Triggerfish uses chromaticity and lightness for object segregation

Humans group components of visual patterns according to their colour, and perceive colours separately from shape. This property of human visual perception is the basis behind the Ishihara test for colour deficiency, where an observer is asked to detect a pattern made up of dots of similar colour with variable lightness against a background of dots made from different colour(s) and lightness. To find out if fish use colour for object segregation in a similar manner to humans, we used stimuli inspired by the Ishihara test. Triggerfish (Rhinecanthus aculeatus) were trained to detect a cross constructed from similarly coloured dots against various backgrounds. Fish detected this cross even when it was camouflaged using either achromatic or chromatic noise, but fish relied more on chromatic cues for shape segregation. It remains unknown whether fish may switch to rely primarily on achromatic cues in scenarios where target objects have higher achromatic contrast and lower chromatic contrast. Fish were also able to generalize between stimuli of different colours, suggesting that colour and shape are processed by fish independently.


Introduction
The survival of visually adept animals depends on their ability to detect and identify prey, predators and conspecifics that are often concealed by shadows and/or camouflaged by disruptive patterns [1][2][3]. Colour vision enhances the ability of animals to detect objects and it has been suggested that colour vision originally evolved as an adaptation for object detection in conditions of changing and patchy illumination [4,5]. Different animals solve these problems in different ways depending on the   [24] (figure 1a), as well as their common occurrence in reef fish skin patterns [27]. The stimuli were mounted on an achromatic background and experiments were performed under natural illumination in blue tanks (electronic supplementary material, figure S1). The receptor quantum catches (figure 1e) were calculated using the illumination measured inside tanks (figure 1b). The RGB values of stimuli were adjusted, so that the dark yellow and the bright yellow shared similar chromaticity for fish, as did the dark blue and the bright blue (figure 1d). The double cone (M + L), M-and L-cone quantum catches of bright colours differed substantially from those of dark colours (figure 1e). On the other hand, the double cone (M + L), M-and L-cone quantum catches of bright blue and yellow were similar to each other, as were the double cone (M + L), M-and L-cone quantum catches dark of yellow and blue colours (figure 1e). Because the lightness vision in R. aculeatus is mediated by double cones and/or by the L cones [17], bright colours are predicted to be easily discriminated from dark colours on the basis of their lightness. Also, because S cones do not contribute to the R. aculeatus lightness vision [17] the bright yellow and bright blue are similar in their lightness, as are the dark yellow and dark blue colours. To quantify the difference between colours, we used the receptor noise limited model [27,28] (for details of calculations see electronic supplementary materials, contrast calculations). Calculations show that all colours used in this study can be discriminated from each other; however, there is a highly salient difference between chromatic properties of yellow and blue and achromatic properties of dark and bright colours (figure 1d,e; see electronic supplementary material, table S1). Square-shaped stimuli (5.0 × 5.0 cm) were composed of hexagonally arranged rows of coloured dots presented against a neutral grey background (figure 2). Dots that are 3 mm in diameter have previously been shown to be visible to R. aculeatus at a viewing distance of 10 cm [25,29]. Stimuli used during the training phase of the experiment were a cross shape comprising bright yellow dots on a background of dark blue or vice versa ( figure 2a,b). In testing, cross stimuli were composed of dots which either had consistent chromaticity, but inconsistent lightness (achromatic noise) (figure 2e,f ) or consistent lightness, but inconsistent chromaticity (chromatic noise) (figure 2g,h). Visual noise was generated by printing, in addition to the dark blue and bright yellow dots, the bright blue and dark yellow dots in random order. Stimuli with chromatic noise had a random arrangement of blue and yellow dots, with the cross being composed from dots having similar lightness and different chromaticity. Stimuli with achromatic noise had a random arrangement of dark and light dots, with the cross being composed of dots having similar chromaticity and different lightness. Distractor stimuli (figure 2c,d,i) were composed of randomly arranged dots with the same number of dots of a given colour as in the corresponding stimuli with a cross. For each stimulus with randomly arranged dots, we designed 25 replicates using a random number generator (Wolfram Mathematica 10.1). These were presented in a randomly shuffled order each session, to prevent the learning of non-cross related features.

