behavioral sciences Review

Categorization: The View from Animal Cognition J. David Smith 1, *, Alexandria C. Zakrzewski 2 , Jennifer M. Johnson 1 , Jeanette C. Valleau 1 and Barbara A. Church 3 1 2 3

*

Department of Psychology, Georgia State University, 738 Urban Life Building, 140 Decatur St., Atlanta, GA 30303, USA; [email protected] (J.M.J.); [email protected] (J.C.V.) Department of Psychology, University at Buffalo, The State University of New York; Park Hall Room 204, Buffalo, NY 14260, USA; [email protected] Language Research Center, Georgia State University, 3401 Panthersville Rd, Decatur, GA 30034, USA; [email protected] Correspondence: [email protected]; Tel.: +1-470-355-3587

Academic Editor: Jennifer Vonk Received: 6 April 2016; Accepted: 31 May 2016; Published: 15 June 2016

Abstract: Exemplar, prototype, and rule theory have organized much of the enormous literature on categorization. From this theoretical foundation have arisen the two primary debates in the literature—the prototype-exemplar debate and the single system-multiple systems debate. We review these theories and debates. Then, we examine the contribution that animal-cognition studies have made to them. Animals have been crucial behavioral ambassadors to the literature on categorization. They reveal the roots of human categorization, the basic assumptions of vertebrates entering category tasks, the surprising weakness of exemplar memory as a category-learning strategy. They show that a unitary exemplar theory of categorization is insufficient to explain human and animal categorization. They show that a multiple-systems theoretical account—encompassing exemplars, prototypes, and rules—will be required for a complete explanation. They show the value of a fitness perspective in understanding categorization, and the value of giving categorization an evolutionary depth and phylogenetic breadth. They raise important questions about the internal similarity structure of natural kinds and categories. They demonstrate strong continuities with humans in categorization, but discontinuities, too. Categorization’s great debates are resolving themselves, and to these resolutions animals have made crucial contributions. Keywords: category learning; categorization; cognitive evolution; comparative cognition; animal cognition

1. Introduction Categorization is a crucial ability for humans and nonhuman animals (hereafter, animals) and a focus of research (e.g., humans [1–8]; animals [9–14]). Forming psychological equivalence classes—categories—allows consistent, adaptive behavior toward ecologically equivalent objects that have imperfect perceptual similarity. Categorization has conferred fitness advantages on vertebrates for hundreds of millions of years. For example, vervet monkeys (Chlorocebus pygerythrus) suffer serious mortality caused by predation—especially martial eagles (Polemaetus bellicosus) (e.g., [15]). Accordingly, vervets have developed call signs that warn group members to behave appropriately at the sight of eagles. In a sense, these calls denote or “name” members of the category eagle. However, you cannot hide all the time. Vervet life must go on. So categorization needs to be discriminating and selective. Members of the category eagle must be avoided but not so vultures. Crying eagle, like crying wolf, must be rare and true so that it can maximally protect against predation while minimally disrupting foraging. So cognitive categories must be efficient: sculpted to suit the ecology, tailored to the shape of the category, sized to the extension of the category. Behav. Sci. 2016, 6, 12; doi:10.3390/bs6020012

www.mdpi.com/journal/behavsci

Behav. Sci. 2016, 6, 12

2 of 24

We discuss the nature of this sculpting within its ecological and evolutionary context. This topic has rarely been discussed. We ground our discussion in the major theories of category representation. We consider the current status of two dominant theoretical debates that arise from these contrasting theories. The debates are about how animals (and humans) might be equipped to categorize within the natural world. From this follows empirical research that illuminates how animals are equipped for categorization and how they are not equipped. The latter understanding is equally important. The empirical data raise important natural-history questions about categorization. They may suggest the pathways that cognitive evolution did and did not take during the emergence of cognitive systems for categorization. They may even reveal something about the structure of natural kinds that could affect animals’ fitness and help steer cognitive evolution. Throughout, we will make it plain that animals have been profoundly important ambassadors to the study of categorization. They have brought crucial theoretical insights. They have helped shape the dialog about human categorization. They have had an important role in the sea change occurring over 10 years within the categorization literature. 2. Three Theories Three theories have dominated recent discussions of human and animal categorization [1,12]. They give useful structure for this chapter. First, there is exemplar theory. This idea is that animals store the category exemplars they encounter (e.g., different specific eagles) as whole, separate memory traces. They compare new items to these, and accept them in the group if they are enough like the stored exemplars [5,7]. In this theoretical framework the organism does not have a general and unified idea of eagles. It has a slide carousel of eagles; a Rolodex of eagles. In the exemplar view, the vervet glances up, sees the thing, and then makes many psychological, similarity-based comparisons to stored memory traces (oh, it is like eagle 56; ooooh, it is very like eagle 208; and look, as it gets really close it looks amazingly like . . . {bad ending}). This sad allegory makes the important psychological point that one’s classification algorithm—that is, the way one’s mind is composed to process category information and behave toward it—can have serious adaptive risks. If your algorithm is too slow, too stepwise, too stringent in its acceptance criteria, the consequences can be serious. This is why there could be a premium on the evolution of efficient systems of categorization in vertebrate minds. Second, there is prototype theory. This idea is that animals average their experiences with disparate exemplar to form the schema, the prototype, the central tendency of the category. They compare new items to this averaged cognitive representation and include them in the category if they are enough like it [16–18]. A prototype-based categorization algorithm also presents adaptive risks. Averaging across disparate exemplars might be difficult computationally, or take more brain, and it could be wasted brain from the perspective of exemplar theory if all one needs in memory are the individuated exemplars. Moreover, the prototype can be a poor guide to category generalization because it cannot reveal anything about the category’s boundaries or limits. These criticisms are completely fair. All theories of categorization have limitations. This could suggest that in the end organisms may turn out not to suffice with just one way of doing category business. Meanwhile, the issue is not resolved by dueling criticisms. It is a matter of empirical science. It is a question of what humans and animals do. Third, there are category rules, grounded in an elegant cognitive-neuroscience literature [19–22]. In this area, rules are the representational vehicle for an explicit learning process with these operational characteristics. Explicit processes apprehend stimuli through focused attention to few (or one) stimulus features. They learn diagnostic features. In humans, at least, explicit processes rely on working memory and the executive functions to evaluate featural hypotheses. This learning is conscious, verbalizeable, and declarative to others. Many have granted rules an important role in human categorization [20,23–27]. Interestingly, even prominent exemplar theorists have done so [25,28,29].

Behav. Sci. 2016, 6, 12

3 of 24

Ashby, Maddox, and their colleagues gave explicit rule learning a neuroscience basis through their COVIS model (Competition between Verbal and Implicit Systems) of category learning [19,21,30]. COVIS models the explicit rule system as reliant on the anterior cingulate gyrus, the prefrontal cortex, the head of the caudate nucleus, and medial temporal-lobe structures also serving declarative memory. This system is modeled to resemble the neural complex that supports attention’s executive control [31]. Until recently, the phylogenetic breadth and depth of this explicit system for category rule learning remained unexplored. One sees, though, that the neuroscience scaffolding for explicit categorization just described has interesting implications for the evolutionary emergence of rule-based category learning. Not all animals have that neural complex equivalently available, and it might have emerged through evolutionary shaping during cognitive evolution. 3. Exemplars and Prototypes The theories just described have provoked the two great theoretical debates in this literature. This review will take those debates in turn (Sections 3–7; Sections 8–12). One cannot overstate the importance of the exemplar-prototype debate in categorization. Exemplar theory (i.e., Rolodex theory) is an elegant theoretical possibility that plausibly could be part of the overall vertebrate system of categorization. We acknowledge our respect for this theory and for all theories within categorization. Exemplar theory (Section 2) has exerted enormous influence over the past 30 years. For decades, exemplar theory grounded the field’s dominant theoretical narrative—that one and only one representational/processing system serves all categorization. This processing system was deemed to be exemplar-based, of course, with organisms referring new items to multiple cognitive reference points (the stored exemplars), and including them in the category if they are enough like them. This singular process had the potential to unify categorization under one mechanism. It made categorization parsimonious. It suggested that all categories are potentially learnable (even random groupings of objects could be manageable as a category through exemplar memorization). It made categorization science “easy”, with just one process in mind and memory to understand. And so exemplar theory was attractive and became powerful. Therefore, it is profoundly important that animal-cognition studies have helped resolve this theoretical debate. 4. Prototypes in Psychological Space Suppose you are an exemplar-based vervet. You see the eagle exemplars of life. You store them in memory, spread out in the mind’s psychological space like a cloud because they are not perfectly similar (Figure 1, Es). Now you see new items that need to be classified (8 . . . 1), and, from the outside working in, you will say, impossible, no way, no, maybe, and so forth, with your potential for category endorsement increasing approaching the cloud. Just as Rosch [32] envisioned, the category would have a typicality gradient. However, this gradient has a theoretically crucial asymptote. A new item cannot ever be close to all the Es at once. It will always be near some but far from others. Even the central item 1 will still not be close to all the Es. So there might be a cap on the exemplar likeness and category belongingness of the prototype, though the prototype—as defenders of exemplar theory have noted—would be the item type most strongly classified into the category. In contrast, suppose you are a prototype-based vervet. Now you average the Es into the central prototype P (Figure 1). The prototype is a point in the mind’s psychological space. A new item can be indefinitely close to it. There is no cap on prototype likeness. The typicality gradient need not flatten. This vervet could experience a DEFCON 1 eagle that is essentially identical to its unitary category representation. Thus, inevitable geometrical facts about the geometry of exemplar-based and prototype-centered representations force differential predictions that let one finally resolve the exemplar-prototype question in an important empirical case.

Behav. Sci. 2016, 6, 12 Behav. Sci. 2016, 6, 12 Behav. Sci. 2016, 6, 12

4 of 24 4 of 24 4 of 24

Figure 1. 1. Schematic Schematic drawing drawing of of aa series series of of hypothetical hypothetical to to be be classified classified items items (8 (8 to to 1) 1) that that are are made made to to Figure Figure 1. Schematic drawingspace of a series of hypothetical to be classified itemsprototype (8 to 1) that made to approach the psychological psychological of aa category category system containing containing central (P)are and shell approach the space of system aa central prototype (P) and aa shell approach psychological space of afrom category system containing central prototype and a shell of training trainingthe exemplars (E). Reprinted Reprinted [33]. Copyright Copyright 2002 by byathe the Association for(P) Psychological of exemplars (E). from [33]. 2002 Association for Psychological of training exemplars (E). Reprinted from [33]. Copyright 2002 by the Association for Psychological Science. Reprinted Reprinted with with permission. permission. Science. Science. Reprinted with permission.

To show this, Smith et al. asked macaques to learn about families of shapes [18]. These shapes To show this, et macaques learn about families of [18]. shapes showusing this, Smith Smith et al. al. asked asked macaques to to about families of shapes shapes [18]. These These shapes wereTo created the influential dot-distortion or learn random polygon technique of Posner, Homa, and were created using the influential dot-distortion or random polygon polygon technique technique of of Posner, Posner, Homa, Homa, and and were created using the influential dot-distortion or random others [17,34]. These shape families were built from a prototype P (Figure 2). There could be lowothers [17,34]. shape families were built from a prototype P (Figure 2). There could be low-level others [17,34]. These These shape families were built from a prototype P (Figure 2). There could be lowlevel distortions of the prototype, quite similar to it, high-level distortions, less similar, and random distortions of the prototype, quite similar to it, high-level distortions, less similar, and random items level distortions of the prototype, quite similar to it, high-level distortions, less similar, and random items completely dissimilar from the prototype that should not be included within the category. The completely dissimilar from the prototype that should be included within the category. The monkeys items completely dissimilar from prototype thatinnot should beexperiments. included within the category. The monkeys successively learned 10 the shape categories each ofnot two successively learned 10 shape categories in each of two experiments. monkeys successively learned 10 shape categories in each of two experiments.

