Two basic tasks have been used to study object categorization by pigeons. Early research used go/no-go tasks, in which a single response is rewarded in the presence of some stimuli (go trials), but not in the presence of other stimuli (no-go trials). In the first published study in this area, by Herrnstein and Loveland (1964), pigeons were rewarded after pecking at a response key when a photograph included people, but they were not rewarded for responses to photographs without people. More recent research has used forced-choice tasks (see Figure 1A), in which several responses are made available at the same time (introduced by Bhatt et al., 1988). Pigeons are rewarded only when they peck at the response key assigned to the presented stimulus.

A large number of studies using both of these tasks have shown that pigeons can learn to categorize objects through feedback and, more importantly, pigeons can generalize discriminative performance to novel objects never seen before. The typical pattern of results is high performance with novel objects, but at a slightly lower level of accuracy than with the original training objects.

Pigeons are capable of learning categories comprising natural objects (Herrnstein and Loveland, 1964; Herrnstein et al., 1976; Herrnstein and De Villiers, 1980; Bhatt et al., 1988; Aust and Huber, 2001, 2002) human-made objects (Bhatt et al., 1988; Wasserman et al., 1988; Lazareva et al., 2004, 2006), scene gist (Kirkpatrick et al., 2014), cartoons (Matsukawa et al., 2004), human face identity (Soto and Wasserman, 2011), gender (Troje et al., 1999; Huber et al., 2000) and emotional expression (Jitsumori and Yoshihara, 1997), and even paintings from different artists (Watanabe et al., 1995; Watanabe, 2001).

The fact that pigeons can accurately classify new objects from known categories suggests that their brains can extract visual properties which are invariant across diverse members of such object categories. However, the information that the pigeon visual system extracts from images is even richer, allowing them to flexibly categorize the same images at different levels. For example, pigeons can learn pseudocategorization tasks (Figure 1B), in which photographs containing objects from several categories are randomly assigned to different sets (Herrnstein and De Villiers, 1980; Wasserman et al., 1988). Focusing on category-relevant visual information would actually hinder performance in pseudocategorization tasks. Thus, the birds must be capable of extracting many different object properties from photographs, some of them invariant across members of the category and others specific to a particular object.

In line with this idea, studies that have directly manipulated object properties in photographs have found that many features simultaneously control pigeons’ performance in a categorization task (e.g., Huber et al., 2000; Aust and Huber, 2002; Lea et al., 2013), with variations in performance being well explained as a linear function of the presence or absence of such features (Huber and Lenz, 1993; Jitsumori and Yoshihara, 1997).

An important aspect of human categorization is that the same object can be flexibly categorized at several different hierarchical levels. For example, the photograph of a human can be categorized at the so-called “basic” level as a person, at the “superordinate” level as an animal, and at the “subordinate” level as “John”. Pigeons, too, have shown the ability to flexibly categorize the same objects (cars, chairs, flowers, and people) at different levels, depending on task demands (Figure 1D; Lazareva et al., 2004). The procedure used to train such flexible categories is illustrated in Figure 2. When the photograph of a human is presented together with four response keys, the pigeons learn to classify it at the basic level (Figure 2A), whereas when the photograph is presented together with two different response keys, the pigeons learn to classify it at the superordinate level of “natural object” (Figure 2B), comprising both people and flowers.

The success of pigeons in the task shown in Figure 2 is evidence for the flexibility in their categorization skills. However, it could be argued that learning to give the same response to two object categories is a far cry from forming a common superordinate representation for them. Other evidence shows that pigeons do in fact learn common superordinate representations in this type of task. For example, when objects from two perceptually dissimilar categories are associated with the same response, new learning obtained with objects from one of the categories automatically transfers to objects from the other category (Wasserman et al., 1992; Astley and Wasserman, 1998, 1999). This transfer suggests that training with a common response leads to the emergence of a single representation for both categories, which then mediates new learning about either of them. Such learning of a common representation for all stimuli associated with the same response is not restricted to superordinate categories, as it can be found after training with basic categories (Vaughan and Herrnstein, 1987) and with pseudocategories composed of two or more perceptually-dissimilar stimuli (Vaughan, 1988). This learning phenomenon, named stimulus equivalence in the behavioral literature (for a review, see Zentall et al., 2014), can also be found when members of a category share a common association with a particular stimulus or reward, instead of with a specific response.

