (1983). QUEST: A Bayesian adaptive psychometric method. Perception & Psychophysics, 33(2), 113–120. https://doi.org/10.3758/BF03202828

Wichmann, F. A., & Hill, N. J. (2001). The psychometric function: I. Fitting, sampling, and goodness of fit. Perception & Psychophysics, 63(8), 1293–1313. https://doi.org/10.3758/BF03194544