Training
To test the hypothesis that fish can segregate shape on the basis of common colour, we trained R. aculeatus via operant conditioning to detect a cross composed of dots presented among a surround of different coloured dots. For the first 5 days in captivity, fish were exposed to the presence of feeding boards in the tank for an hour per day. Fish were initially encouraged to approach cross stimuli by smearing cross stimuli with squid until they approached without food present. Fish were then encouraged to peck stimuli to receive a food reward delivered from above. Group 1 (n = 10) were trained to discriminate a bright yellow cross presented against a surround of dark blue dots, from random arrays of dark blue dots and bright yellow dots (figure 2a,c). Group 2 (n = 10) were trained to discriminate a dark blue cross presented against a surround of bright yellow dots, from random arrays of bright yellow and dark blue dots (figure 2b,d). Stimuli were attached to white acrylic feeding boards (12 × 40 cm), and held 10 cm apart at the presenting tank-end (electronic supplementary material, figure S1). To prevent any sidebiases from developing, the position of positive stimuli (left or right) was changed pseudo-randomly between each trial, never being presented on one side for more than three consecutive trials. A choice counted as a single peck anywhere on a single stimulus. A correct choice for the cross-shaped stimulus was immediately rewarded with a small (1-1.5 mm) piece of squid by tweezers from above at the centre  of the tank. An incorrect choice went unrewarded and punished by immediate trial termination, with no interaction for 30 s. Stimuli were immediately removed following a choice, to prevent multiple choices being made. Depending on the motivation of a fish, between three and six trials were recorded per session. Most fish learnt the task within 8-12 days and made anywhere between 28 and 62 choices. Fish had successfully learnt the task after reaching a probability threshold of ≥70% correctness held over five consecutive sessions, with five to six trials per session (binomial test, n = 28-30, p < 0.05).

Experiment
2.4.1. Do fish predominantly use chromatic or achromatic cues when given a direct choice?
After fish had been trained to select the cross stimulus, we performed unrewarded trials where fish had to select between two cross shapes: one camouflaged with achromatic noise (figure 2e,f ), another camouflaged with chromatic noise (figure 2g,h). These trials were unrewarded to prevent fish from forming a preference for one of the crosses, as this may not be indicative of its saliency. Unrewarded trials were separated by a minimum of three rewarded trials involving training stimuli. Each fish was tested 30 times, over a period of 13-15 sessions.

Do fish predominantly use chromatic or achromatic cues when given an indirect choice?
To further investigate the ability of fish to segregate shape camouflaged by chromatic or achromatic noise, we presented fish with camouflaged crosses (rewarded) (figure 2e,f ), against a random arrangement of dots of different colours (unrewarded) (figure 2i). The proportion of dots of each colour in rewarded and unrewarded stimuli was equal. For Group 1 (originally trained using a bright yellow cross), the rewarded stimulus was initially a yellow cross camouflaged with achromatic noise (figure 2e) followed by, a bright cross camouflaged with chromatic noise (figure 2g). For Group 2 (originally trained using a dark blue cross), the rewarded stimulus was initially a blue cross camouflaged with achromatic noise (figure 2f ) and then a dark cross camouflaged with chromatic noise (figure 2h). All fish conducted a total of 30 choices per set of camouflaged stimuli over a period of five sessions, with the exception of one fish (that conducted n = 20 trials) due to an infected fin.

Can fish generalize shape over difference in colour?
To test whether fish generalize shape over difference in colour, we presented fish with a reverse coloured set of stimuli with a distracter stimulus: fish trained with a bright yellow cross presented with a dark blue cross (figure 2b) and those trained with a dark blue cross (Group 2) (figure 2b) were presented with a bright yellow cross (figure 2a).

Statistical analysis
All statistical tests were conducted using the software package R v. 3.2.2 [30]. All three tests in our experiment were analysed using generalized linear mixed models (GLMM) with a binomial distribution with log link function, from the lmer function in the lme4 package [31]. The outcome (1, correct or achromatically camouflaged stimulus; 0, incorrect or chromatically camouflaged stimulus) was entered as the dependent variable. Rewarded stimuli position (L, left; R, right) and test (chromatic camouflage or achromatic camouflage) were used as fixed factors, and fish identity was a random factor to account for fish being tested multiple times. Analysis was performed separately for fish trained to a bright yellow cross (Group 1) and for fish trained to dark blue cross (Group 2). Initially, the size of the fish (SL) was also included in the model as a covariate, but was found to be insignificant (all models: p > 0.73) and subsequently disregarded. Any fish that was found to have a side bias in a test was excluded from that choice analysis, this included one individual in the first test and two individuals in the second test (see electronic supplementary material, tables S2 and S3a,b).

