Visual discrimination of species in dogs (Canis familiaris)
- First Online:
- Cite this article as:
- Autier-Dérian, D., Deputte, B.L., Chalvet-Monfray, K. et al. Anim Cogn (2013) 16: 637. doi:10.1007/s10071-013-0600-8
- 1.4k Downloads
In most social interactions, an animal has to determine whether the other animal belongs to its own species. This perception may be visual and may involve several cognitive processes such as discrimination and categorization. Perceptual categorization is likely to be involved in species characterized by a great phenotypic diversity. As a consequence of intensive artificial selection, domestic dogs, Canis familiaris, present the largest phenotypic diversity among domestic mammals. The goal of our study was to determine whether dogs can discriminate any type of dog from other species and can group all dogs whatever their phenotypes within the same category. Nine pet dogs were successfully trained through instrumental conditioning using a clicker and food rewards to choose a rewarded image, S+, out of two images displayed on computer screens. The generalization step consisted in the presentation of a large sample of paired images of heads of dogs from different breeds and cross-breeds with those of other mammal species, included humans. A reversal phase followed the generalization step. Each of the nine subjects was able to group all the images of dogs within the same category. Thus, the dogs have the capacity of species discrimination despite their great phenotypic variability, based only on visual images of heads.
KeywordsSpecies discrimination Categorization 2D images Dogs
Social life relates to associations of individuals from the same species (Campan and Scapini 2002; Tinbergen 1953). These associations require a recognition that can be at the level of a species, a specific group, a particular social category or a particular individual (Gheusi et al. 1994). The sensory modalities involved in recognition processes vary among species. In mammals, olfaction, audition and vision are involved to a greater or lesser extent depending of the species (Porter 1987). Nevertheless, the discrimination still seems to be possible using only one sense, when the animal is deprived of others as demonstrated in sheep (Ovis aries: Porter et al. 1997; Ligout and Porter 2004; Ligout et al. 2004). So, numerous studies have demonstrated the ability of animals to discriminate conspecifics from visual cues alone. Examples include rhesus macaques (Macaca mulatta): Pascalis and Bachevalier (1998); Fujita (1987), and others macaques (Macaca fuscata, M. radiata, M. nemestrina): Fujita (1987), in sheep (Ovis aries): Kendrick et al. (1995, 1996), in heifers (Bos Taurus): Coulon et al. (2007), in dogs (Canis familiaris): Racca et al. (2010); Somppi et al. (2012), in birds (Melopsittacus undulatus): Brown and Dooling (1992), and in invertebrates, paper wasps (Polistes fuscatus): Tibbetts (2002).
In humans, faces seem to have a special informational value for individual recognition. In fact, face discrimination has been shown to be more efficient and specific compared to non-face object discrimination (Farah et al. 1998). This specific face processing has also been suggested by the results of brain studies using neuroimaging, single-cell and MEA experiments in temporal cortex (Kendrick et al. 2001a; Perrett and Mistlin 1990; Perrett et al. 1982, 1988; Pinsk et al. 2009; Tate et al. 2006; Tsao et al. 2006). Face discrimination is configurally sensitive, since the presentation of an upside-down face decreases discrimination performance. This so-called “inversion effect” has been shown in humans (Yin 1969), in chimpanzees (Parr et al. 1998) and in sheep (Kendrick et al. 1996; Peirce et al. 2000). The inversion effect seems less obvious in some others species such as dogs (Racca et al. 2010) or in macaques, where studies showed contrasting results (Bruce 1982; Parr and Heintz 2008; Perrett et al. 1988). For the investigation of face discrimination, 2D pictures seem to be appropriate stimuli (for review, see Leopold and Rhodes 2010; Tate et al. 2006).
In some species, individuals recognize more easily or are more attracted by images of individuals belonging to their own species, than those belonging to another species. This so-called “species-specific effect” has been shown in macaques (Fujita 1993; Fujita and Watanabe 1995; Pascalis and Bachevalier 1998; Dahl et al. 2009), chimpanzees (Hattori et al. 2010) and humans (Pascalis and Bachevalier 1998; Dahl et al. 2009; Hattori et al. 2010). Moreover, it has been demonstrated that below the species level, individuals showed less difficulty to discriminate individuals of their own breed (e.g., in heifers; Coulon et al. 2009) or of their own race (e.g., in humans; Malpass and Kravitz 1969; Meissner and Brigham 2001; Young et al. 2009) rather than of others. Those preferences (for the own species or own breed/race) may be linked with sex (Fujita and Watanabe 1995) or with cognitive development (Pascalis et al. 2002). Moreover, cross-fostering experiences in sheep and goat have pointed to the primacy of the species providing maternal care upon other social or interspecific interactions (Kendrick et al. 1998, 2001b).