P P L L H H R R

Figure 2. A dot-distortion category, with the originating prototype (P), low-level distortions (L), highFigure 2. with thethe originating prototype (P), low-level distortions (L), highlevel distortions (H), andcategory, random-unrelated shapes (R) inprototype the top to rows, respectively. Figure 2. AAdot-distortion dot-distortion category, with originating (P),bottom low-level distortions (L), level distortions (H), and random-unrelated shapes (R) in the top to bottom rows, respectively. Adapted from [18]. (H), Copyright 2008 by the American Association. Reprinted high-level distortions and random-unrelated shapes (R) Psychological in the top to bottom rows, respectively. Adapted from [18]. [18]. Copyright by the the American American Psychological Psychological Association. Association. Reprinted Reprinted with permission. Adapted from Copyright 2008 2008 by with permission. with permission.

The filled symbols in Figure 3 show how often monkeys made category-endorsement responses The filled symbols in Figure 3 show how often made category-endorsement for polygon shapes at different distortion levels frommonkeys the prototype—that is, how often theyresponses included The filled symbols in Figure 3 show how often monkeys made category-endorsement responses for polygon shapes at different distortion levels from the prototype—that is, how often they included those shapes within the shape family or category they had learned. The performance on prototype for polygon shapes at different distortion levels from the prototype—that is, how often they included those within the shape family or category they had learned. The performance on prototype items shapes was essentially as high as family it couldor be. There seems have been no cap on category on endorsement those shapes within the shape category they to had learned. The performance prototype items essentially as high as it could be. There seems to have been no cap on category endorsement for thewas prototype as exemplar might predict. items was essentially as high astheory it could be. There seems to have been no cap on category endorsement for the prototype as exemplar theory might predict. More telling,asthe E symbols show howpredict. an exemplar-based macaque should have performed in for the prototype exemplar theory might MoreThe telling, the E model symbols show how an exemplar-based macaque should have performed in this task. exemplar is show capped tooan flat. It’s off by 10%macaque much of should the time. In this literature, More telling, the E symbols how exemplar-based have performed in this task. The exemplar model is capped too flat. It’s off by 10% much of the time. In this literature, this model fit isexemplar quite poor. Moreover, Smith al. conducted model-fitting this task. The model is capped tooetflat. It’s off by additional 10% muchsimulations of the time.and In this literature, this model fit istoquite poor.the Moreover, Smith et al. fit conducted additional simulations and model-fitting studies to try improve exemplar model’s [18]. They failed, because the exemplar model is this model fit is quite poor. Moreover, Smith et al. conducted additional simulations and model-fitting studies to try to improve the exemplar model’s fit [18]. They failed, because the exemplar model is up against the inherent geometry of multiple exemplars spread out in psychological space. Macaques up against the inherent geometry of multiple exemplars spread out in psychological space. Macaques simply are not exemplar-based in this task. simply are not exemplar-based in this task.

Behav. Sci. 2016, 6, 12

5 of 24

studies to try to improve the exemplar model’s fit [18]. They failed, because the exemplar model is up against the inherent geometry of multiple exemplars spread out in psychological space. Macaques simply Behav. are not exemplar-based in this task. Sci. 2016, 6, 12 5 of 24

Behav. Sci. 2016, 6, 12

5 of 24

Figure 3. The proportion of times a macaque endorsed into a learned category to-be-categorized items

Figure 3. The proportion of times a macaque endorsed into a learned category to-be-categorized that were outside the category (Rand.), non-typical category members (High-level distortions), typical items that were (Low-level outside the categoryhighly (Rand.), non-typical category members (High-level distortions), members distortions), typical members (V. Low-level distortions), or prototypical typical members members(Prot.). (Low-level distortions), highly typical members (V. Low-level distortions), or Standard exemplar and prototype categorization models fit the macaque’s performance as well as they could (E symbols and P symbols, respectively). The models differ only in prototypical members (Prot.). Standard exemplar and prototype categorization models fit the macaque’s Figure 3. The proportion of times a macaque endorsed areference learned category to-be-categorized their assumption multiple, specific-exemplar cognitive points for categorization, or a performance as well asofthey could (E symbols and P into symbols, respectively). The modelsitems differ only that were outside the category (Rand.), category members (High-level distortions), typical single, central cognitive reference pointnon-typical for categorization. From [18]. Copyright 2008 by the American in their assumption of multiple, specific-exemplar cognitive reference points for categorization, or members (Low-level distortions), highly members (V. Low-level distortions), or prototypical Psychological Association. Reprinted withtypical permission. a single,members central cognitive reference point for categorization. From [18]. Copyright 2008 by the American (Prot.). Standard exemplar and prototype categorization models fit the macaque’s Psychological Association. permission. performance as well as Reprinted they (Ewith symbols and macaque P symbols,should respectively). The This models differwas onlylike in the The P symbols show howcould a prototype-based perform. model their assumption of multiple, specific-exemplar cognitive reference points for categorization, or a that exemplar-based model just described, except now the psychological assumption was made single,were central cognitive reference point forcategorizing categorization. From [18]. Copyright 2008 by the American macaques comparing items needing to a unitary category representation that waswas like The P symbols show how a prototype-based macaque should perform. This model Psychological Reprinted with permission.model fits extremely closely. There is a strong the average of theAssociation. training exemplars. The prototype the exemplar-based model just described, except now the psychological assumption was made that possibility that macaques are prototype-based in this dot-distortion task, though this does not mean macaques items categorizing to ashould unitary category representation The Pcomparing symbols how needing a prototype-based macaque perform. This model was like thethat was theywere always are. Forshow mathematical reasons, the dot-distortion task provides a particularly strong exemplar-based model just described, except now the psychological assumption was made that the average of the training exemplars. The prototype model fits extremely is a strong contrast between prototype and exemplar processing. However, we point closely. out that There the same macaques were comparing items needing categorizing to a unitary studies category representation conclusion has been reached in numerous earlier animal-cognition [35–39]. possibility that macaques are prototype-based in this dot-distortion task, though this that doeswas not mean the average of the training exemplars. The prototype model fits extremely closely. There is a strong they always are. For mathematical reasons, the dot-distortion task provides a particularly strong possibility thatThe macaques are prototype-based 5. Prototypes: Trouble with Exceptions in this dot-distortion task, though this does not mean contrastthey between and exemplar processing. However,task we provides point outa that the same conclusion alwaysprototype are. For mathematical reasons, the dot-distortion particularly strong A prototype-exception task (Figure 4) also lets one explore animals’ commitment to prototypes. has been reached in numerous earlier animal-cognition studies [35–39]. contrast between prototype and exemplar processing. However, we point out that the same Here, the shapes on the left are though the last two do studies not appear to be. These items are conclusion has been reached inCategory numerousAs, earlier animal-cognition [35–39]. the exceptions. The samewith holdsExceptions for the Category B items to the right. 5. Prototypes: The Trouble

5. Prototypes: The Trouble with Exceptions

A prototype-exception task (Figure 4) also lets one explore animals’ commitment to prototypes. A prototype-exception task (Figure 4) also lets one explore animals’ commitment to prototypes. Here, the shapes on the left are Category As, though the last two do not appear to be. These items are Here, the shapes on the left are Category As, though the last two do not appear to be. These items are the exceptions. The same holds for the Category B items to the right. the exceptions. The same holds for the Category B items to the right.

Figure 4. Examples of categories and stimuli used in Smith et al. [40]. The eight-shape groupings are two categories A and B. The six shapes in the top rows of each category are variations on the theme of the category prototype. The bottom row contains exception items that are variations on the theme of the opposing prototype. From [40]. Copyright 2010 by the American Psychological Association. Reprinted with permission. Figure 4. Examples of categories and stimuli used in Smith et al. [40]. The eight-shape groupings are Figure 4. Examples of categories and stimuli used in Smith et al. [40]. The eight-shape groupings are two categories A and B. The six shapes in the top rows of each category are variations on the theme two categories A andprototype. B. The six the top rows of each category are variations the theme of the category Theshapes bottom in row contains exception items that are variations on the on theme of the category prototype. The bottom row contains exception items that are variations on the theme of the opposing prototype. From [40]. Copyright 2010 by the American Psychological Association. Reprinted with permission. of the opposing prototype. From [40]. Copyright 2010 by the American Psychological Association.

Reprinted with permission.

Behav. Sci. 2016, 6, 12 Behav. Sci. 2016, 6, 12

6 of 24

6 of 24

If animals store separate exemplars, they will potentially task. They will If animals justjust store separate exemplars, they will potentially be be OKOK in in thisthis task. They will be be able able to memorize all the Category A and Category B exemplars. They will learn to respond correctly to memorize all the Category A and Category B exemplars. They will learn to respond correctly to to typical items (those clearly in the category) and exception items (those questionably in the category). typical items (those clearly in the category) and exception items (those questionably in the category). However, they will not if they take a prototype-based approach to categorization. Then they will However, they will not if they take a prototype-based approach to categorization. Then they will call the Category A exceptions Bs, as their appearance suggests (and similarly with the Category B call the Category A exceptions Bs, as their appearance suggests (and similarly with the Category B exceptions). If we observe poor performance with criss-cross exceptions like this, we will know that exceptions). If we observe poor performance with criss-cross exceptions like this, we will know that animals wrongly assumed two prototype-averageable groupings. animalsFigure wrongly assumed two prototype-averageable groupings. 5a shows a macaque’s result over 60,000 trials in this task [40]. He was relentlessly awful 5a shows a macaque’s result over trials thisthem task [40]. Hehe was relentlessly awful onFigure exceptions because he criss-crossed them.60,000 He did not in learn though had 15,000 tries at on them. exceptions because he criss-crossed them. He did not learn them though he had 15,000 tries at them. In addition, he received 8,000 correction trials on these items, because we repeated these stimuli In addition, he received 8000 on these weonrepeated these to make sure he received thecorrection experiencetrials of making theitems, correctbecause response them. Still, he stimuli did not to make sure he received theonly experience of making correct response on them. Still, heand didrespond not learn learn though there were four exceptions. Hethe showed very poor ability to remember though there were only four exceptions. He to showed very pooreagles ability and respond to particular exemplars. If this monkey had face exceptional in to hisremember world, he would make a to particular exemplars. If this monkey had to face exceptional eagles in his world, he would make a fatal fatal mistake. Figure 5 gives the results from three macaques in this study. They all showed poor mistake. Figure 5 gives thefor results from of three macaques in this study.aThey all showed poor exception exception performance thousands trials. They seem to expect prototype to work adaptively. They seemfor to default toward prototype And asoprototype they criss-cross theadaptively. exceptions,They because performance thousands of trials. They averaging. seem to expect to work seem they think all stimuli should perceptually fit their prototypes. We think humans are like this, too. to default toward prototype averaging. And so they criss-cross the exceptions, because they think all Humans often flail withfitpoorly-structured, weird tasks that are depend on exemplar processing. stimuli should perceptually their prototypes. We think humans like this, too. Humans often flail Murphy, Blair, and others have alsothat discussed weaknesses in our classification [41–43]. you with poorly-structured, weird tasks dependthese on exemplar processing. Murphy, Blair, andIf others thought to yourself earlier, oh come on, this exceptions task is silly and unfair to the animals, that have also discussed these weaknesses in our classification [41–43]. If you thought to yourself earlier,isoh youon, agreeing with the macaques. come this exceptions task is silly and unfair to the animals, that is you agreeing with the macaques.

Figure 5. (a–c) The performance Murphby by64-trial 64-trialblocks blocksininthe the category Figure 5. (a–c) The performanceofofmonkeys monkeysHank, Hank, Lou, Lou, and and Murph category tasktask of Smith et al. [40]. Curves TT and E,E,respectively, of Smith et al. [40]. Curves and respectively,show showthe theproportion proportionofofcorrect correctresponses responsesmade made to thetosix typical and and twotwo exception items in in each category. 2010by bythe theAmerican American the six typical exception items each category.From From[40]. [40]. Copyright Copyright 2010 Psychological Association. Psychological Association.Reprinted Reprintedwith withpermission. permission.