In summary, the basic features of object category learning in pigeons are the following. First, pigeons can learn a variety of complex object categories and transfer this learning to novel objects. Second, pigeons can flexibly classify the same object according to different criteria (e.g., pseudocategories and superordinate categories). Third, pigeons extract a rich variety of visual properties from photographic images and use them in combination to learn the structure of object categories. Finally, pigeons learn common abstract representations for all members of the same trained category.

Several factors affect both the speed with which pigeons learn new object categories and the level to which they can generalize this knowledge to unseen objects. One of the factors that has a strong effect on object categorization by pigeons is the similarity relations between objects in the same category (within-category similarity) and between objects in different categories (between-category similarity) included in the same training task. It is generally believed that natural basic object categories have a higher level of within-category similarity than between-category similarity, what is termed “perceptual coherence”. For this reason, several early studies sought evidence as to whether pigeons could perceive and use such perceptual coherence for categorization, in contrast to just learning object categories by rote memorization of the images.

For example, Astley and Wasserman (1992) rewarded pigeons for pecking at photographs from a target category and measured to what extent the pecking response generalized to non-rewarded test objects. Some of these test objects belonged to the target category and others belonged to different categories. Higher responding to objects from the target category would be a indication that pigeons perceive within-category similarity as being higher than between-category similarity. Such categorical generalization was high early in the experiment, but slowly fell as pigeons acquired experience with non-rewarded presentations of the test stimuli.

Several pieces of evidence suggest that the perceptual coherence of object categories biases pigeons to group objects together into basic categories, even when this categorical bias goes against the prevailing task demands and is therefore costly in terms of earned food reward. One example comes from experiments comparing the learning of real categories and pseudocategories (Figures 1A,B). When the perceptual coherence of categories is eliminated by randomly assigning objects to pseudocategories, learning of the task slows down compared to when perceptual coherence is maintained (Herrnstein and De Villiers, 1980; Wasserman et al., 1988).

A categorical bias is also clearly observable in “subcategorization” tasks, in which two different responses are assigned to objects from the same category. In one experiment (Wasserman et al., 1988), illustrated in Figure 1C, objects from one category were assigned to two separate response keys, and objects from a second category were assigned to two other response keys. In this task, if the pigeons randomly choose a response key, then they get 25% correct responses. Pigeons can also learn about which two response keys are associated with each category, in which case they get 50% correct responses, but 50% categorical errors. Thus, this categorization strategy leads to above-chance performance, but it is not the strategy leading to the best payoff. The optimal strategy is learning to identify each individual stimulus and its correct response. When Wasserman et al. estimated the percentage of trials in which the pigeons were following each strategy, they found the results shown in Figure 3A. Although it is not the best strategy, pigeons first learn to categorize stimuli, and only later learn to identify them.

The categorical strategy shown by pigeons in the early blocks in Figure 3A is not optimal, but it does produce better reward payoff than guessing. Soto and Wasserman (2010b; see also Soto et al., 2012) found that a similar categorical bias can be found using a go/no-go subcategorization task. In this task, responses to a group of objects never produce reward, yet early in training pigeons respond to them at the same level as to rewarded objects from the same category. That is, pigeons learn first to categorize objects in subcategorization tasks, regardless of whether or not this strategy produces reward.

The previous experiments all suggest that pigeons perceive the within-category similarity of objects in natural photographs to be higher than their between-category similarity. This result is not trivial; it is important that the category structures that pigeons are biased to learn are exactly those that are likely to be encountered in the natural environment (see Smith et al., 2010). However, even when they are learning artificial categories, pigeons (Cook and Smith, 2006) and primates (Blair and Homa, 2003; Smith et al., 2010) show a bias to learn perceptually-coherent category representations before learning information about specific stimuli.