Ultimately, these studies compared behavioral responses of subjects confronted to images of conspecifics compared to images of a limited number of species, and generally with a limited number of instances of each category. For example, in one of the most exhaustive studies in non-human primates (Dufour et al. 2006), subjects were tested with only four categories of faces (humans, own species and two species of the same genus), with 10 instances per face category. This is also the case in studies involving domestic animals. Species discrimination has been confirmed only for a few categories: in sheep, only with humans, dogs and unfamiliar sheep breeds (Kendrick et al. 1995); in heifers with sheep (5 pictures), horses (3 pictures), goat (1 picture) and dog (1 picture; Coulon et al. 2007). In dogs, the only two studies highlighting their ability to discriminate 2D images of their own species among other species used a procedure involving visual preferential looking between dogs and humans, with 24 pictures per category, humans and dogs, for the first study (Racca et al. 2010) and a similar number for the second one (Somppi et al. 2012).
Because of the small number of stimuli used, these studies could not take into account morphological species-specific diversity. In fact, there is more morphological diversity among breeds in domestic species compared to wild species (Hemmer 1990). The largest morphological variety among all animal species is found in domestic dogs, Canis familiaris (Wayne and Ostrander 2007). There are considerable variations between breeds in size, weight (from less than one kg -Chihuahua- to 100 kg—Mastiff-), color, hair length and texture (stiff or curly), form and position of ears (upright or drooping) and tail, independent profile, etc. (Denis 2007). According to the American Kennel Club and the International Cynologic Federation (FCI), 400–500 breeds of dogs are now registered, some of which are nearly identical morphologically, while others differ to a very large degree (cf. Miklósi 2007 pp 109). Diversity and changes in morphotypes are still increasing with hybridization and the proliferation of dogs associated with their short duration of gestation.
Considering this amazing morphological diversity among dogs, visual species recognition may represent a true cognitive challenge. Kerswell et al. (2010) showed that changes in morphological features could affect communication in young dogs. For example, a shorter snout seems to increase the frequencies with which some social signals are sent or elicited from other dogs; Relationships with eye cover and coat length were also found. Several studies have shown that visual cues are significant for domestic canids in interactions between conspecifics and heterospecifics (Gaunet and Deputte 2011; Hare and Tomasello 1999; Range et al. 2007a; Virányi et al. 2004, 2008). With the constraint of its large morphological diversity, companion dogs are an interesting model for studying visual species discrimination with 2D images of faces.
Behavioral responses of domestic dogs confronted to still 2D pictures have been investigated under various procedures. On the one hand, procedures have explored preferential looking with black and white pictures of dogs, humans and objects (Racca et al. 2010) or with color pictures of dogs, humans, children’s toys or alphabetic characters (Somppi et al. 2012). On the other hand, operant conditioning procedures have explored dogs’ choices when confronted with color pictures of dogs and landscapes (Range et al. 2007b) or confronted with black and white images of humans faces, blank or smiling (Nagasawa et al. 2011). A further paper (Adachi et al. 2007) has examined whether dogs have a cross-modal (acoustic and visual) representation of human individuals. These studies have shown that the particularity of a dog’s visual system does not prevent them being sensitive to still pictures. Compared with that of human beings (Kanwisher and Yovel 2006) or non-human primates (Tate et al. 2006; Tsao et al. 2006), for which an efficient system for the processing of its own species faces was demonstrated, the canine visual system could be considered inferior: in degree of binocular overlap (Sherman and Wilson 1975), in color perception (Jacobs et al. 1993), in brightness discrimination (Pretterer et al. 2004), in accommodative range and in visual acuity (Neuhaus and Regenfuss 1967; Miller 2008). In other ways, such as the capacity to be sensitive to rapid movements (Coile et al. 1989), the canine visual systems may outperform the human visual system.
In short, dogs display a very efficient visual communication system toward conspecifics and also to human beings. We assume that this communication could not work without efficient discrimination of conspecifics. Moreover, the dog is subjected to significant constraints such as a large morphologic diversity and the fact that some dog morphotypes may impair visual communication. Therefore, the present study aims at exploring whether dogs using only visual cues are able to discriminate various morphotypes of conspecifics not only from humans but also from other animal species. To address this issue, we used a great variety of pictures of faces from various dog breeds and cross-breeds, to be discriminated from a variety of pictures of faces from humans and other animal species, domestic and wild.