Behav. Sci. 2016, 6, 12

7 of 24

If animals make a strong prototype assumption in the prototype-exception task, it is a theoretical Sci. 2016, 6,theory. 12 7 of That 24 blowBehav. to exemplar And so here, that theory’s adherents should and can make a rejoinder. is, two of three macaques learned the exceptions by the end—if you train doggedly enough, exemplar If animals make a strong prototype assumption in the prototype-exception task, it is a theoretical processing does finally shine through. blow to exemplar theory. And so here, that theory’s adherents should and can make a rejoinder. That We agree, though we add two First, we you believe that the animal’s assumption is, two of three macaques learned theobservations. exceptions by the end—if train doggedly enough, exemplar as it starts a task is finally a crucial fact about its mind, no matter what it thinks after thousands of trials. processing does shine through. The assumption says whatwethe animal assumes about the world is. animal’s In this case, animals We agree, though add two observations. First, how we believe that the assumption as clearly it a task is a crucial fact about its mind, no matter whattrials it thinks thousands of trials. The madestarts an initial prototype assumption. Second, the late thatafter show exception-item mastery assumptionhave says what the animal assumes about howrelevance. the world is.Natural In this case, animals will clearly madegrant unfortunately no ecological or evolutionary selection never an initial prototype assumption. Second, trials mastery. that show exception-item mastery a monkey 5000 eagle-classification errors on thethe waylate to eagle In nature, your initial approach unfortunately have no ecological or evolutionary relevance. Natural selection will never grant a like will be determinative and decisive, and on failure you do not reach the later phases. Researchers monkey 5,000 eagle-classification errors on the way to eagle mastery. In nature, your initial approach to call a task’s final trials terminal performance, but nature takes a different view. Really, the early will be determinative and decisive, and on failure you do not reach the later phases. Researchers like trials are terminal performance in its grimmer sense, if animals persist in criss-crossing exceptions, as to call a task’s final trials terminal performance, but nature takes a different view. Really, the early they trials clearly aredo. terminal performance in its grimmer sense, if animals persist in criss-crossing exceptions, Thus, participants teach a very important lesson to the human categorization literature. as theyanimal clearly do. Often it will beanimal humans’ initial, intuitive approaches category that reveal the literature. psychological Thus, participants teach a very important to lesson to the tasks human categorization Often will category-learning be humans’ initial, intuitive to category tasks that reveal psychological essence of ittheir system.approaches Often it will be less illuminating to the make them perform essenceofofdifficult, their category-learning system. will be to makecategorization them perform at its thousands unnatural trials. Even Often if theyit learn it,less we illuminating have then studied thousands of difficult, unnatural trials. Even if they learn it, we have then studied categorization at unnatural extreme. its unnatural extreme. Cook and Smith showed that the prototype assumption may extend widely across the Cook and Smith showed that the prototype assumption may extend widely across the vertebrates [44]. In their prototype-exception task (Figure 6), the prototypes had six typical colors. vertebrates [44]. In their prototype-exception task (Figure 6), the prototypes had six typical colors. The typical items in Rows 2–6 had five. The exception items in Row 7 had just one, producing the The typical items in Rows 2–6 had five. The exception items in Row 7 had just one, producing the intuition that theythey should belong totothe Whenwe wesay say “should,” weagreeing are agreeing intuition that should belong theother other category. category. When “should,” nownow we are with with the macaques in Smith et al. [40]. the macaques in Smith et al. [40].

CATEGORY A

CATEGORY B Prototypes

Dimension 1

Dimension 2

Dimension 3

Dimension 4/5

Dimension 6

Exceptions

Figure 6. Illustrating the prototype-exception task used to test pigeons and humans in Cook and Figure 6. Illustrating the prototype-exception task used to test pigeons and humans in Cook and Smith [44]. Each category contained a prototype (Row 1), five typical stimuli (Rows 2–6—five features Smith [44]. Each category contained a prototype (Row 1), five typical stimuli (Rows 2–6—five features shared with their prototype), and an exception (Row 7—five features shared with the opposing shared with their prototype), and an exception (Row 7—five features shared with the opposing prototype). From [44]. Copyright 2006 by the Association for Psychological Science. Reprinted prototype). From [44]. Copyright 2006 by the Association for Psychological Science. Reprinted with permission. with permission.

Behav. Sci. 2016, 6, 12

Behav. Sci. 2016, 6, 12

8 of 24

8 of 24

The black dots in Figure 7 show early learning in this task. Early on, humans and pigeons criss-crossed the exceptions, macaques do. Cook and Smith alsohumans fit formal modelscrissto these The black dots in Figureas7 show early learning in this task. Early on, and pigeons data crossed [44]. The triangle symbols show categorizer should done in the exceptions, as macaques do.what Cook an andexemplar-based Smith also fit formal models to these datahave [44]. The triangle symbols showgradient what anacross exemplar-based should have doneitems in thisis task. The too this task. The typicality exception,categorizer typical, and prototypical flattened typicality gradient across exception, typical, and prototypical items is flattened much. Exemplar much. Exemplar processing is not what humans and pigeons were using attoo this stage. The square processing is not what humans and pigeons were using at this stage. The square symbols show symbols show what prototype processing categorizers should have done. This model fit well.what Possibly processing categorizers should have done. This model fit well. Possibly both species made both prototype species made some kind of a prototype-based assumption at this stage. Thus, something some kind of a prototype-based assumption at this stage. Thus, something like a prototype or like a prototype or abstractionist assumption spans many vertebrate evolutionary lines, giving that abstractionist assumption spans many vertebrate evolutionary lines, giving that assumption assumption evolutionary depth. Wasserman, Kiedinger, and Bhatt found an important converging evolutionary depth. Wasserman, Kiedinger, and Bhatt found an important converging result that resultthey thatgrounded they grounded in especially intuitive and elegant psychological in especially intuitive and elegant psychological models [14]. models [14].

Figure 7. Humans’ and pigeons’ observed and predicted accuracy for prototypes, typical items, and

Figure 7. Humans’ and pigeons’ observed and predicted accuracy for prototypes, typical items, exceptions during the early (top panels) and late (bottom panels) stages of learning within the and exceptions during task the early (top andObserved late (bottom panels) stages of learning within the prototype-exception of Cook andpanels) Smith [44]. performances are shown as unconnected, prototype-exception task of Cook and Smith [44]. Observed performances are shown as unconnected, filled, black circles. The best-fitting predictions of prototype and exemplar models are shown, filled,respectively, black circles. The best-fitting predictions of triangles. prototype and[44]. exemplar models shown, as connected open squares and open From Copyright 2006 are by the respectively, as connected open squares and open triangles. From [44]. Copyright 2006 by the Association for Psychological Science. Reprinted with permission. Association for Psychological Science. Reprinted with permission. Cook and Smith [44] also found that, ultimately, the character of processing changed for both species. Now exemplar memorization finally became successful, exception-item performance Cook and Smith [44] also found that, ultimately, the character of processing changed for improved, the exemplar model fit better, and the abstraction model progressively failed. Smith et al.’s both macaques species. Now exemplar finally became successful, exception-item performance showed this samememorization pattern, suggesting a residual capacity to finally learn exceptions is also improved, theinexemplar model common vertebrates [40]. fit better, and the abstraction model progressively failed. Smith et al.’s

macaques showed this same pattern, suggesting a residual capacity to finally learn exceptions is also 6. Why Prototypes [40]. common in vertebrates Now we should ask why vertebrate species might make a default prototype-based assumption

6. Why Prototypes entering category tasks. To this end, Smith asked how well exemplar and prototype minds would

function the real, world [8]. He considered how a well these prototype-based processes might work for Now weinshould askecological why vertebrate species might make default assumption avoiding eagles, which would be the point of having an eagle category and communicating about entering category tasks. To this end, Smith asked how well exemplar and prototype minds it. would To ask this question was a somewhat radical step in this field. For decades, the literature had the function in the real, ecological world [8]. He considered how well these processes might work for dominant narrative that all categorization is exemplar based. Given that narrative, there would not avoiding eagles, which would be the point of having an eagle category and communicating about it. be different processes in different minds that are differently effective. These possibilities have no To ask question was aofsomewhat radical step in field. Forlesson, decades, the literature place in this a unified field theory exemplar categorization. It isthis an important too, that dominant had the dominant narrative that all categorization is exemplar based. Given that narrative, there would narratives of this kind can have a suppressing, censoring effect. not be different processes in different minds that are differently effective. These possibilities have no place in a unified field theory of exemplar categorization. It is an important lesson, too, that dominant narratives of this kind can have a suppressing, censoring effect.

Behav. Sci. 2016, 6, 12 Behav. Sci. 2016, 6, 12

9 of 24 9 of 24

Smith did extensive simulations [8]. He created thousands of virtual exemplar and prototype simulations who [8]. He created thousands of virtual exemplar prototype monkeys,Smith with did theextensive idea of evaluating categorizes most efficiently. Figure 8and shows one result monkeys, with the idea of evaluating who categorizes most efficiently. Figure 8 shows one result of 84% of many in a series of extensive simulations. Prototype processers—the black dots—averaged many in a series of extensive simulations. Prototype processers—the black dots—averaged 84% correct categorization in Smith’s simulations. Exemplar processors—the clear dots—averaged 80%. in statistically. Smith’s simulations. Exemplar processors—the clearindots—averaged 80%. Thesecorrect meanscategorization were far apart As Smith varied his simulated tasks many ways, prototypes These means were far apart statistically. As Smith varied his simulated tasks in many ways, almost always provided the most efficient strategy. Smith even showed that the prototype effect prototypes almost always provided the most efficient strategy. Smith even showed that the prototype emerges instantaneously, as soon as multiple exemplars have been experienced. That is, even effect emerges instantaneously, as soon as multiple exemplars have been experienced. That is, even usingusing the average of just two exemplars incategorization categorization is better using the average of just two exemplarsas as aa standard standard in is better thanthan using two two exemplars separately as standards. Prototypes very often—not always!—give the clearest signals exemplars separately as standards. Prototypes very often—not always!—give the clearest signals for categorization. for categorization.

Figure 8. The frequency with which the prototype and exemplar process (filled and open circles, Figure 8. The frequency with which the prototype and exemplar process (filled and open circles, respectively) categorized transfer items at different accuracy levels in a family-resemblance task. respectively) categorized transfer items at different accuracy levels in a family-resemblance task. Reprinted from [8]. Copyright 2014 by the Psychonomic Society. Reprinted with permission. Reprinted from [8]. Copyright 2014 by the Psychonomic Society. Reprinted with permission.