Training
All, but one fish from Group 1 learnt to detect the cross shape with a minimum of 70% correct choices (figure 3a; for individual performance of fish during training see electronic supplementary material, figure S2). This indicated that fish could group dots into a shape on the basis of common chromaticity and/or lightness. The fish that did not learn the task was dropped from the experiment. There was no difference between Group 1 (yellow cross; n = 9) and Group 2 (blue cross; n = 10) in the overall performance during the last 5 session of training, i.e. in their ability to learn the task (GLMM; binomial: z = −0.47, n trials = 753, n fish = 19, p = 0.635; figure 3a).

Do fish predominantly use chromatic or achromatic cues when given a direct choice?
When presented with a chromatically camouflaged cross (similar lightness and different colour) and an achromatically camouflaged cross (similar colour and different lightness), fish from both training groups were significantly more likely to choose the chromatic cross (figure 3b, Group 1, GLMM; binomial: z = 2.87, n fish = 8, n trials = 240, p < 0.01; Group 2: z = 3.45, n fish = 10, n trials = 300, p < 0.001; for individual performance see electronic supplementary material, table S2). This suggests that fish relied more heavily on chromatic cues for object segregation.

Do fish prefer chromatic or achromatic cues when given an indirect choice?
When fish were presented with a chromatic cross and a distracter stimulus, they were significantly more likely to choose crosses than the distracter stimulus (figure 3c, chromatic cross; Group 1, GLMM; binomial: z = 5.54, n fish = 9, n trials = 260, p < 0.001; Group 2, z = 4.88, n fish = 8, n trials = 240, p < 0.001). Only fish in Group 1 were also found to be significantly more likely to choose achromatic crosses over the distracter stimulus (figure 3d, achromatic cross; Group 1, GLMM; binomial: z = 3.67, p < 0.001; Group 2, z = 0.56, p = 0.576). Additionally, fish made significantly more correct choices when choosing between chromatic crosses compared with achromatic crosses (Group 1, GLMM; binomial: z = 7.01, n trials = 520, p < 0.01; Group 2: z = 5.58, n trials = 480, p < 0.001). These results suggest that fish were able to detect crosses camouflaged by both achromatic and chromatic noise, and that it was easier for fish to segregate crosses based on chromaticity than on lightness.

Can fish generalize shape over difference in colour?
When fish were presented with stimuli of reversed colour from those presented in training, they were significantly more likely to make a correct choice for the cross-shape stimulus compared with the distracter stimulus (figure 3e, Group 1, GLMM; binomial: z = 4.80, n fish = 8, n trials = 240, p < 0.001; Group 2: z = 5.83, n fish = 10, n trials = 300, p < 0.001). The level of performance during this generalization  four pre-experimental sessions with training stimuli [mean (%) ± s.e. = 82.5 ± 2.0]. It seems most fish generalized the shape of the training stimulus over its colour, rather than relearned the novel stimulus.