Experimental procedure and presentation stimuli
Type of task
Stimulus pictures in each session (12 trials/session)
A single pair of a bowl (S+) versus a black screen (S−)
A single pair of the dog D1 (S+) versus a black screen (S−)
A single pair of the dog D1 (S+) versus a blue screen (S−)
A single pair of the dog D1 (S+) versus the cow C1 (S−)
A set of 12 dogs (S+) versus 12 cows (S−)
A set of 12 dogs (S+) versus 6 cows and 6 others species (S−)
A set of 12 dogs (S+) versus 12 others species (S−)
Reversal reward contingencies
A single pair of the cow C2 (S+) versus the dog D2 (S−)
A single pair of the cow C2 (S+) versus the dog D2 (S−)
A set of 12 cows (S+) versus 12 dogs (S−)
A set of 12 others species (S+) versus 12 dogs (S−)
First training task with “S+ = a bowl of kibble” (Task 0)
Training tasks with “S+ = dog” (Task 1–3)
In the second training task, called Task 1, the pair of pictures included the picture of the dog D1 as S+ and a black screen as S− (Table 1; Fig. 4). Then in Task 2, the pair of pictures included the picture of the same dog D1 as S+ and a blue screen as S− (The same blue as the backgrounds of the animal pictures; Table 1; Fig. 4) and in Task 3, the picture of the same dog D1 as S+ and that of the cow C1 as S− (Table 1; Fig. 4).
Generalization tasks with “S+ = dog” (Task 4–6)
In the generalization tasks, we used dog pictures that were as varied as possible. Within each session, the 12 dog pictures included 3 pictures of each of the four major morphotypes of dog (wolf, hound, mastiff and greyhound types). In Task 4, pairs of dog pictures as S+ and cow pictures as S− were presented to the subjects (Table 1; Fig. 4). Then in Task 5, we introduced species other than cows in the pairs “dog/non-dog”. A session of 12 trials was then constituted with 12 dog pictures as S+ versus 12 “non-dog” pictures as S−, divided into 6 cows and 6 species different than dogs and cows (Table 1; Fig. 4). In Task 6, the diversity in “non-dog” stimuli was even larger, as sessions included 12 dog pictures as S+ versus 12 “non-dog” pictures as S− including different species other than dogs, all different from each other (Table 1; Fig. 4).
Training tasks in reversal reward contingencies with “S+ = non-dog species” (Task 7–8)
After Task 6, the stimulus-reward contingency was reversed, so that S+ was a “non-dog” picture and S− was a dog picture. As in the previous training tasks, in Task 7, the same single pair was used as stimuli, that is, the picture of the cow C2 as S+ and the blue screen as S− (Table 1; Fig. 4). Then in Task 8, the single pair used was the picture of the same cow C2 as S+ and the picture of the dog D2 as S− (Table 1; Fig. 4).
Generalization tasks in reversal reward contingencies with “S+ = non-dog species” (Task 9–10)
In Task 9, sessions included 12 different pairs of different stimuli, 12 cow pictures as S+ versus 12 dog pictures as S− (Table 1; Fig. 4). In Task 10, sessions included 12 different pairs, 12 pictures of “non-dog” species as S+, all species being different from each other, and 12 dog pictures as S− (Table 1; Fig. 4).
The apparatus was installed in a windowless room of 3 m × 5 m. The two computer screens presenting stimuli were inserted in a wooden panel (Fig. 3). The dog could see the two pictures placed at its eye level when assuming a relaxed posture with its neck in the horizontal position. The stimuli were presented at a distance of 1.9 m from the sitting area where subjects made their choice. This distance of 1.9 m was chosen in light of data on the visual acuity of dogs (Neuhaus and Regenfuss 1967) and was validated by preliminary testing. A partition of 0.9 m length was set perpendicularly to the panel between the two screens (Fig. 3a, b); when the dog was in front of one screen, it was unable to see the other screen (Fig. 3c).