Why do prototypes provide the best signals for categorizing? The prototype is closest to all the

Why do prototypes provide the bestyou signals categorizing? Theyou prototype to all the exemplars. If you know it in your mind, knowfor a thing that will help classify ais lotclosest of category members, because in a you sense, is thea least-squares solution to the category problem. exemplars. If you knowthe it prototype, in your mind, know thing that will help you classify a lot of category However, if youthe store exemplars, can be distant from one another, fartoacross psychological members, because prototype, in they a sense, is the least-squares solution the category problem. space, and they may not support accurate categorization as well. On average, prototypes are often However, if you store exemplars, they can be distant from one another, far across psychological space, closer (similar) to exemplars than exemplars are closer (similar) to exemplars. Another point is that and they may not support accurate categorization as well. On average, prototypes are often closer exemplars contain both category-level features and exemplar-idiosyncratic features, so that (similar) to exemplars than exemplars are closer (similar) to exemplars. Another point is that exemplars exemplars are part category signal and part perceptual noise. The noise can deceive, distract, and contain both (though category-level features and exemplar-idiosyncratic sothe thatidiosyncratic exemplars last are part mislead in rare circumstances, when the next exemplar isfeatures, much like category signal and part perceptual noise. The noise can deceive, distract, and mislead (though in rare exemplar, this may support accurate categorization). In contrast, prototypes that are well constructed circumstances, when the next is much like They the idiosyncratic lastand exemplar, support have no idiosyncrasies, butexemplar only category essence. are pure signal they willthis not may deceive. accurate In contrast, that areeither, well constructed no idiosyncrasies, The categorization). advantage they confer is not prototypes rare and specific, but frequent have and general across the but psychological space They of the are category. only category essence. pure signal and they will not deceive. The advantage they confer is With these considerations in mind, on across vertebrate evolution over space 500 million not rare and specific, either, but frequent andreflect general the psychological of the years. category. Prototypes might have given organisms a relentless small edge in recognizing ecologically With these considerations in mind, reflect on vertebrate evolution over 500 important million years. things. It could have meant more food, fewer eagles, better survival, more fitness. Differential fitness Prototypes might have given organisms a relentless small edge in recognizing ecologically important could have shaped our categorization processes by selecting for those with better prototype things. It could have meant more food, fewer eagles, better survival, more fitness. Differential categorization skills. This would explain the prototype effects that many in the categorization fitness could have our categorization processesofbyShepard’s selectingidea for that those with better prototype literature have shaped observed. This could be an illustration cognitive-processing categorization skills. This would explain the prototype effects that in world the categorization literature systems—like physical bodies—are sculpted by the structure of themany natural [45,46]. have observed. This could be an illustration of Shepard’s idea that cognitive-processing systems—like 7. Forcing the Exemplar Assumption physical bodies—are sculpted by the structure of the natural world [45,46]. Perhaps the artificial prototype-exception categories in Smith et al. and Cook and Smith were inherently deceptive to animals, causing them to adopt a prototype-based approach because it seemed intuitively sufficient though it was not [40,44]. Another approach toward Perhaps the artificial prototype-exception categories in Smith et al. andaddressing Cook andanimals’ Smith were

7. Forcing the Exemplar Assumption

inherently deceptive to animals, causing them to adopt a prototype-based approach because it seemed intuitively sufficient though it was not [40,44]. Another approach toward addressing animals’

Behav. Sci. 2016, 6, 12

10 of 24

Behav. Sci. 2016, 6, 12

10 of 24

categorization strengths and weaknesses is to force them to make the exemplar assumption that one categorization strengths and weaknesses is might to forcespontaneously them to makeprefer. the exemplar assumption that one wants to evaluate, not the assumption they wantsThe to evaluate, not the assumption they might spontaneously prefer. exclusive-or (XOR) task suits this purpose (Figure 9). In this task, prototype averaging will Behav. Sci. 2016, 6, 12 10 of 24 The exclusive-or (XOR) taskof suits purpose (Figure this task, prototype averaging will completely fail. The “average” thethis A and B items will 9). be In identical, yielding no distinguishing completely fail. The “average” of the A and B items will be identical, yielding no distinguishing category information. In this exemplar is extremely useful. Really, all the categorization strengths andtask, weaknesses is tomemorization force them to make the exemplar assumption that oneitems category information. In this task, exemplar memorization is has extremely Really, all and the items are exceptional, the task is chaotic. One simply to learnuseful. the four stimuli what wants to evaluate, not theperceptually assumption they might spontaneously prefer. The exclusive-or task suits this purpose (Figure 9). has In this willwhat are exceptional, the task (XOR) is perceptually chaotic. One simply to task, learnprototype the four averaging stimuli and responses they deserve. completely fail. The “average” of the A andAs B items will responses they deserve. Bs be identical, yielding no distinguishing category information. In this task, exemplar memorization is extremely useful. Really, all the items B A are exceptional, the task is perceptually chaotic. One simply has to learn the four stimuli and what responses they deserve. As Bs B

A

A A

B B

Figure (XOR) category categorytask. task.Category CategoryAAmembers membershave haveboth bothred red and blue sectors Figure 9. 9. An An exclusive-or exclusive-or (XOR) and blue sectors or or neither and both yellow and green sectors or neither. Category B members have both blue and neither and both yellow and green sectors or neither. Category B members have both blue and yellow Figure 9. An exclusive-or (XOR) category task. Category A members have both red and blue sectors yellow or neither and both red andsectors green sectors or Single neither. Singledo sectors doblue notcategory predict sectors or neither and redand and green or neither. sectors notboth predict or sectors neither and bothboth yellow green sectors or neither. Category B members have and category membership and categories do not share family resemblance. membership and categories do not share family resemblance. yellow sectors or neither and both red and green sectors or neither. Single sectors do not predict category membership and categories do not share family resemblance.

Smith confirmedthese these well-known facts through simulations [8].task, In this task, Smith confirmed well-known facts through simulations (Figure(Figure 10) [8]. 10) In this prototype Smith confirmed these well-known facts through simulations (Figure 10) [8]. In this task, prototype (black dots) isItuseless. It has to be. Exemplar processing (clear dots) is highly processingprocessing (black dots) is useless. has to be. Exemplar processing (clear dots) is highly effective. prototype processing (black dots) is useless. Itexemplars has to be. Exemplar processing (clear dots) is highly The effective. So there could be times when storing confers a vital performance advantage. So there couldSobethere times when storing exemplars confers aconfers vital performance advantage. The XOR effective. could be times when storing exemplars a vital performance advantage. The task XOR task is a particularly strong test of exemplar processing in categorization, because itnogives is a particularly test of exemplar in processing categorization, because it gives animals way XOR task isstrong a particularly strong testprocessing of exemplar in categorization, because it gives animals no way to fall back to some prototype-based assumption. to fall back to some prototype-based assumption. animals no way to fall back to some prototype-based assumption.

Figure Frequencywith withwhich which the the prototype prototype and process (filled and and openopen circles, Figure 10. 10. Frequency andexemplar exemplar process (filled circles, respectively) categorized transfer items at different accuracy levels in an exclusive-or category Figure 10. Frequency with whichitems the prototype exemplar (filled and category opentask. circles, respectively) categorized transfer at differentand accuracy levelsprocess in an exclusive-or task. Reprinted from [8]. Copyright 2014 by the Psychonomic Society. Reprinted with permission. respectively) categorized transfer items at different accuracy levels in an exclusive-or category task. Reprinted from [8]. Copyright 2014 by the Psychonomic Society. Reprinted with permission. Reprinted from [8]. Copyright 2014 by the Psychonomic Society. Reprinted with permission. Accordingly, Smith, Coutinho, and Couchman gave thousands of XOR trials to macaques [47]. Accordingly, Smith, and They Couchman gave thousands of the XOR trials to macaques They found these tasksCoutinho, very difficult. were only 75% accurate by end of training. They [47]. Accordingly, Smith,allCoutinho, andtheCouchman gave thousands of XORtimeouts. trials to Remember, macaques [47]. trials. During theyaccurate received by 15,000 Theymissed found about these 33% tasksofvery difficult. They experiment, were only 75% the 20-s end of training. They missed They each found these tasks very They were 75% category, accurate and by the end of training. could have been difficult. a the diving member ofthey an only XOR eagle the consequences of anThey about 33%error of all trials. During experiment, received 15,000 20-s timeouts. Remember, each missed about 33% of all trials. During the experiment, they received 15,000 20-s timeouts. Remember, error far more serious than a timeout. Once again, macaques showed a poor capacity to store and error could have been a diving member of an XOR eagle category, and the consequences of an error each respond error could have been a diving member of an XOR eagle category, and the consequences of an differentially to individual exemplars.

error far more serious than a timeout. Once again, macaques showed a poor capacity to store and respond differentially to individual exemplars.

Behav. Sci. 2016, 6, 12

11 of 24

Behav. Sci. 2016, 6, 12

11 of 24

far more serious than a timeout. Once again, macaques showed a poor capacity to store and respond differentially individual exemplars.implications. The world might be full of exceptions, XORness, This XORtofailure has important failure has important implications. The world might be full of exceptions, and This otherXOR poorly-structured categories that defeat prototype processing and demand XORness, exemplar and other poorly-structured categories that defeat prototype processing and demand exemplar memory—for example, the XOR berry task shown in Figure 11. How long would a foraging macaque memory—for example, the XOR task shown in Figure 11. adaptive How longdisaster. would aIfforaging macaque last being only 75% correct on berry this task? This would be an the world were last being only 75% correct on this task?likely This would would have be an evolved adaptivetoward disaster. If the world were exemplar-demanding like that, animals exemplar adeptness, exemplar-demanding like that, animals likely would have evolved toward exemplar adeptness, unless unless exemplar processing is computationally too difficult (which all exemplar theorists would exemplar processing is computationally too difficult (which all exemplar theorists would reject). reject). That they are not exemplar adept suggests the world is mainly not exemplar demanding like That(though they arecertainly not exemplar adeptmay suggests the world is mainly not exemplar like that that the world sometimes be exemplar demanding, anddemanding animals sometimes (though certainly the world may sometimes be exemplar demanding, andthe animals sometimes exemplar exemplar processors, as when they manage the individuals within dominance hierarchy of processors, their group).as when they manage the individuals within the dominance hierarchy of their group). Deadly

Safe

Safe

Deadly

3 Lobes

4 Lobes

Purple

Plum

Figure 11. A knowledge, nature Figure 11. A hypothetical hypothetical ecological ecological exclusive-or exclusive-or (XOR) (XOR) category category task. task. To To our our knowledge, nature almost never presents this kind of foraging or predator-avoidance problem to living organisms, and almost never presents this kind of foraging or predator-avoidance problem to living organisms, and it it is a graceful state of nature that this is so. is a graceful state of nature that this is so.

Instead, suppose the world is mostly prototype-averageable. Then prototype processing might Instead, suppose the world is mostly prototype-averageable. Then prototype processing might be selected for and achieved during cognitive evolution, unless it is computationally too difficult be selected for and achieved during cognitive evolution, unless it is computationally too difficult (which the data disconfirm). Animals would be sensitive to prototypes. They would make default, (which the data disconfirm). Animals would be sensitive to prototypes. They would make default, family-resemblance assumptions. They do. They could safely be poor exemplar processors. They are. family-resemblance assumptions. They do. They could safely be poor exemplar processors. They So might their human cousins be. They are. This idea explains far more of categorization’s natural history. are. So might their human cousins be. They are. This idea explains far more of categorization’s This idea also sharply delimits the explanatory role of exemplar theory, but its adherents have a natural history. rejoinder. The universe of logically-possible categories contains mainly poorly-structured categories. This idea also sharply delimits the explanatory role of exemplar theory, but its adherents have In its overall Venn diagram, prototype categories are just a sliver. So then, how could it possibly be a rejoinder. The universe of logically-possible categories contains mainly poorly-structured categories. that animals are especially adept at this tiny sliver of the possible tasks they might have to confront? In its overall Venn diagram, prototype categories are just a sliver. So then, how could it possibly be Is that not an adaptive disaster and a lousy way to equip an organism for life in the natural world? that animals are especially adept at this tiny sliver of the possible tasks they might have to confront? No, because the logical universe is not relevant to evolution or behavior. Only the natural world Is that not an adaptive disaster and a lousy way to equip an organism for life in the natural world? is relevant. In the natural world, it may well be that prototype categories are dominant, that animals No, because the logical universe is not relevant to evolution or behavior. Only the natural world is are adept at what they need, that their prototype adeptness is not a coincidence and not an relevant. In the natural world, it may well be that prototype categories are dominant, that animals are evolutionary disability. In fact, the natural kinds of the monkey ecology seem to bear this out. Their adept at what they need, that their prototype adeptness is not a coincidence and not an evolutionary world does not give them chaotic XOR berry tasks. It gives eagles, leopards, snakes, Masai, insects, disability. In fact, the natural kinds of the monkey ecology seem to bear this out. Their world does mates, and so forth—all coherent and possibly prototype-averageable things. Drawing on many not give them chaotic XOR berry tasks. It gives eagles, leopards, snakes, Masai, insects, mates, and so psychological and anthropological studies, Malt provided a seminal review of the structure that may forth—all coherent and possibly prototype-averageable things. Drawing on many psychological and be given to humans and animals through natural kinds [48]. anthropological studies, Malt provided a seminal review of the structure that may be given to humans This discussion suggests that there has been an evolutionary tuning between prototypes in mind and animals through natural kinds [48]. and nature. Animals have made us look at what is naturally true, what performances are naturally This discussion suggests that there has been an evolutionary tuning between prototypes in mind optimal, in ways the literature had not considered. In this area, understanding categorization’s and nature. Animals have made us look at what is naturally true, what performances are naturally natural history is crucial, and animals have been great behavioral ambassadors for bringing these optimal, in ways the literature had not considered. In this area, understanding categorization’s natural ecological issues to the fore. Indeed, animals could even be saying tacitly that really vertebrates history is crucial, and animals have been great behavioral ambassadors for bringing these ecological haven’t been doing much exemplar processing for hundreds of millions of years. This is a literatureissues to the fore. Indeed, animals could even be saying tacitly that really vertebrates haven’t been changing insight. In the end, a unitary exemplar account of categorization simply cannot stand. doing much exemplar processing for hundreds of millions of years. This is a literature-changing Nonetheless, exemplar theory could have some role within vertebrates’ overall categorization insight. In the end, a unitary exemplar account of categorization simply cannot stand. capability. There are times—as in managing dominance relations (for monkeys in a troop, for humans in a modern workplace)—when individuals must be encoded individually. There are sparse categories (e.g., the category of Crosby, Stills, Nash, and Young songs) whose members are encoded individually. We have often stated explicitly that the structure of laboratory category tasks can also

