Main

First described for the perception of phonemes in human speech6, categorical perception was later shown to function in the perception of auditory signals in other animals7,8,9. With regard to colour, animals may discriminate among colours but nevertheless treat them as similar10,11,12, and colour categorization13 may influence decision-making thresholds14. Thus, animals can categorize colours, forming discrete groups of similar yet discriminable variants across the visible spectrum. However, the hallmark of categorical perception15,16—increased discrimination of variants between categories relative to variants from within—has not been demonstrated for natural variation in colour-based signals.

Carotenoid-based colouration is commonly used in visual signalling across many taxa, although its function in assessment signalling is best described in mate choice in birds and fish17,18. Individuals vary in their ability to acquire19 and metabolize20 carotenoids; therefore, variation in carotenoid-based colouration has been linked to variation in the quality of the signaller. Carotenoid-based beak colouration in male zebra finches ranges from light orange to dark red21, beak redness correlates positively with variation in cell-mediated immunity21,22, and females show a mating preference for males with red versus orange beaks23. Previous studies18 have tested how receivers respond to both ends of a carotenoid-based colour continuum, but whether they perceive variation continuously (responding differently to any detectable differences in colour) or exhibit categorical perception is unknown.

To create stimuli that vary continuously along a spectrum from red to orange, we selected eight Munsell colours (Pantone) previously used to describe the colour of zebra finch beaks24,25,26. We modelled chromatic distance (ΔS) using the receptor noise-limited model of colour discrimination27 (Extended Data Tables 1, 2). To ensure that the selected colours are approximately equidistant from one another when accounting for zebra finch spectral sensitivity28 and ambient light (Extended Data Figs. 1, 2 and Supplementary Information), we visualized ΔS in a chromaticity space29 in which the Euclidean distance between stimuli plotted in an x–y plane equals chromatic distance for a trichromatic viewer (Fig. 1).

Fig. 1: Munsell colours used to create stimuli.
figure 1

Colours were approximately equally spaced in chromaticity space and were closer to their nearest neighbour than to any other colour. Dots show mean chromaticity coordinates for each colour; ellipses show one standard deviation in the X1 and X2 dimensions (n = 4 measurements per colour; Supplementary Information); numbers between dots show chromatic distance (ΔS) between colours (mean ± s.d.). Ellipse colour corresponds to relevant Munsell colour (exact colours in the figure may vary). Inset, foraging grid of an example trial.

Source Data.

Full size image

We used a food-reward protocol to test for categorization (which is sometimes referred to as ‘labelling’6) and discrimination of the eight stimuli spanning the orange–red colour spectrum. We created discs of Munsell paper comprising two semi-circular halves of either the same or different colours (hereafter ‘solid’ and ‘bicolour’, respectively). Once females had been trained to flip over these discs to access a food reward, we trained them to flip bicolour discs first, before any solid discs; in essence, birds learned to recognize ‘bicolour’ versus ‘solid’ discs, rather than particular colour combinations. In experimental trials, we presented females with a foraging grid that had twelve wells, six of which were covered with discs: two solid discs each for each of the two colours, and two bicolour discs comprising the same colours as the solid discs (Fig. 1 inset). Birds passed a trial if they flipped both bicolour discs before any solid discs, indicating that they perceived the two colours on the bicolour disc as different.

We first performed categorization experiments to establish the location of potential perceptual boundaries. We tested females using bicolour discs that included colour 1 in combination with all other colours (that is, 1|2, 1|3, 1|4 and so on) and, separately, colour 8 in combination with all other colours (that is, 8|7, 8|6, 8|5 and so on). We determined the proportion of trials passed for each comparison. For both the 1|X and 8|X comparisons (where X is any other colour), pass frequency increased when the chromatic distance increased between colours 1 or 8 and the comparison colour X. The greatest difference occurred, however, when comparing the pass frequencies for 1|5 and 1|6, and likewise between 8|5 and 8|6, suggesting that there is a potential boundary between colours 5 and 6 (Fig. 2). A linear mixed model (Table 1) demonstrated that comparing colour 1 or 8 with the colour immediately preceding versus immediately following the putative boundary resulted in the same change in pass frequency as would moving 10.5 ΔS units, approximately equal to three colour steps within a category (mean ΔS between colour steps = 3.6).