Subjects were tested using a simultaneous discrimination paradigm. The experimental protocol was adapted from previous experiments in cows (Rybarczyk et al. 2001; Coulon et al. 2007, 2009, 2010), sheep (Ferreira et al. 2004) and dogs (Pretterer et al. 2004). Each trial consisted of a choice between two stimuli, one of them (S+) being rewarded while the other (S−) was not. Prior to conducting the experiments (3–5 weeks), the dogs were trained using an operant positive conditioning method with a clicker followed by a food reward. The experimenter first taught them to sit and stay motionless on the “sitting area” (Fig. 3); then to move toward one of the stimuli after hearing a verbal order (namely “image”) given by the experimenter; then to put the paw on the tablet in front of the stimulus chosen (Fig. 3c). The pair of stimuli used during this initial shaping was a bowl full of kibble (S+) versus a uniformly black screen (S−). The position (left/right) of the picture with the bowl of kibble varied randomly from trial to trial. The criterion for passing this shaping period was that the dog without any assistance immediately returned to the “sitting area” and retook the sitting position after the order “place!/here!” was given by the experimenter behind him, sitting motionless before hearing the order “image!” given by the same experimenter, and then in a delay of less than 10 s gets up to put his paw in front of the chosen stimulus.
The experimenter wore dark glasses and stood motionless, arms by his sides, 1 m behind the dog. The experimenter gave the order “place” to the dog. The experimenter looked at his feet and changed the stimuli with a remote control. He then gave the order “image” in a neutral tone. The experimenter raised his eyes only when the dog ran toward a picture and made his response. Then, the experimenter activated the clicker if the dog’s choice was correct and dropped a food reward behind him. After making an incorrect choice, the dog simply returned to the “sitting area” where he resumed his sitting position, facing the screens (Fig. 3). Thus, great care was taken to avoid visual, tactile, or acoustical cues that could inform the subject about the location of the rewarded stimulus. Before the experiment, the experimenter was trained to stay motionless by means of a video under observation by two other experimenters. The whole experiment was video recorded and then checked for possible cueing of the dogs.
All sessions were the same for the nine dog subjects. The criterion for a subject to pass from a given task to the following one was set at 10 correct trials out of 12, for two consecutive sessions. These thresholds were chosen to considerably decrease the probability of passing a session by chance. The probability of getting at least 10 successful trials out of 12 trials by chance, when the probability for one trial is 0.5, was P = 0.019. The probability of obtaining two consecutive successful sessions was 0.0192 = 0.00037. When the subject succeeded with a task, it was moved on to the next one.
The training and generalization tasks included discrimination tasks of increasing complexity (Table 1). Without this progressive complexity, the subjects lost their motivation for the experiments (personal observation in pilot studies). They laid in front of the screens showing distress (moaning, intention to leave the room, yawns, scrapings, e.g., Beerda et al. 1997).
One to four consecutive sessions of 12 trials were given in the morning, depending on the dog’s motivation. There were at least 24 h between daily blocks of sessions. At the beginning of the test, a dog was taken from the kennel and led to the experiment room after a relaxing walk. The owner was not present in the experimental room unlike other studies using dogs (Range et al. 2007b; Racca et al. 2010; Nagasawa et al. 2011; Somppi et al. 2012); the attention of the dogs was considerably reduced in all preliminary tests when the owner was present, even if he or she was hidden from the dog’s view.
Data analysis and statistical tests
For each trial, we recorded the success or the failure of the dog to choose S+. For each task, we recorded the number of sessions that each subject needed to reach the criterion. The number of sessions constituted the main variable. The data analysis was carried out using R 2.13.1 (R Development Core Team 2010).
Since the same dogs were tested repeatedly in different tasks, data were not independent, so we took into account the individual dog as a “random effect” in the analysis. The comparison of the number of sessions to reach the criterion between tasks for all subjects was analyzed by means of a generalized linear mixed model using lme4 package version 0.999375-35 for R (Bates and Maechler 2010). This model aimed at explaining the variation of the number of sessions that the dogs required for a given task. Using this model, it was possible to predict the expected mean value of the number of sessions for each task. Task 3 was chosen as a reference because it was the first task where the subject had to discriminate between two pictures of animals. Since the number of sessions to criterion was a count, we used a Poisson regression; since the observations were made on the same dogs, the effect of individual was taken into account as a random effect (Ogura 2011). In order to find the most relevant model for describing the average number of sessions required for a task, we used the minimal Akaike Information Criterion (AIC) method (Akaike 1973; Ogura 2011). The most relevant model was obtained for the Poisson regression model with dogs as random effect on the intercept and “tasks as a factor” for a fixed effect. The normality of the distribution of the residuals was assessed by graphic representation and the Shapiro test. We considered that a difference was significant when P value (P) was lower than 0.05.