Behav. Sci. 2016, 6, 12

12 of 24

Nonetheless, exemplar theory could have some role within vertebrates’ overall categorization capability. There are times—as in managing dominance relations (for monkeys in a troop, for humans in a modern workplace)—when individuals must be encoded individually. There are sparse categories (e.g., the category of Crosby, Stills, Nash, and Young songs) whose members are encoded individually. We have often stated explicitly that the structure of laboratory category tasks can also move participants toward the pole of prototype abstraction or toward the pole of exemplar memorization [49]. It might be a useful rule of thumb for readers to see that category representations are memories. Different category-learning systems might map onto different memory systems (e.g., exemplars onto episodic memories). This is why the systems debate considered now looms so important within our field. 8. Category Rules and the Single-System, Multiple System Debate Now, we turn to the evolutionary history of the third major representational system that may support category learning—category rules. Cognitive scientists and neuroscientists have considered whether humans and animals approach categories with an ability to selectively attend, find a single predictive cue of category membership, form a cognitive rule based on that diagnostic feature, and use it to classify systematically and economically. The idea of a rule-based system for categorization has been deeply controversial in the categorization literature (e.g., [50–52]), and it has provoked a lasting debate about single vs. multiple systems of category learning (i.e., a single exemplar system vs. multiple systems comprising exemplars, rules, and prototypes). Clearly, category rules are completely different from stored exemplars as category representations. Rules are cross-exemplar summaries. Rules can exist absent any exemplar storage and applied absent any exemplar comparisons. So category rules challenge the narrative of a unitary exemplar-based theory as clearly as prototypes do. Moreover, note that rules are abstractions that in some species could even go toward logical propositions, symbolic functioning, and the verbal declaration of category rules. Consequently, there is a lot of theoretical food for thought in the possibility that rules might be a sub-system within animals’ and humans’ overall system for category learning. Considering the prototype-exemplar debate evolutionarily, animal-cognition studies revealed continuity across vertebrate lines. However, considering the rules-exemplar debate we will find, instead, strong evolutionary divergence. These stages may suggest the original character of vertebrates’ category learning. They make us reckon with the psychological meaning of constructs like “rule” and “selective attention.” They suggest the nonhuman primates as a transitional group standing in evolution near the horizon of true category rules. By documenting this evolutionary progression, animals have brought important theoretical insights to this area as well. 9. Rule-Based Categorization—Theoretical Context The category-rule literature draws on a distinction between implicit and explicit processes. The implicit process apprehends stimuli holistically with diffuse attention that encompasses multiple stimulus features. It learns by associating responses to whole stimuli. These associations are behavioral-implicit. They are neither conscious nor verbalizeable. The COVIS model of Ashby and colleagues modeled the implicit process as reliant on associative-learning mechanisms. As described in Section 2, the explicit process emphasizes attention to single stimulus aspects. In humans, it relies on working memory and the executive functions to evaluate hypotheses. It learns through conscious hypothesis testing and its contents are verbalizeable. However, verbalizability may not be a requirement for explicit categorization, as will be discussed below in the context of animal-cognition research. Ashby and his colleagues have clearly acknowledged this. Though ours is not a neuropsychological review, we point out that process distinctions like that between implicit and explicit categorization are grounded in the study of cognitive neuroscience (review in [22]). According to one neuroscience framework, humans’ explicit/analytic processes are mediated in part by a broad neural network that includes the anterior cingulate gyrus, prefrontal cortex, the head of the caudate nucleus, and medial temporal lobe structures that serve declarative

Behav. Sci. 2016, 6, 12

13 of 24

memory. This brain system is related to a broader neural complex serving the executive control of attention [31]. The implicit/nonanalytic processes depend heavily on the striatum and are catalyzed by the reinforcement-mediated strengthening of dopamine-related synapses [19,20]. One can consider exemplar and prototype processing from the same neuroscience perspective. It is known that in humans’ processing in category tasks changes as one increases the number and density of exemplars in categories. For sparse categories, something more like exemplar memorization prevails. Behav. Sci. 2016, 6, 12 of 24 For densely populated categories, something more like prototype abstraction prevails 13[17,41,53]. Antzoulatos and Miller [54]For explored neuroscience of this transition monkeys. found that memorization prevails. densely the populated categories, something moreinlike prototypeThey abstraction early prevails on, when the monkeys could acquire more specific stimulus-response associations, or oneincould [17,41,53]. Antzoulatos and Miller [54] explored the neuroscience of this transition say exemplar-response associations, striatum was the earlier predictor of category responding. monkeys. They found that early on, when the monkeys could acquire more specific stimulus-Later on, asresponse the number of exemplars the task became more a general taskwas of categorization, associations, or oneincreased could sayand exemplar-response associations, striatum the earlier predictor of category cortex responding. on, ascategory the number of exemplars increased and the task the activity in the pre-frontal beganLater to predict responding earlier. This result parallels became more categorization a general task of categorization, activityofinresponses the pre-frontal cortex began to predict idea that implicit represents the linking to whole stimuli—or one might category responding earlier. the idea that implicit categorization represents the say whole exemplars—and it is This also result more parallels dependent on the striatum. linking of responses to whole stimuli—or one might say whole exemplars—and it is also more In our view, categorization science is strengthened greatly by the study of its neuroscience basis dependent on the striatum. and by the application of systems neuroscience to dissociating its different processes and sub-systems. In our view, categorization science is strengthened greatly by the study of its neuroscience basis Clearly, of neuroscience is pointing toward the presence in human and primate brains andthis by application the application of systems neuroscience to dissociating its different processes and subof multiple, dissociable processes within the overall category-learning system. systems. Clearly, this application of neuroscience is pointing toward the presence in human and primate brains of multiple, dissociable processes within the overall category-learning system.

10. Rule-Based Categorization—Methodology

10. Rule-Based Categorization—Methodology The implicit and explicit category learning systems have been differentiated using rule-based implicit and explicit category learningtasks, systems been differentiated rule-based (RB) and The information-integration (II) category as have depicted in Figure 12.using Given the vertical (RB) and information-integration (II) category tasks, as depicted in Figure 12. Given the category boundary in the right panel, X-axis variation—only—distinguishes the twovertical categories. category boundary in the rightthe panel, X-axis variation—only—distinguishes the two The B)). The participant must discover dimensional rule (X < 50 (Category A); X >categories. 50 (Category participant must discover the dimensional rule (X < 50 (Category A); X > 50 (Category B)). Of course Of course this whole stimulus space is not shown. Rather, the rule must be discovered based this whole stimulus space is not shown. Rather, the rule must be discovered based on feedback from on feedback from successive trials. The informative dimension should be attended sharply; the successive trials. The informative dimension should be attended sharply; the non-informative non-informative dimension The panel shows anGiven II categorization task.category Given this dimension ignored. The leftignored. panel shows anleft II categorization task. this major-diagonal major-diagonal category boundary, both dimensions partially informative. must integrate boundary, both dimensions are partially informative.are Subjects must integrate theSubjects information offered the information offered by both to produce a correct categorization response (again based on feedback by both to produce a correct categorization response (again based on feedback from successive trials). from Neither successive trials). Neither dimension can be ignored. dimension can be ignored.

Figure 12. 12.(a) An information-integration information-integration category structure, depicted within 100 an× abstract Figure (a) An category structure, depicted within an abstract 100 100 ˆstimulus 100 stimulus The open-circle and plus-sign respectively, indicateACategory space.space. The open-circle and plus-sign symbols, symbols, respectively, indicate Category and B stimuli; andand (b) a(b) rule-based category structure, depicteddepicted in the same way.same way. A andCategory Category B stimuli; a rule-based category structure, in the

RB and II tasks are matched for important aspects of category structure, including between-

RB and II tasks are matched for important aspects of category structure, including category exemplar separation and within-category exemplar cohesion. Consequently, they are also between-category exemplar separation and within-category exemplar cohesion. Consequently, they matched for the a priori perceptual difficulty of the categorization problem. In fact, the two category are also matched the a priori perceptual of the categorization problem. In fact, structures arefor essentially identical, except difficulty for a 45-degree rotation through stimulus space. If thethe X two and Y dimensions were truly integral (e.g., the color attributes of brightness and saturation, in which attentional processes cannot selectively target one dimension [55]), the tasks would be identical for learners, because there would be no privileged dimensional axes and neither category task would be

Behav. Sci. 2016, 6, 12

14 of 24

category structures are essentially identical, except for a 45-degree rotation through stimulus space. IfBehav. the XSci.and Y 6,dimensions were truly integral (e.g., the color attributes of brightness and saturation, 2016, 12 14 ofin 24 which attentional processes cannot selectively target one dimension [55]), the tasks would be identical rule-based. Thus, forthere many reasons, theprivileged RB and IIdimensional tasks serve as strong mutualcategory controls.task Theywould differ for learners, because would be no axes and neither only in the potential value of reasons, selectivethe attention rule serve formation in one task but not the other. be rule-based. Thus, for many RB andand II tasks as strong mutual controls. They differ Inthe fact, this is why the dissociation a comparative contribution. only in potential value ofRB-II selective attentioncan andmake rule formation in one task but notBy therotating other. the dimensional axis of category tasks, from II can to RB, onea comparative can ask whether the cognitive systemsthe of In fact, this is why the RB-II dissociation make contribution. By rotating different species are dimensionally aligned or polarized. If they are, then the dimensional RB task dimensional axis of category tasks, from II to RB, one can ask whether the cognitive systems of different may support robust andaligned rapid learning, just If asthey a polarizing filter will strongly admit light when species are dimensionally or polarized. are, then the dimensional RB task may support alignedand to rapid the light’s polarization If the filter cognitive system of a species is not aligned dimensionally robust learning, just as a axis. polarizing will strongly admit light when to the polarized, both II and RBIftasks will be learned to of theasame level at the same speed—because task’s light’s polarization axis. the cognitive system species is not dimensionally polarized,the both II dimensions in aatsense. This speed—because evaluation can be regarding any and RB taskswill willbe beintegral learned to to the the species same level the same themade task’s dimensions species that cantoparticipate tasks, making RB-II framework valuable will be integral the speciesininoperant-discrimination a sense. This evaluation can be made the regarding any speciesa that can comparative assay. participate in operant-discrimination tasks, making the RB-II framework a valuable comparative assay. Forexample, example, Figure 13 shows gave humans RBRB andand II tasks [56].[56]. The For shows the theresult resultwhen whenSmith Smithetetal.al. gave humans II tasks circle stimuli varied in the width and tilt of internal bars. Participants were told that they should The circle stimuli varied in the width and tilt of internal bars. Participants were they should decidewhether whethereach eachstriped stripedcircle circlebelonged belongedtotoCategory CategoryAAor orB.B.They Theyreceived receivedthe theRB RBand andIIIItasks tasksin in decide counterbalancedorder order(RB-II (RB-IIor orII-RB). II-RB).Humans Humansclearly clearlyshowed showedthe thecrucial crucialresults resultsofoflearning learningthe theRB RB counterbalanced taskwith with distinctive sensitivity and producing strong RB performance task distinctive easeease and and sensitivity and producing a strongaRB performance advantage.advantage. Extensive Extensive by Ashby,and Maddox, and their have colleagues havethis produced this result times research byresearch Ashby, Maddox, their colleagues produced result many timesmany (e.g., [57]). (e.g., [57]). Moreover, Maddox, and Ashby have described many related implicit-explicit dissociative Moreover, Maddox, and Ashby have described many related implicit-explicit dissociative phenomena in humans’ category learning Humans use explicit-reasoning and processes hypothesisinphenomena humans’ category learning [21]. Humans use[21]. explicit-reasoning and hypothesis-testing to testing this processes to achieve this performance advantage. Their areand verbal, and conscious, achieve performance advantage. Their solutions are verbal, andsolutions conscious, reportable to others. and reportable to others. Humans powerful instantiateand a particularly powerful and agile form ofThere category-rule Humans instantiate a particularly agile form of category-rule learning. may be learning. There be brain well organized for the ruleand hypothesis-testing processes of brain circuits wellmay organized forcircuits the ruleand hypothesis-testing processes of RB category learning. But RB category learning. thenItwhich species have them? Ithelps mayhumans also be that helps humans then which species haveBut them? may also be that language holdlanguage rule-based hypotheses holdapply rule-based hypotheses and apply them. But thenare what kindswithout of rulelanguage? processes From are possible and them. But then what kinds of rule processes possible either without language? From either perspective, the evolutionary emergence of rule-based category perspective, the evolutionary emergence of rule-based category learning is intriguing and deserving learning of study. is intriguing and deserving of study.