Differences in between-category similarity also play a role in category learning. For example, Aust and Huber (2002) concluded that how much responding to the trained category “person” generalized to similar or related categories (such as dolls, primates, mammals, and birds) depended on how many features were shared by the categories.

When pigeons are concurrently trained to classify the same categories at both basic and superordinate levels, it is usually found that they learn the basic task faster for some categories and the superordinate task faster for other categories (Lazareva et al., 2004, 2010; Lazareva and Wasserman, 2009). Lazareva et al. (2010) obtained estimates of the similarity among four object categories by analyzing generalization data through multidimensional scaling. Then, they showed that such similarity estimates could predict whether the basic or superordinate levels would show an advantage for different pairs of categories. A superordinate-level advantage is seen when the two categories in a superordinate set are perceptually similar, whereas a basic-level advantage is seen when the two categories in a superordinate set are perceptually dissimilar. This result is interesting because it is in line with one of the hypotheses put forward to explain the basic-level advantage in humans (Rosch et al., 1976).

Factors related to the training regime also affect category learning and generalization. One such factor is the number of different objects in each category presented during training (Kendrick et al., 1990; Astley and Wasserman, 1992; Wasserman and Bhatt, 1992). As shown in Figure 4A, learning of the categorization task is slowed and transfer of performance to new images is enhanced with a higher number of training exemplars. In the extreme case, in which training images are never repeated, pigeons can still learn the object categories, but learning is slower than when the training exemplars are repeated (Bhatt et al., 1988).

Another important training factor for studies using a go/no-go task is whether responses are rewarded to images showing the category, in what is called a feature-positive task, or to images showing no category, in what is called a feature-negative task. For example, Edwards and Honig (1987; see also Aust and Huber, 2001, 2002) trained pigeons to discriminate photographs of various scenes from photographs of the same scenes with people in them. Their results, reproduced in Figure 5A, show that pigeons were quite fast in learning the feature-positive discrimination, in which responses to people were rewarded, but they were slow in learning the feature-negative discrimination, in which responses to scenes without people were rewarded. In fact, learning of the feature-negative discrimination was as slow as learning a pseudocategorization task, suggesting that pigeons do not show any benefit from perceptual coherence when responses to the category are not rewarded.

Patterns of generalization also vary for feature-positive and feature-negative tasks. Aust and Huber (2001) trained pigeons with the “people” category in feature-positive and feature-negative tasks. After training, pigeons were presented with new combinations of background scenes and people that involved contradictory information. For example, either familiar or novel people, which were associated with one outcome during training (e.g., reward), could be presented on familiar backgrounds, which were followed by the opposite outcome during training (e.g., no reward). The authors found that feature positive training led to generalization of the response learned for people to these conflicting test stimuli, whereas feature negative training led to no preference to respond or to inhibit responding to conflicting test stimuli. Again, this finding suggests that learning about the whole category is possible only when responses to the category are rewarded, but not when responses to the category are not rewarded.

More recent experiments, motivated by the model described in the previous section, have led to more direct evidence for the role of error-driven learning in object categorization by pigeons. The important insight provided by the model is that different tasks can be used to manipulate the connections between different types of elements (category-specific and stimulus-specific) and responses.

One example is the blocking design illustrated in Figure 6A. In the blocking condition (Soto and Wasserman, 2010b,d), objects from the same category are first assigned to different responses in a pseudocategorization task (Phase 1). According to the model, accurate performance in this task requires strong connections between stimulus-specific elements and the correct responses. Once the pseudocategorization task is learned, it is possible to transform it into a true categorization task by dropping half of the trials, as shown in the middle panel of Figure 6A (Phase 2). Under normal circumstances, experience with this new categorization task should lead to strong control by category-specific elements and good generalization to new objects when they are presented during a test (Phase 3). In a control condition, pigeons were exposed only to this categorization task and a generalization test (Phases 2 and 3). In the blocking condition, however, the stimulus-response mapping is already known at the beginning of Phase 2; thus, pigeons should make few, if any, errors in predicting the correct response for each of the stimuli in this phase. No prediction error means no category learning; so, the model predicts less generalization of categorical performance to new objects in the blocking group than in the control group.