In order to complete the analysis, using the same method with a generalized linear mixed model (Bates and Maechler 2010), with the dog as a “random-effect”, and the task as a “fixed-effect”, a comparison of the numbers of sessions with the same dogs was carried out for successive periods (A, B, C, D and E). The successive periods were defined as periods of monotonic increase or decrease periods in the number of sessions. This procedure establishes the best way to explain the task effect as a factor, or as an ordinal value, for each period. A significant task effect as ordinal value highlighted whether response tendencies are significantly increasing or decreasing.
The protocol (schedules and duration of the session blocks of our experiment) was approved by the Ethical Committee of VetAgro-Sup (Lyon, France) registered as number 1058, complying with French law.
Training tasks with “S+ = dog” (Task 1–3)
For the training tasks with “dog as S+”, the nine dogs reached the criterion of success for all the tasks (Figs. 5, 6). Dogs needed 2–13 sessions in training Task 1 (a single pair: the dog D1, S+, vs. a black screen S−: Median = 5; Fig. 6), 2–3 sessions in training Task 2 (a single pair: the dog D1, S+, vs. a blue screen S−: Median = 2; Fig. 6) and 2–12 sessions in training Task 3 (a single pair: the dog D1, S+, vs. the cow C1, S−: Median = 3; Fig. 6).
Estimate of parameters
*With bounty at task 10
·Without bounty at task 10
Generalization tasks with “S+ = dog” (Task 4–6)
For the generalization tasks with dog as S+, all nine dogs reached the criterion of success for all the tasks (Figs. 5, 6). Dogs needed 2–13 sessions in generalization Task 4 (12 dogs S+ vs. 12 cows S−: Median = 4; Fig. 6), 2–10 sessions in generalization Task 5 (12 dogs S+ vs. 6 cows + 6 other species, S−: Median = 6; Fig. 6) and 2–6 sessions for generalization Task 6 (12 dogs S+ vs. 12 “non-dog” species S−: Median = 4; Fig. 6).
Training tasks in reversal reward contingencies with “S+ = non-dog species” (Task 7–8)
For the training tasks, when the stimulus-reward contingency was reversed, the nine dogs met the success criterion for all the tasks (Figs. 5, 6). Dogs needed 2–4 sessions in reversal training Task 7 (a single pair: the cow C2, S+ vs. a blue screen S−: Median = 2; Fig. 6) and 3–27 sessions in reversal training Task 8 (a single pair: the cow C2, S+ vs. the dog D2, S−: Median = 17; Fig. 6).
The generalized linear mixed model showed that the number of sessions needed for Task 8 (a single pair: the cow C2 vs. the dog D2) was significantly (P < 0.001) greater than that for Task 3 (a single pair: the cow C1 vs. the dog D1; Table 2). In contrast, the number of sessions needed for Task 7 (a single pair: the cow C2 vs. a blue screen) was significantly (P < 0.05) lower than for Task 3 (Table 2; Fig. 6).
Generalization tasks in reversal reward contingencies with “S+ = non-dog species” (Task 9–10)
For the generalization tasks in reversal reward contingencies, the nine dogs met the success criterion for all the tasks (Figs. 5, 6). Dogs needed 2–14 sessions in the reversal generalization Task 9 (12 cows S+ vs. 12 dogs S−: Median = 4; Fig. 6), and 2–3 sessions in the reversal generalization Task 10 (12 “non-dog” species S+ vs. 12 dogs S−: Median = 3; Fig. 6).
The generalized linear mixed model showed that the number of sessions needed for Task 9 was not significantly different from Task 3 (Table 2). With respect to Task 10, the results were different depending on whether the dog Bounty was taken into account or not. Bounty needed a considerable number of sessions to reach criterion on Task 10 (39 sessions). With Bounty included, median performance in Task 10 was significantly more difficult than in Task 3 (P < 0.05; Table 2; Fig. 6). Without Bounty’s data, Task 10 was marginally less difficult than the reference Task 3 (P < 0.1; Table 2; Fig. 6).