Fig. 2: Categorization trials suggested a boundary between colours 5 and 6.
figure 2

The boundary is indicated by vertical lines on the x axis. Females were 31% more likely to pass 1|6 than 1|5 trials and 34% more likely to pass 5|8 than 6|8 trials. n = 26 birds in three independent cohorts. Box plots depict median (horizontal line inside box), 25th and 75th percentiles (box), 25/75th percentiles ±1.5× interquartile range (whiskers), and outliers (circles). Horizontal grey line indicates the expected pass frequency if birds flip discs by chance.

Source Data.

Full size image
Table 1 Mixed model results for categorization and greyscale discrimination data
Full size table

We next determined whether discrimination of colour differences crossing the putative 5–6 category boundary was increased compared to discrimination of equal colour differences that did not cross the boundary. We presented females with colour pairs that did or did not cross the hypothesized boundary, and that were two colour spaces apart (that is, 1|3, 2|4, 3|5 and so on; Fig. 3a). The pass frequency for comparisons that crossed the 5–6 boundary was 26 ± 6 percentage points higher (mean ± s.d.) than comparisons that did not (paired t-test, t25 = 9.26, P < 0.0001; Fig. 3b). We found similar results when stimuli were one or three colour spaces apart (Extended Data Figs. 3, 4) and when we combined all categorization and discrimination data into a single linear mixed model (Extended Data Table 3).

Fig. 3: Discrimination of stimuli that are two colour spaces apart increased across the 5–6 boundary.
figure 3

a, b, Mean pass frequency across all individuals (a) and for each individual (b) was greater for comparisons that crossed the boundary versus those that did not. a, Green boxes in the grey-shaded area are comparisons that cross the boundary. Sample sizes (number of birds, across three independent cohorts) are shown within each box. Michelson contrast is shown in parentheses. Boxes, whiskers, circles and horizontal grey line are as described in Fig. 2.

Source Data

Full size image

Our colour stimuli varied in brightness because real zebra finch beaks of different colour are also not of equal brightness. Several lines of evidence support the conjecture that brightness alone does not completely explain our results. First, we built a linear mixed model comparable to the one presented in Table 1, but which—in addition to chromatic distance (∆S)—included Michelson contrast (a measure of brightness ratios) between colour pairs to explain pass frequency instead of the binary variable indicating the 5–6 boundary. This model performed substantially worse (ΔAkaike information criterion = 13) than the model that included the binary parameter of crossing the boundary. Additionally, our data show that several discrimination comparisons had either similar Michelson contrasts but different pass frequencies (4|6 compared to 6|8 (Fig. 3) and 5|6 compared to 6|7 (Extended Data Fig. 3)), or equivalent pass frequencies with different Michelson contrasts (4|6 compared to 5|7 (Fig. 3) and 3|6 compared to 4|7 and 5|8 (Extended Data Fig. 4)). Lastly, we performed one-apart discrimination experiments (n = 18 birds) using shades of grey (Extended Data Fig. 5 and Supplementary Information) selected to match the red–orange colours in zebra finch double cone28 quantum catch (Extended Data Table 1), an estimate of perceived brightness in passerines30. In these greyscale experiments, pass frequency was significantly predicted by Michelson contrast: birds demonstrated increased discrimination for both 5|6 and 6|7 compared to greyscale pairs with a lower contrast (Table 1 and Extended Data Fig. 5). Together, these pieces of evidence suggest that brightness may contribute to category formation but cannot alone explain the perceptual categories that we observed.

Our results demonstrate categorical perception of colour associated with an assessment signal. We found that a category boundary influences the perception of two colour stimuli as similar or different, and results in differential discrimination between stimuli depending on their location relative to the category boundary. Discrimination improved with increasing chromatic distance between colours (that is, variants within a category were not perceived as identical), but increased most sharply across the category boundary (that is, variants from across the boundary were perceived as most different).