The predictions of the model and the performance of pigeons with novel objects in each condition are shown in Figures 6B,C, respectively. It can be seen that pigeons showed the predicted pattern of results. This blocking effect, analogous to effects found in Pavlovian conditioning (Kamin, 1969), is direct evidence that object category learning in pigeons is driven by reward prediction error.

The blocking effect also helps to explain some contradictory results in the literature. For example, Sutton and Roberts (2002) used a design very similar to that of Astley and Wasserman (1992) to study the “perceptual coherence” of object categories, but found that generalization was the same to objects from any category, not only the target category. We have shown (Soto and Wasserman, 2010b) that Sutton and Roberts’ results can be explained as a blocking effect, in which elements common to all of the object categories acquire control over performance early in training.

Other studies have found evidence of an overshadowing effect in category learning (Soto and Wasserman, 2012a; Soto et al., 2012). Figure 7A shows a schematic representation of the training tasks given to pigeons in one of these experiments (Soto and Wasserman, 2012a). On each trial, two different objects were presented to the pigeons. In the overshadowing condition, these objects came from two categories that were both informative about the correct response. For example, in Figure 7A, both airplanes and chairs were consistently associated with Response 1. Here, the category-specific elements of both categories should acquire control over behavior quite fast, quickly reaching a point in which performance is good and learning stops. At this point, the two categories overshadow each other: each acquires only a proportion of the response control that they would have gained if they had been presented alone. In the control condition, two objects are presented in each trial, but a single target category is informative about the correct response. In the example in Figure 7A, butterflies and cars are informative about correct responses, but people and flowers are not. In both conditions, category learning was tested by presenting pigeons with new objects from the trained categories. As shown in Figure 7B, the model predicts that performance with the target categories (red bars) should be impaired in the overshadowing condition compared to the control condition. As shown in Figure 7C, this prediction of the model matched the pigeons’ behavior. Furthermore, performance with the competing categories (blue bars) was also close to the model’s predictions.

Given the accumulated evidence suggesting that error-driven learning plays an important role in object categorization by pigeons and the fact that this form of learning is widespread across species and tasks (Siegel and Allan, 1996; Bitterman, 2000; Macphail and Bolhuis, 2001), it seems likely that similar mechanisms underlie object categorization in primates, including humans.

A repetition advantage effect for category-specific properties seems to be as important in people as it is in pigeons. For example, the categorical bias effects and category size effects that are pervasive in the pigeon literature can also be found in people and other primates (Homa et al., 1973; Smith and Minda, 1998; Minda and Smith, 2001; Smith et al., 2010). As indicated earlier, such effects result naturally from the interaction of a repetition advantage for category-specific information and error-driven learning (see Soto and Wasserman, 2010b).

The results of behavioral experiments suggest that error-driven learning plays an important role in object categorization in people. Just as with pigeons, when people are trained to solve a discrimination task by memorizing individual objects in photographs and their assigned responses, they are impaired in detecting a change in the training circumstances in which all of the presented objects are sorted according to their basic-level categories (Soto and Wasserman, 2010d). That is, people show a category blocking effect, as illustrated in Figure 6D (see also Gluck and Bower, 1988; Shanks, 1991; Nosofsky et al., 1992).