Estimate of parameter with the standard error in brackets
Progression across the tasks differed between individual dogs (Fig. 5). One dog (Vodka) presented an extreme pattern. This dog took more time for the learning Task 0 (with 29 sessions), but then, it succeeded rapidly with all of the following tasks, needing only 5 sessions for Tasks 1 and 2, and after that, no more than 3 sessions for each task. This dog needed fewer and fewer sessions to meet the criterion for the subsequent tasks. Ultimately, this dog was among those which needed the lowest number of generalization sessions (i.e., 12 sessions = 144 trials). On the contrary, Bounty needed increasing numbers of sessions to reach the criterion for the later tasks, with a peak at 39 sessions for the final generalization Task 10. Bounty was among those needing the greatest number of generalization sessions (i.e., 56 sessions = 672 trials).
Our results explore the dog’s ability to visually discriminate 2D pictures of the faces of various species depending on whether they represent dogs or not. Behavioral studies investigating the capacities of dogs to use visual cues for face identification are still relatively sparse compared with humans and other animals such as non-human primates, sheep and heifers (Leopold and Rhodes 2010). Compared to previous studies investigating such abilities in domestic dogs (Range et al. 2007b; Racca et al. 2010; Somppi et al. 2012), our study is the only one using as stimuli species other than dogs and humans, that is, domestic species (cats, cows, sheep, horses, etc.) and wild species (tigers, birds, rodents, etc.).
Moreover, in our study, the dogs were confronted by a large diversity of stimuli: for the images of dog faces, the four morphological types of dogs were used in balanced proportions, from the smallest such as Chihuahua (1 kg) to the largest such as mastiff (100 kg). Also images of “non-dog” species included about forty different species in roughly equal proportions. As a result, our subjects were confronted by more than 144 pictures of morphologically different dogs versus others species (144 being the number of trials performed by the dog which was the fastest to complete the successive tasks), whereas the number of stimuli in “non-dog” category was less than 30 in previous studies (Racca et al. 2010, Somppi et al. 2012).
Thus, our study may suggest that dogs can form a visual category of “dog pattern”, as assumed in rhesus macaques (Yoshikubo 1985). We may then hypothesize that there may exist some invariants in dog morphotypes that allow the nine subjects to group pictures of very different dogs into a single category despite the great diversity in canid species. The rapid generalization from a single training instance (a single pair: a dog picture versus a cow picture in Task 3) to multiple new instances in Task 4 goes against the “category size effect”, which has been thoroughly explored in pigeons (cf. Soto and Wasserman 2010): in general, when animals are trained on category discriminations, generalization is quite poor when only a few instances are used in training, let alone only one. There is one well-known exception, where training pigeons to respond to a single oak leaf silhouette image led to instant generalization to all oak leaf silhouettes (Cerella 1979). Cerella interprets this unexpected result by saying that the oak leaf pattern is “transparent” to the pigeons, that is, they do not have to learn its extension. Could this hypothesis apply with the “dog pattern”? This might be possible in light of the performances in the first sessions of generalization (Task 4), and especially in the first trials of the first session of that task: the performances of the dogs were generally above chance for the majority of them.
But there might be another explanation for this very rapid generalization: the performance could be the consequence of (1) progressive training with the same picture of the dog D1 versus a black screen (Task 1), and then versus a blue screen (Task 2) and further a single cow C1 (Task 3); (2) the fact that the instances in Task 4 included only cows as S− (not other species) against dogs as S+. When in Task 5 a greater diversity appeared in the trials with other species than cows presented against dogs, the subjects had more trouble in reaching the criterion, although this difference was not significant. Dogs’ morphology varies more than that of cows (Wayne and Ostrander 2007). In Task 4, the dogs may have developed a strategy “not to choose the cow picture” in order to choose the picture of a dog. But this strategy was no longer possible in Task 5 with various species (not only cows) pitted against various dogs. It may be that ultimately it was only in Task 5 that the subjects understood the “dog categorization” required by the experimenter. The “dog category” is an insight which has been especially explored in various species. For human babies, cats are treated as a kind of dog, but dogs are not treated as a kind of cat (Eimas et al. 1994; Quinn and Eimas 1996). Experiments conducted on humans and pigeons confronted by pictures of dogs and cats showed that pigeons and humans do not form categories using the same features (Ghosh et al. 2004; Goto et al. 2011). We assume that such differences may exist between dogs and humans, and further investigations are needed to support the idea and the nature of “a dog species pattern” in dogs as Cerella (1979) suggested with the oak leaf pattern for pigeons. According to the five levels of categorization of Herrnstein (1990), from simple discrimination based on perceptual cues to categorization based on complex concepts, we cannot conclude more than that dogs based on their categorization of dog faces on perceptual cues.