Discussion
We have shown that R. aculeatus groups dots to segregate shape and generalizes shape irrespective of colour and lightness. These findings support the hypothesis of a similarity of object detection strategies among different animals and humans.
R. aculeatus was able to detect the cross shape when it was camouflaged both with chromatic and achromatic noise. However, fish were better at distinguishing the cross when it was chromatically consistent compared with crosses with chromatic variability/noise. From this, we conclude that R. aculeatus relied more heavily on chromatic cues for object segregation. Previous studies have demonstrated that a number of animals including the honeybee (Apis mellifera) [14], birds [15,32] and humans [33] mainly rely on lightness (achromatic) cues for detecting and discriminating shape and small targets, while chromatic cues are primarily used for discriminating colours of stimuli subtending large visual angles. A recent study also demonstrated that R. aculeatus chose stimuli based on achromatic cues rather than chromatic cues when viewing small stimuli [29]. Newport et al. also found that R. aculeatus learnt larger stimuli via chromaticity, rather than pattern/shape or luminance/lightness [29]; however, in their study, rewarded conspicuous stimuli were discriminated against similarly conspicuous distracter stimuli. The present study is a detection task, rather than a discrimination task, therefore different results may be expected as discrimination tasks require memory of specific objects, whereas detection tasks do not [13].
The importance of chromatic cues for object segregation can be explained as an adaptation to detection and identification of objects in conditions of spatially and temporally variable illumination. Spatial variation of illumination renders lightness unreliable and, hence, chromaticity becomes a more stable cue for segregation of object shape [4,6]. The patchiness of illumination is characteristic of forest habitat and it has been proposed that variations of lighting conditions explains the greater weight given to chromatic signals by primates [4] and other forest-dwelling species [34][35][36]. In shallow aquatic environments, wave motion produces patchy illumination, which may explain the usefulness of chromatic cues for segregation of objects for reef fish. However, whether fishes that dwell in habitats at greater depths with uniform illumination [37] rely on chromaticity for object segregation, remains unclear.
Our conclusion that chromaticity is a dominant cue for object segregation is based on comparing extreme chromatic contrast (yellow-blue) to extreme achromatic contrast (bright-dark). Both differences correspond to the range of 30-50 just noticeable differences (JNDs) (electronic supplementary material, table S1) and are probably close to the saturation of the saliency of contrasts. Generally, animals rely on a more salient cue and, therefore, in the case of unsaturated colours strong achromatic contrast is likely to be more salient than chromatic contrast. It would be interesting to investigate how the ability of fish to segregate objects depends on the relative amounts of chromatic and achromatic contrast, and how the saliency of contrast depends on chromatic and achromatic contrasts. For humans, chromaticity and lightness belong to different modalities; therefore, comparing the two on the same scale is a difficult task [38]. Human observers can make reliable pair-wise contrast matches between gratings that differ along chromatic and achromatic axes [38]. Aside from previous behavioural work on the vision of the honeybee [39] and crow [40], the comparison of the saliency of chromatic and achromatic cues of different contrasts for shape segregation has not yet been performed for humans, nor for most animals.
Similar to primates, fish generalize shape over colour, which probably helps in the recognition of objects when colour changes depend on illumination, viewing angle and distance to object [41]. The fact that fish generalize shape over colour, suggests that similar to humans, fish also process colour and shape separately. In humans, the separation of colour and shape is achieved by independent processing of different aspects of visual stimuli in the visual cortex [7].
Previous studies have shown that fish and other animals are capable of performing tasks that are thought to require complex cortical processing in humans. For example, archerfish can recognize faces [42] and various species of fish amodaly complete objects [43,44]. Fish can also be tricked by optical illusions including the Ebbinghaus illusion [45], illusory motion [46] and lightness illusion [46]. Additionally, fishes and bees have been found to perceive illusory contours [47][48][49]. The ability of animals to carry out complex visual tasks and perceive visual illusions, which we also perceive, supports the hypothesis that fish and other 'lower animals', including insects, use similar neural strategies for object detection and discrimination. However, since 'lower animals' do not have a visual cortex, the neural implementation of these 'algorithms' in fish may be more down-stream, even starting with the retina [50,51].
Natural lighting conditions have a strong influence over which visual cues are most salient to observers. R. aculeatus appears to depend more on chromatic cues for object segregation and this may be due to the presence of high achromatic noise in shallow marine habitats. However, our conclusion is derived using extreme chromatic difference between colours. Further investigation involving a range of intermediate colours with less extreme chromatic contrast and greater differences in intensity is necessary to fully understand the importance of chromatic and achromatic cues. Finally, the non-cortical structures in fish that are responsible for independently processing the visual signals of colour and shape seem to exhibit a similar neural strategy implemented by relevant cortical structures found in humans.
Ethics. The capture of live Rhinecanthus aculeatus was approved by the Great Barrier Reef Marine Parks Authority (permit no. G12/35688) and the Queensland Fisheries Department (permit no. 161624). The experiment received approval from the University of Queensland's Animal Ethics Committee (AEC approval no. SBS/111/14/ARC).