We have proposed (Soto and Wasserman, 2012c) that underlying these behavioral similarities is an evolutionarily conserved learning mechanism that might be implemented in the basal ganglia, which are homologous structures in birds and mammals (Reiner, 2002; Reiner et al., 2005). Many studies implicate the basal ganglia in visual categorization and other visual discrimination tasks in people and other primates (Ashby and Ennis, 2006; Seger, 2008; Shohamy et al., 2008). The basal ganglia receive input from most sensory areas and send output to motor areas, which allows for the sensory integration and response selection functions necessary for category learning. The input nuclei in the basal ganglia, collectively known as striatum, receive dopaminergic input from the subtantia nigra pars compacta (Durstewitz et al., 1999; Nicola et al., 2000; Reiner et al., 2005) and the plasticity of cortical-striatal synapses depends on the presence of this dopaminergic input (Centonze et al., 2001; Reynolds and Wickens, 2002). As there is considerable evidence that the activity of these dopaminergic neurons is correlated with reward-prediction error (Montague et al., 1996; Schultz, 1998, 2002; Waelti et al., 2001; Suri, 2002), cortical-striatal synapses (pallial-striatal synapses in birds) may mediate the error-driven learning of associations between visual representations and responses. As proposed by our model, learning in the striatum during object categorization tasks would require activity of the presynaptic visual neurons (stimulus elements in the model), activity of the postsynaptic striatal neurons (actions in the model), and the presence of a dopaminergic signal (reward prediction error in the model).

Following the human literature, much research in object recognition by pigeons has focused on whether or not this species can show recognition that is invariant to changes in identity-preserving variables, such as rotation, scaling, illumination, etc. In general, the results of psychophysical experiments all point to the same conclusion: pigeons’ object recognition after training with a single object image is controlled by a variety of properties that are irrelevant to object identification. In order to show invariant object recognition, pigeons require training with variations in such irrelevant properties.

For example, experiments that have explored whether pigeons show view-invariant object recognition after being trained with only one object view have uniformly found significant costs of object rotation on accuracy, regardless of the type of object used to generate the experimental stimuli (Cerella, 1977; Lumsden, 1977; Wasserman et al., 1996; Peissig et al., 1999, 2000, 2002; Friedman et al., 2005). Similarly, other experiments have found that, after experience with a single image view, pigeons’ object recognition is affected by variations in size (Larsen and Bundesen, 1978; Pisacreta et al., 1984; Peissig et al., 2006), shading (Cabe, 1976; Cook et al., 1990; Young et al., 2001), and position (Kirkpatrick, 2001).

Although object recognition in people is far from being completely invariant (Jolicoeur, 1987; Hayward and Tarr, 1997; Tarr et al., 1998; Kravitz et al., 2008), it is clear that humans show greater invariance than do pigeons (Biederman and Ju, 1988; Biederman and Cooper, 1992; Biederman and Gerhardstein, 1993; Hayward, 1998). For example, people, but not pigeons, have been shown to exhibit view-invariant recognition when they are tested with the appropriate stimuli (Biederman and Gerhardstein, 1993) and show view-invariant recognition of novel views of an object which are interpolated between experienced views (Spetch and Friedman, 2003). Furthermore, some factors that are known to foster view-invariance in people do not have the same effect in pigeons. People show rotation costs when recognizing bent-paperclip objects (e.g., Edelman and Bülthoff, 1992), but these costs are reduced when a single diagnostic geometrical volume (“geon”) is added to each object (Tarr et al., 1997). The same results are not observed in pigeons (Spetch et al., 2001), which show decrements in performance as a function of rotational distance regardless of the object components.

On the other hand, pigeons show generalization behavior that is closer to true view invariance as the number of training views is increased (Wasserman et al., 1996; Peissig et al., 1999, 2002). This finding is essentially another manifestation of the category size effect described earlier and can be explained in the same way: that is, as arising from a repetition advantage effect for view-invariant properties, which are repeated often across different views and therefore are frequently paired with the correct responses.

If this explanation of view-invariance learning in pigeons is correct, then it should be possible to arrange conditions in which training with multiple views of an object does not lead to higher invariance, by reducing the advantage of view-invariant properties over other properties during training. Soto et al. (2012) recently tested this hypothesis by training pigeons with object views similar to those shown in Figure 8A. In the training images for the overshadowing condition, across variations in viewpoint, there is a pronounced feature that is not view-invariant and that can perfectly predict object identity: the orientation of the main axis. The repetition of this feature across views should produce something akin to the category overshadowing effect explained earlier and impair view-invariance learning. On the other hand, in the control condition, pigeons are trained with the same views of less elongated objects; this training eliminates the competing non-invariant feature of main-axis orientation, which should result in higher invariance. Figure 8B shows that performance with new views was above chance for the control condition and below chance for the overshadowing condition, just as predicted.