We did not explicitly test whether categorical perception arises at the level of the photoreceptor or higher, such as in the brain. However, a wavelength discrimination function derived from electroretinographic data from the pigeon Columba livia31, a reasonable proxy for the zebra finch, suggests that photoreceptor sensitivity alone probably does not explain categorical perception (Extended Data Fig. 6). We also did not directly test female preference for males in relation to the categories found here. Nonetheless, these findings have important implications for the understanding of colour perception and encourage further work exploring whether and how categorical perception interacts with selection on signal form and function, particularly in the context of assessment (Extended Data Fig. 7).

Methods

The goal of this study was to test whether female zebra finches perceive colours along a red-to-orange spectrum in a continuous or categorical fashion. We selected eight colours from red to orange that correspond to male zebra finch beak colours and that are roughly equidistant from one another in a chromaticity space based on zebra finch spectral sensitivity. These eight colours were made into disc stimuli that were either solid in colour (that is, made of two halves of the same colour) or bicolour (that is, one colour on one half of the disc and another colour on the other half). Disc stimuli were used in a food-reward protocol. Female zebra finches were first trained to flip over discs covering wells in a foraging grid to find a food reward (millet seeds, as used previously32), using discs made from the colours at the endpoints of our red–orange continuum (colours 1 and 8). Initially, all wells (both solid and bicolour) were baited with millet seeds, to reward the birds any time they successfully flipped a disc. After being trained on this task, females were further trained to flip only bicolour discs; we used the same stimuli made of colours 1 and 8, but this time baited only wells that were covered with bicolour discs. Once females passed six out of seven consecutive trials under this protocol, indicating they had learned to search for food only under bicoloured discs, we then conducted trials using different combinations of our eight selected colours. If females perceived two colours as distinct, they would flip the bicoloured discs first to gain a food reward; if they did not perceive two colours as distinct, they would not preferentially flip the bicoloured discs first. For details, see ‘Behavioural trials’.

Birds used in this study

All birds in this study were sexually mature female zebra finches (age range: 3–50 months at start of experimental testing) from a colony maintained by R. Mooney at Duke University (IACUC A258-14-10). After transfer from the colony, birds were housed singly in cages (46 × 23 × 23 cm3, Prevue Pet) outfitted with two wooden perches and a cuttlebone. Lighting was controlled during trials (see ‘Behavioural trials’) and food was removed 5 h before trials to ensure that birds would be motivated to perform the task. Outside of trial times, birds were kept on a 15 h:9 h light:dark cycle (consistent with the light cycle of the birds’ home colony), with overhead lighting provided by fluorescent bulbs (Ecolux with Starcoat SP 35/41, colour temperature 3,500–4,100 K, General Electric) with ballast (Hi-Lume 3D/Eco-10, Lutron Electronics) operating at 50–60 Hz. Birds were given zebra finch food (Kaytee Forti-Diet Pro Health Finch diet) and water ad libitum. Rooms were maintained at 25–27 °C. Testing was done under Duke University IACUC protocol A004-17-01.

Selecting stimulus colours

Previous work has shown that the range of beak colouration in zebra finch males can be represented by red and orange shades in the Munsell colour system24,25,26. Munsell colours are defined by three parameters: hue, value (brightness) and chroma (saturation). Previous work on zebra finches identified a large set of Munsell colours that, by the human eye, approximate the colours of zebra finch beaks. In particular, this set of Munsell colours consists of colours with hues 6.25R–3.75YR, values 3.5–6 and chromas 10–1424,26. Notably, the values used in these previous studies spanned a range of hues from yellow–orange to red, and specifically did not use Munsell colours of the same brightness because real male beaks of different hues are also of different brightness.

These Munsell colours24,25,26 are based on beak colours across two different colony populations of zebra finches and capture most of the variation within those populations, although the authors of the studies note that the beaks of occasional individuals were found to be outside that range. Because the goal of our study was to examine how female zebra finches perceive a range of colours that are similar to the range spanned by the colours of male zebra finch beaks, we used this previous work as a starting point to choose 40 Munsell colour samples from within the set outlined above. Reflectance spectra from each colour sample were measured using an integrating sphere with a built-in tungsten-halogen light source (ISP-REF; Ocean Optics). All measurements were taken with reference to a Spectralon 99% white reflectance standard (Labsphere).