Another unusual feature of our study was the use of a reversal of the stimulus-reward contingency. This procedure has not often been used in category discrimination studies, particularly in dog studies. However, reversal learning in categorization studies can be used to strengthen the demonstration of categorization abilities. Spence’s theory (1960) predicts that the more the initial discrimination is learned, the more difficult it will be to learn the reversed stimulus-reward contingency. This theory has been supported by numerous studies (e.g., in pigeons: Vaughan 1988; heifers: Coulon et al. 2007). In contrast, our dogs succeeded easily in the first reversal task (Task 7). This could be explained by the “overtraining reversal effect ORE” (Sperling 1965; Sweller 1973), where under some conditions, and contrary to Spence’s theory, overtrained subjects (as our dogs could be considered by numerous successive training) acquire the reversal discrimination more easily than those trained only to criterion. On the other hand, the ORE effect may not be the only explanation of the small number of sessions required for the first reversal task. We may suppose, as observed in the individual performances in the first session of Task 7, dogs rapidly learned to choose the screen containing an image, not just a dog, versus a blue uniform screen. This explanation seems consistent with the peak in the following task 8 (a same single pair: the cow C2 as S+ versus the dog D2 as S−), showing the persistence of the initial habit as described by Spence (1960).
Each procedure that can be used to investigate categorization abilities has methodological advantages and disadvantages. With preferential looking procedures, some factors, such as attractiveness and interest in certain categories of stimuli which carry more informative value, can affect the preferences for a category (Buswell 1935; Faragó et al. 2010). This interest could be the result of the diversity within a category: for instance, as between photographs of dogs and those of humans, dogs could be attracted by the diversity of dog stimuli while human faces belonging to a single human race (Caucasian) presented less diversity to them (Somppi et al. 2012). Category-dependent gazing behavior could indeed be a consequence of differences in the physical complexity of the stimuli. As a matter of fact, even though they used sophisticated eye tracking procedures, the authors conclude that they “cannot draw any conclusions as to whether the attention of dogs was directed mainly by stimulus features or semantic information, or both” (Somppi et al. 2012). This bias could also explain contrasting results in different studies using the same paradigm: Somppi et al.’s dogs fixated on a familiar image longer than on novel stimuli regardless of the category (dogs or humans), whereas it was not the case in the study of Racca et al. (2010). The use of an operant positive conditioning method has the advantage of allowing the subjects to make an unambiguous choice. To the best of our knowledge, our study is the first one using such a procedure to demonstrate species discrimination in dogs.
It remains true, however, that the conditioning procedure also has some disadvantages, because reinforced (or not) choices may affect the motivation of the dog, and ultimately the outcome of the trials. However, this motivation effect may influence the number of sessions needed to reach the criterion more than the final outcome of the trials. The repetition of tasks may lead to “learning set” formation (Harlow 1949). Learning set formation refers “to the learning of visual and other types of discrimination problems more quickly as a function of training on repeated series of such problems” (Schrier 1984). Due to the succession of the tasks, dogs had greater discrimination experience when it came to reversal Task 9 than in the equivalent Task 4.
Our results do not allow us to determine which dog morphotypes or which species were easier to discriminate by our dogs because 1) we used crossbreed dogs as stimuli, whose morphological types were not always well defined 2) the number and diversity of stimuli presented to the dogs was high and was not similar because they depended on the dog’s facility to reach the criterion. In our study, the “Category size effect” (Soto and Wasserman 2010), instead of reducing the learning speed of our dog subjects, may in fact have facilitated their extraction of the “Common elements” (Soto and Wasserman 2010) in dog faces, despite the large phenotypic diversity. “Error-driven learning” (Soto and Wasserman 2010) may have played a role in the rapid generalization shown in the difficult reversal task. This “error-driven” learning is probably based on the analysis of the stimuli themselves rather than on a strategy based on choosing a side. In the latter case, this would have not led to the rapid improvement observed. Thus by considering the stimuli, the dogs may have learned to reverse their responses based on “error-driven” learning. This kind of learning, in relation to the “Common elements” theory (Soto and Wasserman 2010), leads us to assume that our dog subjects only formed a “dog category” rather than a “dog category” and a “non-dog category”. It is likely that having already formed a “dog category” based on a reward contingency in the generalization tasks, the subjects may have shifted their strategy in relation to the two stimuli, and in recognizing the “dog category”, chose the “non-dog” stimuli based on the reward associated to it. While dog faces were likely categorized, the “non-dog” stimuli were only “identified” (Soto and Wasserman 2010). It is then likely that the reversal learning task has helped the dogs to strengthen their categorization of the dog species.