Humans do sometimes show rotational costs in object recognition tasks (Hayward and Tarr, 1997; Tarr et al., 1998), which diminish after training with multiple views (Mash et al., 2007). These findings raise the possibility that view-invariance learning in people might follow similar principles as in pigeons, being driven by prediction errors. A role for error-driven learning has been found in human object categorization (Soto and Wasserman, 2010d) and there is evidence that categorization and identification depend on similar neural representations and computations (e.g., Hung et al., 2005).

This possibility has remained unexplored in the primate literature, which has focused instead on looking for evidence of unsupervised learning of invariant object representations (Cox et al., 2005; Li and DiCarlo, 2008, 2010, 2012). It must be noted that the evidence gathered so far does not rule out a role for reward prediction error in invariance learning. In the monkey experiments carried out by Li and DiCarlo (2008, 2010), for example, animals were rewarded for looking at the presented objects. In similar human experiments (Cox et al., 2005), people were engaged in a task that involved “correct” and “incorrect” responses and learning was not observed when experience was delivered passively (Li and DiCarlo, 2012). Thus, these experiments do involve presentation of explicit and implicit rewards and clearly raise the possibility that learning is driven by reinforcement (Li and DiCarlo, 2010). Although one study (Li and DiCarlo, 2012) reported evidence of unsupervised learning independent of reward magnitude and timing, it did not show that reward is not necessary for invariance learning. On the other hand, Yamashita et al. (2010) have provided evidence that reward-based discrimination, and not simple exposure, is necessary for invariance learning at least under some circumstances.

The role of unsupervised learning mechanisms in object recognition by pigeons has also remained unexplored. As our discussion of the primate literature shows, one reason is that it is quite difficult to study unsupervised learning in isolation from the influence of reward, particularly in nonhuman animals. This is an important issue that should be addressed by future research.

Despite the difference in invariant recognition shown by people and pigeons, there is considerable evidence that the two species rely on similar image information during object recognition tasks (Wasserman and Biederman, 2012). For example, both primates and pigeons seem to extract nonaccidental properties from images of geons and rely heavily on them for recognition (e.g., Biederman and Bar, 1999; Vogels et al., 2001; Kayaert et al., 2005; Gibson et al., 2007; Lazareva et al., 2008). Gibson et al. (2007) trained pigeons and people to discriminate four simple objects, each shown from a single viewpoint. Using the Bubbles technique (Gosselin and Schyns, 2001), it was determined that both species relied more heavily on image properties that are relatively invariant across changes in viewpoint, such as cotermination and other edge properties, than on properties that vary across changes in viewpoint, such as shading. This result is depicted in the leftmost group of bars in Figure 9.

Results such as those shown in Figure 9 do not mean that pigeons rely only on view-invariant properties for object recognition. As mentioned earlier, pigeons are sensitive to changes in object viewpoint, size, location, and shading, which means that all of these properties are extracted and used by pigeons during object recognition tasks. The inability of pigeons to show one-shot view invariance is not the result of an inability to extract view-invariant representations. Instead, it is more likely that pigeons extract a rich variety of visual properties from images and can only gradually learn to focus on those that are relevant for a given task through a reinforcement learning mechanism.

Several experiments have found evidence that pigeons represent not only local shape properties, but also the spatial structure of objects (Van Hamme et al., 1992; Wasserman et al., 1993; Kirkpatrick-Steger and Wasserman, 1996; Kirkpatrick-Steger et al., 1998). In one study, Van Hamme et al. (1992) trained pigeons to recognize line drawings of objects, similar to those shown in Figure 10A, in which half of an object’s contour was deleted. This technique allowed the experimenters to train the pigeons with one contour image and to test them with its complement, which shared no local features with the training stimulus. As shown in Figure 10B, pigeons recognized these complementary contours with considerable accuracy, suggesting that their visual system could infer object structure from the partial contours seen during training.