For each of the Munsell colours, we calculated relative photon catches (that is, how many photons are absorbed by a given cone type in response to a visual scene33) for zebra finch short-, medium- and long-wave cones. Photon catches were calculated over the wavelengths 400−700 nm, using zebra finch spectral sensitivity curves28, an ambient light spectrum and the reflectance spectrum of each colour. Thus, photon catch Q for receptor type r in response to colour c was calculated using:

$${Q}_{r,c}(\lambda )\propto \mathop{\mathop{\int }\limits^{700}}\limits_{400}{S}_{r}(\lambda )\hspace{-0.5pt}\ast \hspace{-3pt}{R}_{c}(\lambda )\hspace{-0.5pt}\ast \hspace{-3pt}I(\lambda ){\rm{d}}\lambda $$

where Sr is the sensitivity of receptor type r, Rc is the reflectance of colour c, λ denotes the wavelength, and I is the irradiance of the illuminant. We use proportionality throughout, because we did not require an absolute quantum catch, and the constant factors remain the same across different receptor types. As for our measure of ambient light, we used a standard tungsten bulb illuminance spectrum (CIE Illuminant A, colour temperature 2,856 K; spectrum in Supplementary Information), which is very similar to the spectrum of ambient light provided by the tungsten bulbs in our experiment (see below, Extended Data Fig. 1 and Supplementary Information). We used sensitivity data from the zebra finch28 to calculate photon catches for the short-, medium- and long-wave and double cones (Extended Data Table 1). We did not calculate photon catches for the ultraviolet cone because (1) male zebra finch beaks reflect minimal ultraviolet light34, (2) female zebra finches do not require ultraviolet radiation to discriminate and rank red–orange colouration35, (3) our own measurements confirmed that Munsell paper does not reflect strongly in the ultraviolet portion of the spectrum, and (4) the reflected ultraviolet radiance from our Munsell chip stimuli, under experimental lighting conditions, was essentially zero (see Extended Data Fig. 2).

In tests of categorical perception, stimuli should be roughly equidistant from one another in perceptual space. No metric can fully describe the perceptual space of human colour vision, let alone that of a zebra finch. Therefore, to calculate the chromatic distance between colour stimuli, we used the receptor noise-limited (RNL) model of colour discrimination27, which uses photon catches calculated from the spectral sensitivity of a relevant visual system to calculate ΔS, a measure of chromatic distance between two colours (equivalent to just-noticeable differences, JNDs). We visualized ΔS using a perceptually uniform, two-dimensional space based on both hue and saturation/chroma, in which the Euclidean distance between two colours is equivalent to the RNL model-derived chromatic distance (ΔS) between colours29. Plotting chromatic distances in this two-dimensional space is only relevant for trichromatic vision, which in our case was appropriate given that we did not incorporate quantum catches from ultraviolet cones (see above). The RNL model27 states that colour discriminability is primarily limited by photoreceptor noise; thus a Euclidean distance of 1.0 in the chromaticity space is approximately equal to one standard deviation of receptor noise, or one JND. Given that this method uses photon catches from single cones (short-, medium- and long-wave), we performed a separate analysis to examine perceived brightness based on double cones (see below).

In summary, the eight colours that we used (see Supplementary Information for reflectance spectra) were based on previously published comparisons with male zebra finch beaks and are approximately equally spaced in a chromaticity space based on zebra finch spectral sensitivity. Throughout, we refer to these colours as 1 (the darkest, red end of the range) through 8 (the brightest, orange end of the range).

Sensitivity analyses

To ensure that the eight colours that we selected are not equidistant in chromaticity space simply owing to the exact parameters of the zebra finch spectral sensitivities and the tungsten lighting conditions that we used, we additionally plotted the selected colours in chromaticity space using different conditions (all spectra and sensitivity curves are provided in the Supplementary Information), including: the zebra finch ultraviolet-light-sensitive (UVS) cone type retina28 under both (1) tungsten and (2) daylight (D65) illuminants; (3) the zebra finch UVS cone type retina using the spectrum of halogen light present in the experimental rooms; (4) the zebra finch UVS cone type retina after applying von Kries adaptation based on a neutral grey background36 to account for colour constancy mechanisms; (5) another UVS cone-type retina visual system (the starling Sturnus vulgaris37); (6) the average UVS cone-type retina (data from previously published supplementary online material33); and (7) the average violet-light-sensitive (VS) cone-type retina (data from previously published supplementary online material33), which is the other primary type of retina found in birds (Extended Data Table 2). Overall, chromatic distances between selected Munsell colours remained relatively consistent even when changing the spectral sensitivity and lighting conditions, as well as when accounting for colour constancy, with the distance between 5 and 6 never being the largest jump. Thus, chromatic distance alone cannot explain the category boundary that we observed.