Figure 13. (a) Proportion of correct responses in each 10-trial block for 30 humans who performed Figure 13. (a) Proportion of correct responses in each 10-trial block for 30 humans who performed 600 600 trials of a rule-based (RB) and an information-integration (II) category task, in that order; trials of a rule-based (RB) and an information-integration (II) category task, in that order; (b) (b) Proportion of correct responses in each 10-trial block for 30 humans who performed 600 trials of Proportion of correct responses in each 10-trial block for 30 humans who performed 600 trials of an II an II and RB category task, in that order. Reprinted from [58]. Copyright 2012 by Elsevier. Reprinted and RB category task, in that order. Reprinted from [58]. Copyright 2012 by Elsevier. Reprinted with permission. with permission.

11. Rule-Based Categorization—Comparative Results 11.1. Pigeons Smith et al. evaluated pigeons’ RB-II performance in a normal operant-discrimination situation [59]. The stimuli were circular sine-wave gratings varying in bar spatial frequency (width) and

Behav. Sci. 2016, 6, 12

15 of 24

11. Rule-Based Categorization—Comparative Results 11.1. Pigeons Smith et al. evaluated pigeons’ RB-II performance in a normal operant-discrimination situation The stimuli were circular sine-wave gratings varying in bar spatial frequency Behav.[59]. Sci. 2016, 6, 12 15 of(width) 24 and orientation (tilt). Figure 14a shows performance by session from the onset of learning for eight orientation Figure 14a shows performance session from the(learning onset of learning forforward eight RBfrom RB and eight II (tilt). learning pigeons. These forwardby learning curves graphed and eight II learning pigeons. These forward learning curves (learning graphed forward from the the beginning of acquisition) show no hint of any RB performance advantage. However, pigeons acquisition) hint of any performance advantage. However, pigeonsbirds wereto be were beginning removed of from the task show uponno reaching the RB criterion, leaving the weaker performing removed from the task upon reaching the criterion, leaving the weaker performing birds to be graphed subsequently. This adds waviness to the acquisition curves. Accordingly, Figure 14b shows graphed subsequently. This adds waviness to the acquisition curves. Accordingly, Figure 14b shows performance by sessions backward from the criterion block. In these backward learning curves, the performance by sessions backward from the criterion block. In these backward learning curves, the birds’birds’ approach to criterion is is aligned, is no no evidence evidenceofofanan advantage. Pigeons approach to criterion aligned,but butstill still there there is RBRB advantage. Pigeons showed no difference in the speeds of learning or the final performance levels in the two tasks. To the showed no difference in the speeds of learning or the final performance levels in the two tasks. To contrary, these graphs seem to reflect that pigeons were learning two category tasks of the the contrary, these graphs seem to reflect that pigeons were learning two category tasks of the samesame character and difficulty. Pigeons showed no preference forfor one-dimensional character and difficulty. Pigeons showed no preference one-dimensionaltask tasksolutions solutionsor orfor fortasks tasks something affording something rule formation. affording like rule like formation.

Figure 14. Pigeons performing information-integration (II)and andrule-based rule-based (RB) tasks. (a) Proportion Figure 14. Pigeons performing information-integration (II) (RB) tasks. (a) Proportion of correct responses in each session fromthe theonset onset of learning forfor eight II-learning pigeons of correct responses in each session from learningforward forward eight II-learning pigeons (open-triangle symbols) and eight RB-learning pigeons (filled-square symbols); proportion of (open-triangle symbols) and eight RB-learning pigeons (filled-square symbols); (b)(b) proportion of correct correct responses in each session from the criterial block backward for eight II-learning pigeons and responses in each session from the criterial block backward for eight II-learning pigeons and eight eight RB-learning Reprinted from Copyright [58]. Copyright 2012 Elsevier.Reprinted Reprinted with RB-learning pigeons. pigeons. Reprinted from [58]. 2012 byby Elsevier. withpermission. permission.

This result points to a notable weakness in pigeons’ selective-attention capacity. It is as though

This result points of to this a notable weakness in pigeons’ selective-attention capacity.toward It is asone though within the context task they could not direct attentional resources selectively within the context this taskresult theycomplements could not direct selectively toward dimension. Thisof surprising that ofattentional Pearce et al.resources [60], who also demonstrated a one striking This failuresurprising of selectiveresult attention by pigeons, that leading them toetquestion pigeons even dimension. complements of Pearce al. [60],whether who also demonstrated possess central processes. a striking failure ofattention-allocation selective attention by pigeons, leading them to question whether pigeons even This result also suggests that the psychological privilege accorded dimensional analysis and onepossess central attention-allocation processes. dimensional category rules is not a vertebrate-wide cognitive adaptation. Pigeons’ category learning This result also suggests that the psychological privilege accorded dimensional analysis and might illuminate a phylogenetically ancient associative categorization system that is still one-dimensional category rules is not a vertebrate-wide cognitive adaptation. Pigeons’ category predominant in some vertebrate lines. We could be seeing the character of the stem vertebrate learning might illuminate a before phylogenetically ancientand associative categorization system that it. is still category-learning systems dimensional analysis category rules evolutionarily augmented predominant in some vertebrate lines. We could be seeing the character of the stem vertebrate Pigeons’ indifference to the task’s rotation in stimulus space is like humans’ indifference to task category-learning systems before dimensional analysis category rules evolutionarily it. rotations within non-separable stimulus spaces thatand defeat the analytic processes of augmented selective Pigeons’ indifference toduring the task’s rotation in stimulus space is liketreat humans’ attention [61,62]. Possibly category learning in these tasks, pigeons stimuli indifference as integral to wholes, not applying attentional, dimensional, or rulethat selection/focus. patternsofcould task rotations within non-separable stimulus spaces defeat the Pigeons’ analyticdata processes selective reflect that they bring the same implicit category system to bear on both RB and II category tasks. That system would not be affected by whether the category task is dimensionally aligned or not, and learning would proceed equivalently either way. Pigeons’ performance is consistent with a cognitive

Behav. Sci. 2016, 6, 12

16 of 24

attention [61,62]. Possibly during category learning in these tasks, pigeons treat stimuli as integral wholes, not applying attentional, dimensional, or rule selection/focus. Pigeons’ data patterns could reflect that they bring the same implicit category system to bear on both RB and II category tasks. That system would not be affected by whether the category task is dimensionally aligned or not, and learning would proceed equivalently either way. Pigeons’ performance is consistent with a cognitive organization by which they gradually associate behavioral responses to regions of perceptual space or to unanalyzed stimulus wholes, while withholding (or lacking the capacity for) stimulus analysis, selective attention, and rule formation. This interpretation recalls Pearce’s configural theory [12]. There might be advantages to having a unitary category-learning system based on responding to configural wholes. It would reduce strategy competition from multiple processes competing to solve tasks. One would need no referee or metacognitive agency to adjudicate this competition. One could avoid the superstitious rules that humans sometimes “discover” when their rule system prevails wrongly. A configural system might also be especially adept at learning complex or non-linear category decision boundaries that would defeat a dimensionally aligned, rule-based system. Moreover, many natural kinds might well be organized essentially as high-dimensional II tasks (recalling Rosch’s probabilistic feature theory) and thus well suited to holistic (configural) responding [32,63]. 11.2. Macaques (Macaca Mulatta) Smith et al. tested rhesus macaques (an Old-World non-human primate) on RB-II category tasks [56]. The macaques performed these tasks in counterbalanced order (RB-II; II-RB). Smith et al. found similar results in a companion experiment with capuchin monkeys (Cebus apella—a New World primate) [64]. These monkeys at the Language Research Center (LRC) had been trained to respond to computer-graphic stimuli by manipulating a joystick that controlled a cursor. The LRC pioneered the exploration of primates’ cognitive systems using computer-based stimulus presentation and joystick-controlled behavioral responses by the monkeys. This work originated and continues with the express goal that these cognitive-testing techniques can enrich animals’ lives cognitively by providing forms of cognitive stimulation like that humans deliberately seek out (as when your child seeks out Minecraft!). Macaques’ capacity for explicit cognition is not a salient feature of their cognition. They have relatively smaller frontal cortices than humans [65], producing performance decrements on tasks that foster response competition or need response inhibition [66,67]. Therefore, one reasonable hypothesis was that they might have a minimal rule-based category-learning system [12]. In that case, RB and II tasks should be learned about equally well and fast, because the tasks are matched in category structure and inherent difficulty unless the organism can apply dimensional attention or rule-based categorization selectively to the RB task. Macaques’ proportion correct by block (100 trials per block) is shown in Figure 15. Macaques learned the RB task faster and to a higher final performance level than the II task. In the RB-II task order (Figure 15a), macaques improved from a very low RB performance level (as they initially acclimated to the task and stimuli) to a very high level later on. In the II-RB task order (Figure 15b) they were sharp and sensitive RB categorizers nearly from the task’s outset, having earlier accomplished their acclimation. Strikingly, taking only the first 600 trials of their second task, they were 90% and 67% correct on the RB and II task, respectively. Dimensional analysis clearly had privilege within their category-learning system. Despite lesser frontal-cortical brain development and their lack of language, macaques have similarities to humans in their RB-II performance. These monkey results suggest answers to several of the comparative questions raised earlier. It is not the case that language is a necessary condition for the psychological privilege toward rules or dimensional analysis to emerge. Early multiple-system theories tied explicit categorization and category rules to verbal hypotheses [19,27]. The monkey results cut this tie. Dimensional hypotheses or criteria for categorization can apparently be represented successfully in mind as non-verbal proto-rules or cortical loops of sustained dimensional-focusing activity.

Behav. Sci. 2016, 6, 12 Behav. Sci. 2016, 6, 12

17 of 24 17 of 24

Figure 15. (a) Proportion of correct responses in each 100-trial block for three monkeys who performed Figure 15. (a) Proportion of correct responses in each 100-trial block for three monkeys who 6000 trials of a rule-based (RB) and an information-integration (II) category task, in that order; and performed 6,000 trials of a rule-based (RB) and an information-integration (II) category task, in that (b) proportion of correct responses in each 100-trial block for three monkeys who performed 6000 trials order; and (b) proportion of correct responses in each 100-trial block for three monkeys who of an II and an RB category task, in that order. Reprinted from [58]. Copyright 2012 by Elsevier. performed 6,000 trials of an II and an RB category task, in that order. Reprinted from [58]. Copyright Reprinted with permission. 2012 by Elsevier. Reprinted with permission.