Furthermore, when both shape and spatial relations can be used as cues to solve a recognition task, pigeons rely on both of them and show a trade-off between their reliance on one source of information vs. the other; that is, the more a pigeon relies on shape for recognition, the less it relies on spatial information, and vice-versa (Kirkpatrick-Steger and Wasserman, 1996). Such trade-offs can be explained as another form of overshadowing: when two object properties are equally reliable for identification, they compete for control of performance.

Comparative studies have revealed similarities and differences in high-level vision by pigeons and people not only at the behavioral level, as described in the previous section, but also at the neurobiological level. Although primate and avian visual systems are each organized into two main visual pathways, the tectofugal pathway is used for complex visual discrimination tasks in pigeons, whereas the thalamofugal pathway is used for such tasks in primates (Shimizu and Bowers, 1999; Wylie et al., 2009). Still, these pathways show similar functional organization, which has led to the proposal that they might be analogous (Shimizu and Bowers, 1999). For example, the avian tectofugal pathway and its pallial targets are organized into parallel subdivisions in charge of processing motion and shape (Wang et al., 1993; Shimizu and Bowers, 1999; Laverghetta and Shimizu, 2003; Nguyen et al., 2004; Fredes et al., 2010), which is similar to the organization of the primate thalamofugal pathway and its cortical targets (Mishkin et al., 1983; Ungerleider and Haxby, 1994).

Furthermore, there is evidence that one of the main mechanisms thought to be responsible for visual shape processing in the primate thalamofugal pathway is also at work in the avian tectofugal pathway. This mechanism, first proposed by Hubel and Wiesel (1962, 1968), relies on feedforward processing across visual areas that are hierarchically organized in terms of the complexity of the visual information that they represent. Neurons at each level of the system integrate information from neurons at the previous level to build selectivity for shape features of increasing complexity and tolerance to variables such as size and location (for a short review and references, see Soto and Wasserman, 2012c). Li et al. (2007) found that the receptive fields of neurons in the pigeon nucleus isthmus (sensitive to oriented gratings) are constructed by feedforward convergence of receptive fields from neurons in the tectum (which have center-surround organization), as proposed by the hierarchical model of Hubel and Wiesel (1962, 1968). Also in accord with hierarchical processing, there is a large increase in receptive field size from early to later areas in the avian tectofugal pathway (Engelage and Bischof, 1996).

Thus, hierarchical and feedforward processing of shape information–a central mechanism for most current neurocomputational theories of object recognition in primates (e.g., Fukushima, 1980; Perrett and Oram, 1993; Riesenhuber and Poggio, 1999, 2000; Rolls and Milward, 2000; Serre et al., 2007)– might be widespread across vertebrate visual systems. If this is true, then behavioral differences between pigeons and people must be explained by some other mechanism. We Soto and Wasserman (2012c) recently offered a proof of concept for this hypothesis, by showing that a hierarchical model of object recognition in the primate ventral stream (a version of the HMAX model described in Serre et al., 2007), coupled with a reinforcement learning model (see Section The Role of Error-driven Reinforcement Learning), can explain much of the available behavioral data in object recognition by pigeons reviewed in sub-sections Invariance in Object Recognition by Pigeons and What Information Is Extracted From Images by Pigeons?

The success of this model was surprising for two reasons. First, the model could better explain pigeon behavior than human behavior. Just as pigeons but unlike people, the model’s recognition was strongly affected by changes in viewpoint, size, and shading. In the case of size, the model could even reproduce the logarithmic relation between physical and perceived object size that has been found in pigeons (Peissig et al., 2006). Furthermore, invariant recognition was not fostered by variables that seem to do so for people, such as adding geons to paperclip objects.