Behavioural trials

Room set-up

Birds were housed in individual cages, with up to 12 birds in each room. Before trials, we placed opaque barriers between adjacent cages so that birds could not see their closest neighbours perform any tasks. Individuals were able to see other birds across the room, although they could not see the task that birds were performing—this set-up prevented birds from seeing how other birds were passing the task while ensuring that they could see other birds, which was beneficial given that zebra finches are highly social. During trials, the room’s overhead lights were turned off and birds were allowed approximately 10 min to acclimatize to the experimental lighting. During experimental trials, each cage was illuminated from above by a halogen lamp (colour temperature 2,900 K, model number H&PC-61361, Philips Lighting) approximately 81 cm from the foraging grid (spectra in Supplementary Information). The light passed through vellum paper hung 8 cm from the light source to ensure that each cage had even and diffuse illumination (Extended Data Fig. 1). All trials were recorded on video (Logitech Webcam Pro 9000, Logitech). On trial days, food was removed from the cages at 09:00 and trials began at 14:00, to ensure that birds would be motivated to attempt the task.

Disc design

Our experimental stimuli were discs 2.5 cm in diameter, made of two semi-circles of Munsell colour sheets, glued with their straight edges together to create a full circle. The discs were covered with a clear epoxy cover. A clear vinyl disc (1.3 cm diameter, 0.3 cm high) was attached to the bottom of the disc to ensure it fit precisely into the wells of the foraging grid. We created two types of discs: solid, in which both semi-circular halves were the same colour; and bicolour, in which the two semi-circular halves were made of different colours. Bicolour discs were named for the colours of their two halves—that is, a disc made up of half colour 1 (the far red) and half colour 8 (the far orange) would be referred to as 1|8.

The experimental foraging grid consisted of two grey plastic blocks (13.5 × 9 cm2 and 2.5 cm high) positioned adjacent to one another. Each block contained six identical wells (1.3 cm diameter, 0.8 cm deep). Birds first learned to search for food beneath the discs in five stages. In the first stage, we placed millet seeds in four randomly selected wells, with no discs present. In the second stage, four discs (two bicolour, and one each of the two solid colours that comprised the bicolour discs halves) were on the grid, adjacent to the baited wells. In the third stage, discs were placed halfway covering the baited wells. In stage four, discs were placed tipped into the baited wells, so as to hide the seed but with discs fitted only loosely into the well. In the fifth and final stage, discs were placed to completely cover the baited well, so a bird could only access the seed by flipping the disc off of the well using its beak. For each of the training tasks, success was defined as obtaining the seeds from any of the baited wells within a 20-min period. A subject had to pass three consecutive trials of each stage to progress to the next stage.

Once the birds had learned the basic task of searching for food under the discs, they were trained to associate only bicolour discs with a food reward, a stage that we call ‘bicolour association’. Zebra finches were trained in bicolour association using a total of six discs: two 1|8 bicolour discs (which were baited), and two solid discs each of colour 1 and 8 (which were not baited). Importantly, our behavioural data show that birds learned to recognize bicolour versus solid discs, rather than particular colour combinations of bicolour discs, as shown by their ability to extrapolate the colour-based task to a greyscale task (described below). Trials lasted for 2 min, during which time birds were allowed to flip discs, and we recorded the order of the first two discs that they flipped. Following the 2-min observation period, the grid and any remaining millet were left in for up to 20 min. To pass a bicolour association trial, birds had to flip over both bicolour (1|8) discs before either solid disc was flipped within the 2-min observation period. Birds had to pass six out of seven consecutive trials before we determined that they had learned the task, after which they progressed to data collection. If an individual failed more than one training trial out of seven, they were given additional training trials until the pass criterion was reached or until we determined the bird was unlikely to learn the task. In total, 26 out of 30 birds (87%) that reached the bicolour association trials passed and went on to data collection trials.