Whereas paradigms using spontaneous responses have to deal with subjects’ motivation and attention, requiring a reduction in the duration of the session and likely the number of stimuli used, paradigms that use operant conditioning also have to deal with frustration. Failing to respond correctly leads to no reward. The effect of this frustration may have varied consequences depending on the subject’s temperament (Svartberg and Forkman 2002). Differences in temperament and emotional stability may explain some differences in the dogs’ performance. Another temperament trait, perseverance, might be useful in interpreting individual differences in performance. We have observed that often errors came in bouts, thus making the session unsuccessful (only 2 errors were accepted if the session was to reach the criterion for success). This trait could explain both the rapid success of some individuals (Babel, Bag, Cyane, Vodka) when they applied the right rule and stuck to it, and difficulties in reaching the criterion when perseveration was applied to errors, especially when dogs persistently stuck to their preferred side (“Laterality effect”: Bahia, Bounty). Perseveration is likely an important temperament trait in dogs, as it has “doggedness” as a synonym. Another important aspect for interpreting dogs’ differences in performances, and one which is quite specific to dogs, is their degree of obedience training. On the one hand, a very obedient dog may have initial difficulties in adapting to the experimenter and the new commands used in the procedures. This “smart obedient” dog will then become a “slow learner” (e.g., Vodka). On the other hand, a less obedient dog might have difficulties in adapting to the constraints of the experiment and fail to be attentive to the screens, and then be a “slow learner” throughout the experiments (e.g., Bahia). Another aspect that was mentioned by Morgan (1898) is the fact that an initially “slow learner” might become the “fastest learner” when the stimulus-reward contingency is reversed as he might be more flexible once he has understood a rule (such as Vodka). A perseverant dog might have difficulties in reversing the rule (such as Bounty).
Moreover, our study also shows that dogs are able to discriminate unfamiliar dog faces (and likely “non-dog” faces) in pictures from different viewpoints (front, profile, etc.). This capacity has already been shown in domestic species, in sheep (Kendrick et al. 2001a; Ferreira et al. 2004) and heifers (Coulon et al. 2007, 2009, 2010). As we were using a large variety of “non-dog” species, we had to reduce or increase the size of natural stimuli to adjust it to a standard surface. We did that in order to prevent the dogs making their discrimination on the size of the surface of colored pixels different from the uniform blue background. Although some authors suggested that still images as they may change the natural size of the stimuli reduce their informational content (Bovet 1999; Van der Velden et al. 2008), our dogs succeeded in spite of this drawback. This recalls the performances of pigeons which were not impaired when the size of stimuli was modified (Lombardi and Delius 1990).
In our study, we presented stimuli against a standardized background even though dogs seem able to discriminate a dog picture with heterogeneous background such as landscape (cf Range et al. 2007b). This ensured that our dog subjects’ performances were based on elements found within the pictures, or the contour of them, rather than on features present in the background with no relation to the categories tested.
In conclusion, we have demonstrated that dogs are able to discriminate their own species in 2D pictures of faces alone, from different viewpoints, as shown previously in several other species (for a review Leopold and Rhodes 2010). The species discrimination demonstrated in our study might be considered as an “open class” categorization (Herrnstein 1990), as the dog faces presented covered the great variability in dog breed. This phenotypical diversity includes both contour and intrafigural features of dog faces. A natural further step would be to determine the salient features of dog faces, both common and stimulus specific elements (Soto and Wasserman 2010) on which the subjects relied to make their discrimination and their categorization. It might also be possible that the species “open class” categorization that we proposed might be of a higher level such as a conceptual one (Herrnstein 1990). This would be the case if dogs were also able to group in the same category familiar conspecifics and in other category non-familiar ones.
The fact that dogs are able to recognize their own species visually and that they have great olfactory discriminative capacities insures that social behavior and mating between highly morphologically different breeds is still potentially possible and therefore that, although humans have stretched Canis familiaris to its morphological limits, its biological entity has been preserved.
We thank Professor Charles T. Snowdon for his useful comments and careful editing on the manuscript. Thanks are also due to VetAgro-Sup which enabled our project to be carried out, to vet students Cindy Ribolzi and Florent Roques for their assistance in experimental procedure, to owners of our subjects who entrusted their dogs to us and to Royal Canin® for providing food rewards for dogs.