The analysis in in categorization categorizationisisnot nota Theresults resultsalso alsoshow show that that the the privilege privilege toward toward dimensional dimensional analysis ahumanly humanlyunique uniquething. thing.It It extends beyond humans, at least to other species in the order Primates. extends beyond humans, at least to other species in the order Primates. This This implies rule-based categorization had an earlier evolutionary origin than with humans. implies that that rule-based categorization had an earlier evolutionary origin than with humans. It raises Itthe raises the additional questions of that whendimensionally that dimensionally analytic system of categorization emerged, additional questions of when analytic system of categorization emerged, why why it was adaptive, and what the most primitive steps toward category rules may have been. it was adaptive, and what the most primitive steps toward category rules may have been. 12. 12. Category Category Rules Rules and and Representational Representational Portability Portability Zakrzewski Zakrzewskiand andher hercolleagues colleagueshave havebegun begunresearch researchtotoexplore explorethese theseissues issues(e.g., (e.g.,[68]). [68]).She Sheasks asks whether RB and II category knowledge are qualitatively different in cognitive content, and whether whether RB and II category knowledge are qualitatively different in cognitive content, and whether they different in the way for humans and macaques. If so, thisIfwould strengthen the homology theyare are different in same the same way for humans and macaques. so, this would strengthen the between humans’ and monkeys’ RB category learning. In particular, we have tested the idea homology between humans’ and monkeys’ RB category learning. In particular, we have testedthat the the generalization of II and of RBII category knowledge shouldshould differ substantially—for humans and idea that the generalization and RB category knowledge differ substantially—for humans primates—for theoretical reasons described now. now. and primates—for theoretical reasons described Correct responding in II tasks is believed Correct responding in II tasks is believedtotobe begoverned governedby byimmediate immediatereinforcement reinforcementthat thatlinks links correct responses to training stimuli. Therefore, II category learning should be yoked to those original correct responses to training stimuli. Therefore, II category learning should be yoked to those original training trainingstimuli, stimuli,and andititshould shouldgeneralize generalizetotonew newstimuli stimuliwith withconstraints constraintsdetermined determinedby byinterstimulus interstimulus distances and dissimilarities. Moreover, II learning is implicit. So there will be no freestanding category distances and dissimilarities. Moreover, II learning is implicit. So there will be no freestanding knowledge resident inresident working build ato bridge correctto generalization for novelfor stimuli. category knowledge inmemory workingto memory buildto a bridge correct generalization novel For this reason, too, II category knowledge should have limited generalizability to novel stimuli. stimuli. For this reason, too, II category knowledge should have limited generalizability to novel stimuli. However, However,correct correctresponding respondingininRB RBtasks—for tasks—forhumans—is humans—isthought thoughttotodepend dependon onan anorganizing organizing rule resident in working memory (this answer for the non-human primates is not known but it rule resident in working memory (this answer for the non-human primates is not known but it should should be discoverable). This rule could be more abstract than II category learning is. It could be discoverable). This rule could be more abstract than II category learning is. It could have have substantial stimulus independence. A color rule, instance,might might survive survive changes changes in substantial stimulus independence. A color rule, forforinstance, in other other perceptual features (shape, size, etc.). According to this theoretical account, RB category learning perceptual features (shape, size, etc.). According to this theoretical account, RB category learning should shouldbe bemore moregeneralizable. generalizable.Casale Casaleetetal. al.demonstrated demonstratedthis thisgeneralizability generalizability[69]. [69]. We Wetook tookaastep step farther to demonstrate this generalizability in humans and macaques. farther to demonstrate this generalizability in humans and macaques. Figure our experimental situation [68].[68]. We broke the category structures generally used Figure1616shows shows our experimental situation We broke the category structures generally inused RB-II research (Figure 12) into training and transfer distributions. In this way, we could train in RB-II research (Figure 12) into training and transfer distributions. In this way, we couldwith train the category stimuli in the lower part of the stimulus space, and then suddenly test generalization by with the category stimuli in the lower part of the stimulus space, and then suddenly test

generalization by presenting stimuli from the untrained part of the space. In this experiment, we

Behav. Sci. 2016, 6, 12

18 of 24

Behav. Sci. 2016, 6, 12

18 of 24

presenting stimuli from the untrained part of the space. In this experiment, we maintained trial-by-trial maintained trial-by-trial reinforcement after each trial, so that we are measuring transfer with reinforcement after each trial, so that we are measuring transfer with ongoing learning. ongoing learning.

Figure 16. (a) A training-generalization information-integration category structure, depicted in the Figure 16. (a) A training-generalization information-integration category structure, depicted in the same way as Figure 12. The bottom pair and top pair of stimulus ellipses, respectively, were the defined same way as Figure 12. The bottom pair and top pair of stimulus ellipses, respectively, were the categories during training and generalization; and (b) a training-generalization rule-based category defined categories during training and generalization; and (b) a training-generalization rule-based structure, depicted in the same way. category structure, depicted in the same way.

The monkeys monkeys transferred RB learning seamlessly. seamlessly. They They stayed stayed at very very high high performance performance levels through the transition, transition, so they generalized generalized with no difficulty. difficulty. This is consistent with the idea that they were were applying applying some some form form of of category category knowledge knowledge that was was more more abstract abstract with with more more stimulus stimulus independence, something beyond an associative response to the original training stimuli. In contrast, independence, something beyond an associative response their II learning faltered, faltered, requiring requiring additional additionaltrials trialstotorecover. recover.This Thisis is consistent with some form consistent with some form of of category knowledge that lessindependent independentofoftraining trainingstimuli stimuliand andless less prepared prepared to to generalize. category knowledge that is isless Humans’ result was strikingly similar, with flawless flawless RB transfer transfer but faltering II transfer. In this one respect, respect, it it does does seem seem that that RB RB learning learning is is similarly similarly distinctive distinctive for for humans humans and and macaques, macaques, in the sense sense that it was immediately portable to new and untrained stimuli. it was immediately portable to new and untrained stimuli. However, immediately generalized, it did so so in However,remember rememberthat thatininthis thiscase, case,though thoughRB RBknowledge knowledge immediately generalized, it did the context of always being propped up byup immediate trial-by-trial reinforcement. Accordingly, we are in the context of always being propped by immediate trial-by-trial reinforcement. Accordingly, asking what happens in RB and II generalization if we knock the props. we are now asking now what happens in RB and II generalization if weaway knock away the props. This experiment experimentisisconducted conductedinin three phases. First, we train participants onCategory the Category three phases. First, we train participants on the A andA B and B stimulus distributions in one partperceptual of the perceptual space. After theycriterion, reached we criterion, we stimulus distributions in one part of the space. After they reached gradually gradually wean them away from trial-by-trial reinforcement, so that they were performing wean them away from trial-by-trial reinforcement, so that now theynow were performing blocks ofblocks trials of trials no reinforcement and receiving their summary atof the endtrial of each trial block. with no with reinforcement and receiving their summary feedbackfeedback at the end each block. This is a This is a technique called deferred-rearranged reinforcement that we borrowed from earlier animal technique called deferred-rearranged reinforcement that we borrowed from earlier animal research research (e.g., andresearch human [49]. research [49]. Participants with noAfter feedback. (e.g., [70]), and[70]), human Participants completedcompleted trial blockstrial withblocks no feedback. each After they received the reinforcements fromtrials all correct trials and then together block,each theyblock, received together thetogether reinforcements from all correct and then together the timeouts the timeouts allThrough error trials. thiswe technique, ablefamiliar to present familiar stimuli on from all errorfrom trials. thisThrough technique, were ablewe towere present stimuli on most trials, most trials,existing sustaining existing category knowledge the task reinforcing enough so that sustaining category knowledge and makingand the making task reinforcing enough so that everybody everybody on task. Under these conditions, we also couldsalt alsoinsalt in transfer trials—from stayed on stayed task. Under these conditions, though,though, we could transfer trials—from the the untrained region of category the category space—with no announcement or separation the ongoing untrained region of the space—with no announcement or separation fromfrom the ongoing task, task, soexisting that existing category knowledge just transfer spontaneously—if In this so that category knowledge wouldwould just transfer spontaneously—if it could.itIncould. this way, we way, we also avoided the shock of suddenly moving participants to a testing phase with all novel also avoided the shock of suddenly moving participants to a testing phase with all novel stimuli. We stimuli. this complementary approach was worthwhile uswhether evaluate the whether the thoughtWe thisthought complementary approach was worthwhile for lettingfor us letting evaluate category category knowledge that derives from original RB and II training spontaneously,inherently inherently different knowledge that derives from original RB and II training is is spontaneously, in generalizability. generalizability. The answer for humans is clear-cut [71]. In RB tasks, humans under deferred-rearranged feedback are completely able to transfer their rule knowledge to the occasional novel, untrained stimuli that are presented. Their II category knowledge, whatever form it exists, does not show this

Behav. Sci. 2016, 6, 12

19 of 24

The answer for humans is clear-cut [71]. In RB tasks, humans under deferred-rearranged feedback are completely able to transfer their rule knowledge to the occasional novel, untrained stimuli that are presented. Their II category knowledge, whatever form it exists, does not show this feature of portability at all. Perhaps this is because the category knowledge—the category learning—is welded associatively to the training stimuli. In the II case there is no verbal or conscious or symbolic representational material freestanding that can be applied to new stimuli. Now Zakrzewski and her colleagues are testing macaques in this paradigm [72]. Preliminary results show that monkeys’ II learning—just as humans’ II learning—is unable to transfer to untrained stimuli while under deferred reinforcement. However, the crucial test is whether monkeys can transfer RB knowledge. Preliminary results suggest that this knowledge is substantially less portable than humans’ RB knowledge is. Their transfer weakness bears on the extent to which primates’ rule-based performances fully involve something like category rules. Non-human primates may share some aspects of humans’ rule-based categorization, but not all. This result is potentially important to multiple-systems theory for suggesting that there are primitive and partial forms of rule use in categorization. There is a lot to uncover, now, in order to understand why monkeys’ failure in RB transfer occurred. Perhaps it is because of a smaller working memory capacity. Perhaps it is because of a lack of language. Some of these reasons may be uncovered by work in developmental research in this area of RB category learning. Monkeys’ provisional failure and humans’ definite success in transferring RB knowledge while under deferred reinforcement show how distinctive humans’ ability really is. Humans clearly have a distinctive explicit system for rule-based categorization that monkeys are only on the way towards. 13. General Discussion We described three theoretical perspectives on categorization (Section 2). From these arose the two theoretical debates that have dominated our literature: the debate on exemplars vs. prototypes (Sections 3–7) and the debate on single category-learning systems vs. multiple systems (Sections 8–12). These theories and debates provide a good grounding in the categorization literature. Then, we explored the contributions of recent animal-cognition research to this literature to reach these conclusions. Consider the exemplar-prototype debate first. Animals are poor at encoding specific exemplars that deserve particular responses. The disability is so serious it could be fatal to animals encountering exemplar-forcing natural categories. Therefore, it is unlikely that exemplar processing is the unitary system through which animals categorize. Assuming evolutionary continuity, the same may be true for humans. In fact, humans show converging performance failures when exemplar processing is forced on them (e.g., [5,73,74]). Murphy discussed these failures pointedly, reminding researchers to carefully attend to them and to seek to understand ecologically-valid categorization [42,43]. Doing so can give insight into when and why exemplars or prototypes are used in categorization. (2) Animals commonly pre-assume that categories will be organized by family resemblance and, possibly, prototype-based. Observing this pre-assumption, one must consider why it is appropriate or adaptive in the natural world. Animal studies encourage us to reach that understanding (Sections 6 and 7) [8]. (3) Given animals’ exemplar weakness and prototype strength, one is led to ask whether animals’ systems for categorization reflect the structure of natural categories. Natural categories were the proving ground for categorization and could have provided a shaping evolutionary force. Anderson [75], Briscoe and Feldman [2], Feldman [76], Shepard [45,46], and Smith [8] have pointed to the possibility of these shaping or tuning forces, but the perspective of animal cognition brings them back to life. If family-resemblance categories were common in nature [32,48,77–79], then cognitive evolution would sensibly have drifted toward the family-resemblance pre-assumption.