Data collection trials involved the same grid and 12-well set up as in training. Six wells were covered with discs (two bicolour, four solid), and only the two bicolour discs were baited. We used the ‘sample’ function in R38 to create a set of randomized locations for each disc in each trial, up to 6 trials for each colour combination. Discs were placed in haphazard orientations (that is, the direction of the line bisecting the middle was not consistent from one trial to the next). During data collection, birds were allowed up to 2 min to flip discs, after which time the grid and any remaining seed were removed. We recorded both the order in which the first two discs were flipped, as well as the latency to flipping the first disc. Observers were not blind to the tasks; however, up to three of five (usually, one or two) observers collected data on a given day and no single observer consistently collected data for a given trial (for example, 5|6, 1|6). Results were consistent across the five observers who collected data during the experiment.

We conducted 3–5 trials per bird per day. Overall, we collected data from 26 birds (trained in three independent cohorts of n = 10, n = 6 and n = 10) over the course of 7,015 experimental trials. In 10.1% of trials (n = 712), birds flipped either no discs or one disc and then stopped flipping discs. In these cases, we did not count these trials as a failure, but rather removed the trial from the dataset. Most commonly, birds did not flip any discs or only flipped one disc if other birds in the room suddenly went quiet or if they were startled by the behaviour of another bird. Removing these incomplete trials from the dataset ensured that we did not bias our data towards an increased failure rate. At the end of each trial day, we performed a motivation check on each bird by returning the bird’s food dish to the cage and observing for up to 10 min to ensure that they would eat seeds. Birds typically began eating within 30 s of having their food returned, indicating that individuals still had high feeding motivation even at the end of our trials.

Categorization experiments

Categorization experiments tested whether birds responded in a continuous or discontinuous manner to colour variation, identifying the location of boundaries if they were present. As in all trials, we presented birds with two bicolour discs and two of each solid colour disc (six discs total spread randomly across 12 possible wells on the grid). Each day, a bird was given one refresher trial (1|8 bicolour and 1 and 8 solid) and five experimental trials of a single colour combination. The next day, the bird would receive a refresher trial and five trials on a different colour combination, and so on. Colour combinations used in categorization trials included the combination of colour one and all other colours (that is, 1|2, 1|3, 1|4…) as well as the combination of colour eight and all other colours (that is, 8|7, 8|6, 8|5…). Additionally, as a control for the possibility that birds used olfactory cues to locate seeds under certain discs, we performed 1|1 and 8|8 trials to check that, in the absence of any variation in visual information, birds would not perform better than chance at flipping discs placed on baited well. Birds were shown colour combinations in an order that alternated relatively distant colour combinations (those that were five to seven steps apart) with relatively closer colour combinations (those that were one to four steps apart). Each bird was given two experimental days (10 total trials) with each colour combination, and one experimental day (5 total trials) with each of the 1|1 and 8|8 comparisons. The second experimental day for each combination occurred only after the first experimental day for all other combinations had occurred (that is, experimental days for a specific colour combination did not immediately repeat).

Preliminary data analysis identified a putative boundary between colours 5 and 6. To analyse these data, we calculated the proportion of trials that each bird passed for each comparison (15 total comparisons per bird). Using only data from categorization trials, we built a linear mixed model (throughout, linear mixed models were calculated using R package lme439) that included pass rate as its response variable, Euclidian chromaticity distance (∆S) between the two colours being compared and whether the comparison crossed the putative 5–6 boundary as fixed effects, and bird ID as a random effect. The model included random slopes in addition to random intercepts, to account for variation among birds in the strength of the effect of crossing the 5–6 boundary.

We visually inspected the residuals of this model using a quantile–quantile plot, a histogram and a scatterplot of residuals versus predicted values to confirm that the residuals of the model were approximately normally distributed around zero and that they were homoscedastic.