Second, although this model uses a “bag of features” to mediate object representation, the results of several simulations showed that such representations can be much richer than one would initially assume. As shown in Figure 10C, the model has no problem reproducing the ability of pigeons to recognize objects from their complementary contours (Van Hamme et al., 1992). This result was originally interpreted as showing that a feature-based representation (such as that proposed by Cerella, 1986)– lacking explicit information about the spatial relations among features–could not explain object recognition in pigeons. This interpretation is only partially correct, because the simulated results suggest that the feature pool in the model can implicitly represent information about spatial structure.

The model also reproduces the bias to rely on nonaccidental properties in geon recognition found in people and pigeons (Gibson et al., 2007), as depicted in Figure 9. The model is successful despite the fact that it was not designed to do so, as in the case of other theories of object recognition (structural description theories; see Biederman, 1987). Instead, the bias emerges in the hierarchical model from simple principles of biological visual computing and because the features in the model have been trained through exposure to natural images (see Serre et al., 2007). Coterminations and elongated edges are both quite common in natural images (Geisler et al., 2001) and they could reliably distinguish between the objects used by Gibson et al. (2007).

The success of the hierarchical model in explaining the pigeon behavioral data has no equal in the current literature. Together with the results of neurophysiological studies (Engelage and Bischof, 1996; Li et al., 2007), the success of this model suggests that feedforward and hierarchical processing of visual information play important roles in object recognition by pigeons, as they do in primates.

Up to this point, we have focused on the mechanisms of visual object recognition that are likely to be shared by pigeons and people. However, the evolutionary lineages of both species diverged more than 300 million years ago; surely, we can expect their visual systems to show important differences due to adaptive specialization.

For example, it is likely that there are specialized mechanisms¹ of face perception in people and other primates (Pascalis and Kelly, 2009). However, a comparative analysis requires taking into account the fact that face recognition is a complex form of behavior, likely to result from the interaction of many mechanisms, including general processes shared with other species (de Waal and Ferrari, 2010; Shettleworth, 2010). Determining which aspects of human face perception are due to specialized vs. general mechanisms requires comparative research; here, pigeons are becoming a key species to determine the role of general recognition processes (Soto and Wasserman, 2011).

Only a handful of behavioral studies have compared human face recognition by pigeons and people. They have led to a complex pattern of results, suggesting that some properties of face perception in people are likely to be the result of specialized processes, whereas others might result from general processes. Regarding specialized processes, it has been found that, while people and other primates show an advantage in discriminating upright faces over inverted faces, the same advantage is not found in pigeons (Phelps and Roberts, 1994). It is widely believed that faces are perceived in a “holistic” or “configural” way to a larger extent than other objects (for reviews, see Maurer et al., 2002; Richler et al., 2008) and inversion effects have been proposed as a manifestation of holistic face perception (Farah et al., 1995). That is, holistic processing might be a specialized mechanism for face perception in primates.

Surprisingly, other studies have shown similarities in the way people and pigeons process human faces. For example, both species use information near the eyes and chin to discriminate gender and they use information near the mouth to discriminate emotion (Gibson et al., 2005). Also, in both people (e.g., Schweinberger and Soukup, 1998; Fox and Barton, 2007; Ellamil et al., 2008; Fox et al., 2008) and pigeons (Soto and Wasserman, 2011), recognition of emotional expression depends on variations in identity, whereas recognition of identity is relatively independent of variations in emotion. It is possible that the origin of this latter interaction in people is decisional rather than perceptual (Soto et al., 2014), which would make the similarity across species easier to reconcile with the existence of specialized face perception processes in primates.

Overall, these results challenge the common assumption that a specialized human face perception system must underlie all observed aspects of human face recognition, being somehow “encapsulated”, or free from the influence of more general processes. Furthermore, they serve to underscore the fact that the evolution of a face recognition system did not solely involve the specialization of perceptual processes, but also the specialization of the human face as an efficient transmitter of social signals (Smith et al., 2005; Schyns et al., 2009). The human face could have been specialized through evolution to transmit signals that would be easily decoded by existing visual processes. If such visual processes are also present in birds, then the fact that some aspects of face recognition are similar in pigeons and people seems less surprising.