Discrimination experiments

A second requirement for demonstrating categorical perception is that subjects show increased discrimination between stimuli that cross a category boundary compared to equally spaced stimuli that do not cross a boundary. Thus, for discrimination trials, bicoloured discs were made of colour pairs that were one space apart (that is, 1|2, 2|3, 3|4…; Extended Data Fig. 3), two spaces apart (that is, 1|3, 2|4, 3|5…; Fig. 3) or three spaces apart (that is, 1|4, 2|5, 3|6…; Extended Data Fig. 4). The experimental set-up and criteria for passing were the same as for categorization trials. For each bird, mean pass rates were calculated for all discrimination trials that did not cross the putative boundary and, separately, for those that did. We then calculated the difference between these means for each bird and used a two-sided t-test to determine whether the mean difference in pass rate was significantly different from zero.

Combining categorization and discrimination data in a single statistical model

To present our data in the most easily interpreted format, we present separate analyses for categorization and discrimination trials in the main text. To bolster our conclusions, maximize statistical power and contain our analysis in a single model, we built a linear mixed model that includes all discrimination and categorization data. As reported in the main text, this model contains random effects of bird ID and includes crossing the 5–6 boundary as a random slope. As in the main text, significance tests were performed using ANOVA comparisons. The results of the single model can be found in Extended Data Table 3.

Greyscale discrimination experiments

The set of eight Munsell colours that we used in this experiment were selected primarily based on their colour: in particular, the colour of the Munsell colours aligned with those of actual beaks and were equally spaced in a chromaticity space that describes the hue and saturation of colours and from which brightness has been factored out. However, brightness is also an important part of how colour is perceived. To examine the effects that brightness has on structuring zebra finch perception of colours, we performed a second experiment in which we used shades of grey that matched our eight Munsell colours in brightness, as perceived by zebra finches.

To select appropriate shades of grey, we used an integrating sphere to measure reflectance spectra from a set of 72 grey paint swatches (Behr brand, Behr Process). We then calculated the quantum catch of the zebra finch double cone28 for each of our eight Munsell colours and for each of the grey paint swatches; the double cone is believed to be the primary way in which birds encode brightness information30. By finding shades of grey that matched each Munsell colour in double cone quantum catch, we selected a set of eight grey shades that were equivalent in bird-perceived brightness to the colour stimuli that we had been using.

Using this set of eight shades of grey, we created experimental discs for the 1|8-grey and each one-apart combination of grey shades. We followed the same experimental procedure as above to examine the birds’ discrimination ability when hue information had been removed (Extended Data Fig. 5).

Wavelength discrimination function

To investigate whether the categorical perception that we observed could be solely due to the wavelength discrimination function (WDF) of avian photoreceptors, we examined a WDF that was derived using electroretinography in the pigeon C. livia31. We know of no experimentally or behaviourally derived WDF for zebra finches. The pigeon is a reasonable substitute, however, given that the spectral sensitivity peaks for its medium- and long-wavelength sensitive cones (505 and 565 nm, respectively40) are very similar to those of the zebra finch (507 and 565 nm, respectively28).

We plotted electroretinography-derived WDF data31 alongside the spectral sensitivity peaks of the zebra finches (Extended Data Fig. 6). The units of the WDF are arbitrary units, in keeping with the original publication, but show the general pattern of wavelength discrimination. One complication is that it is not possible to know precisely where in the visible spectrum the avian perception of ‘orange’ and ‘red’ would occur. However, in humans, orange and red both occur above 580 nm, in the region in which stimulation is primarily of the long-wavelength cone and secondarily the medium-wavelength cone. The WDF showed a discriminability peak at approximately 605 nm, followed by a relatively smooth decrease in discriminability in the region of the spectrum in which avian viewers probably see orange and red (Extended Data Fig. 6). The shape of these curves suggests that photoreceptor sensitivity on its own probably does not explain categorical perception. From this analysis alone, however, we cannot rule out that hue discrimination based on spectral sensitivity curves contributes to categorical perception. Ultimately, whether categorical perception arises at the level of the photoreceptor or retina, or is a higher-order process, does not affect our findings or interpretation, but suggests avenues for future research.

Data reporting

No statistical methods were used to predetermine sample size. The experiments were not randomized and the investigators were not blinded to allocation during experiments and outcome assessment.

Reporting summary

Further information on experimental design is available in the Nature Research Reporting Summary linked to this paper.

Data availability

The datasets generated and analysed during the current study are available in the Duke Data Repository: https://doi.org/10.7924/r4rx96r99.