(1)

Behav. Sci. 2016, 6, 12

(4)

(5)

(6)

(7)

(8)

(9)

(10)

20 of 24

But if the ecology were different and exemplar-forcing, the drift of cognitive evolution might have been different, and animals might have achieved a different tradeoff between exemplar-specific and abstractive-general classifications. Thus, in this area, animals have made a very important theoretical statement about categorization. The hope that there can be a pan-exemplar description of categorization is essentially gone. In this conclusion, animal-cognition studies endorse converging findings from many cognitive, neuroscience, and neuropsychological studies showing that exemplar processing is insufficient to fully describe human categorization [1,21]. Animal-cognition research even reaches out to help refine our human theories. We stress that by this conclusion we are not rejecting exemplar theory, or exemplar processes, as having a role in animals’ and humans’ overall categorization system. We are rejecting, based on everything about categorization’s natural history, that there can be a unitary, comprehensive exemplar theory. Animals have also made important contributions to the debate about single and multiple categorization systems. Pigeons’ equivalent performance in RB and II tasks is a reminder that some organisms, sometimes, may perceive the world integrally (in Garner’s sense) [55], not separably, and holistically, not analytically. If so, then RB and II tasks will be identical in learning and performance as pigeons show. One does not have to selectively attend. One does not have to analyze and narrow in on one dimension. One can just make the correct response to whole stimuli after those responses have been associatively strengthened through reinforcement. Attention is optional in a significant, unappreciated sense. Moreover, pigeons suggest that in the stem vertebrate categorization system selective attention may have played a lesser role, though we would not deny pigeons’ some form of selective attention in some situations. In fact, it is known that pigeons can, over multiple training sessions, gradually change the focus of something like attention in the service of improving performance within operant discriminations (e.g., [80]). This shows that the idea of selective attention is a nuanced construct, for pigeons gradually shift the axis of their discriminative learning over sessions, but they seem not to recruit dimensional attention in an RB task so as to gain a performance advantage there. Monkeys, in contrast, do recruit some attentional or rule process that grants them a meaningful RB advantage. Why? Zakrzewski’s RB-II research shows that the notion of a category rule is also a nuanced construct that deserves more theoretical attention. Rules for humans bring many cognitive affordances, including the important advantage of making category knowledge freestanding and easily portable to novel, untrained stimuli. Our preliminary results show that monkeys share that affordance only partially, and there may be other affordances of rules that they do not share. The full theoretical description of category rules needs to include an understanding of the degrees of separation between monkeys’ rule-like processes and humans’ rule-based processes, and consider the possible evolutionary pathways that may connect them. In the end, comparative RB-II research carries the headline of evolutionary emergence and discontinuity, from pigeons’ integral categorization to humans showing robust, explicit, declarative, language-based rule systems. Even here, animals illuminate the theoretical issues of human categorization by showing how profound the cognitive changes have been in the evolution of category rules. Clearly, the hope for a pan-single system category theory is lost to the human literature as well. Humans do have a profoundly sophisticated rule system for category learning, one that is apparently served by different brain processes and neural circuits, and one very different from that expressed in other vertebrate lines. It is time for our whole field to fully respect the contributions of cognitive neuroscientists who have demonstrated this system fully and elegantly. Perhaps it was inevitable that categorization—an essential backbone of animal cognition—would turn out to be important enough to have multiple cognitive systems supporting it, so that even the same organism can bring different processes to bear on different category tasks. We may

Behav. Sci. 2016, 6, 12

21 of 24

capture the richness of our field, and the richness of this capacity, more fully if we acknowledge this possibility. (11) Animal-cognition research has also cast light on the different high-level choices that theorists can make approaching a cognitive capacity. They may try to enforce a structural parsimony and unitarianism onto the science from the top down, or they can hold back and accept the psychological processes and systems they glean when research occurs from the bottom up. In the area of comparative categorization, as in the area of the cognitive neuroscience of categorization, the latter approach has proven more successful. Minds just may not be parsimonious, no matter our wishes in the matter. Thus, animal-cognition research has made important contributions to our understanding of human and animal categorization, by prompting the field to give categorization an evolutionary depth, and phylogenetic breadth. In this way we have been led to consider the state of nature, the survival needs of animals, evolutionary tuning forces, adaptive approaches, deep continuities across vertebrate lines, but also strong emergence across them. We believe that the perspective from evolutionary depth could also illuminate other areas of human cognition. And we believe that animal-cognition research will continue to provide a synergistic complement to human research in future years. Acknowledgments: The preparation of this article was supported by Grants HD061455 and HD060563 from NICHD and BCS-0956993 from NSF. Author Contributions: J. David Smith conceived of the idea and was primary author of the paper. Alexandria C. Zakrzewski, Jennifer M. Johnson, Jeanette C. Valleau, and Barbara A. Church contributed background research and ideas for the manuscript. They also helped write and prepare the manuscript. Conflicts of Interest: The authors declare no conflict of interest.

References 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12.

13.

Ashby, F.G.; Maddox, W.T. Human Category Learning 2.0. Ann. N. Y. Acad. Sci. 2011, 1224, 147–161. [CrossRef] [PubMed] Briscoe, E.; Feldman, J. Conceptual Complexity and the Bias/Variance Tradeoff. Cognition 2011, 118, 2–16. [CrossRef] [PubMed] Brooks, L.R. Nonanalytic Concept Formation and Memory for Instances. In Cognition and Categorization; Rosch, E., Lloyd, B.B., Eds.; Erlbaum: Hillsdale, NJ, USA, 1978; pp. 169–211. Knowlton, B.J.; Squire, L.R. The Learning of Categories: Parallel Brain Systems for Item Memory and Category Knowledge. Science 1993, 262, 1747–1749. [CrossRef] [PubMed] Medin, D.L.; Schaffer, M.M. Context Theory of Classification Learning. Psychol. Rev. 1978, 85, 207–238. [CrossRef] Murphy, G.L. The Big Book of Concepts; MIT Press: Cambridge, MA, USA, 2002. Nosofsky, R.M. Attention and Learning Processes in the Identification and Categorization of Integral Stimuli. J. Exp. Psychol. Learn. Mem. Cognit. 1987, 13, 87–108. [CrossRef] Smith, J.D. Prototypes, Exemplars, and the Natural History of Categorization. Psychon. Bull. Rev. 2014, 21, 312–331. [CrossRef] [PubMed] Cerella, J. Visual Classes and Natural Categories in the Pigeon. J. Exp. Psychol. Hum. Percept. Perform. 1979, 5, 68–77. [CrossRef] [PubMed] Herrnstein, R.J.; Loveland, D.H.; Cable, C. Natural Concepts in Pigeons. J. Exp. Psychol. Anim. Behav. Process. 1976, 2, 285–302. [CrossRef] Medin, D.L. A Theory of Context in Discrimination Learning. In The Psychology of Learning and Motivation; Bower, G., Ed.; Academic Press: New York, NY, USA, 1975; Volume 9, pp. 263–314. Pearce, J.M. Discrimination and Categorization. Animal Learning and Cognition. In Handbook of Perception and Cognition Series, 2nd ed.; Mackintosh, N.J., Ed.; Academic Press: San Diego, CA, USA, 1994; Volume 18, pp. 109–134. Smith, J.D.; Minda, J.P.; Washburn, D.A. Category Learning in Rhesus Monkeys: A Study of the Shepard, Hovland, and Jenkins (1961) Tasks. J. Exp. Psychol. Gen. 2004, 133, 398–414. [CrossRef] [PubMed]

Behav. Sci. 2016, 6, 12

14. 15. 16. 17. 18. 19. 20. 21. 22. 23. 24. 25. 26. 27. 28. 29. 30. 31. 32. 33. 34. 35. 36. 37. 38. 39. 40.

22 of 24

Wasserman, E.A.; Kiedinger, R.E.; Bhatt, R.S. Conceptual Behavior in Pigeons: Categories, Subcategories, and Pseudocategories. J. Exp. Psychol. Anim. Behav. Process. 1988, 14, 235–246. [CrossRef] Cheney, D.L.; Seyfarth, R.M. How Monkeys See the World; University of Chicago Press: Chicago, IL, USA, 1990. Posner, M.I.; Keele, S.W. On the Genesis of Abstract Ideas. J. Exp. Psychol. 1968, 77, 353–363. [CrossRef] [PubMed] Homa, D.; Sterling, S.; Trepel, L. Limitations of Exemplar-Based Generalization and the Abstraction of Categorical Information. J. Exp. Psychol. Hum. Learn. 1981, 7, 418–439. [CrossRef] Smith, J.D.; Redford, J.S.; Haas, S.M. Prototype Abstraction by Monkeys (Macaca mulatta). J. Exp. Psychol. Gen. 2008, 137, 390–401. [CrossRef] [PubMed] Ashby, F.G.; Alfonso-Reese, L.A.; Turken, A.U.; Waldron, E.M. A Neuropsychological Theory of Multiple Systems in Category Learning. Psychol. Rev. 1998, 105, 442–481. [CrossRef] [PubMed] Ashby, F.G.; Ell, S.W. The Neurobiology of Human Category Learning. Trends Cognit. Sci. 2001, 5, 204–210. [CrossRef] Maddox, W.T.; Ashby, F.G. Dissociating Explicit and Procedural-Learning Based Systems of Perceptual Category Learning. Behav. Process. 2004, 66, 309–332. [CrossRef] [PubMed] Seger, C.A.; Miller, E.K. Category Learning in the Brain. Annu. Rev. Neurosci. 2010, 33, 203–219. [CrossRef] [PubMed] Ahn, W.K.; Medin, D.L. A Two-Stage Model of Category Construction. Cognit. Sci. 1992, 16, 81–121. [CrossRef] Feldman, J. Minimization of Boolean Complexity in Human Concept Learning. Nature 2000, 407, 630–633. [CrossRef] [PubMed] Medin, D.L.; Wattenmaker, W.D.; Hampson, S.E. Family Resemblance, Conceptual Cohesiveness, and Category Construction. Cognit. Psychol. 1987, 19, 242–279. [CrossRef] Regehr, G.; Brooks, L.R. Category Organization in Free Classification: The Organizing Effect of an Array of Stimuli. J. Exp. Psychol. Learn. Mem. Cognit. 1995, 21, 347–363. [CrossRef] Shepard, R.N.; Hovland, C.I.; Jenkins, H.M. Learning and Memorization of Classifications. Psychol. Monogr. 1961, 75, 1–42. [CrossRef] Erickson, M.A.; Kruschke, J.K. Rules and Exemplars in Category Learning. J. Exp. Psychol. Gen. 1998, 127, 107–140. [CrossRef] [PubMed] Nosofsky, R.M.; Palmeri, T.J.; McKinley, S.C. Rule-Plus-Exception Model of Classification Learning. Psychol. Rev. 1994, 101, 53–79. [CrossRef] [PubMed] Ashby, F.G.; Waldron, E.M. On the Nature of Implicit Categorization. Psychon. Bull. Rev. 1999, 6, 363–378. [CrossRef] [PubMed] Rossi, A.F.; Pessoa, L.; Desimone, R.; Ungerleider, L.G. The Prefrontal Cortex and the Executive Control of Attention. Exp. Brain Res. 2009, 192, 489–497. [CrossRef] [PubMed] Rosch, E. On the Internal Structure of Perceptual and Semantic Categories. In Cognitive Development and the Acquisition of Language; Moore, T., Ed.; Academic Press: New York, NY, USA, 1973; pp. 111–144. Smith, J.D. Exemplar Theory’s Predicted Typicality Gradient Can be Tested and Disconfirmed. Psychol. Sci. 2002, 13, 437–442. [CrossRef] [PubMed] Posner, M.I.; Goldsmith, R.; Welton, K.E., Jr. Perceived Distance and the Classification of Distorted Patterns. J. Exp. Psychol. 1967, 72, 28–38. [CrossRef] Aydin, A.; Pearce, J.M. Prototype Effects in Categorization by Pigeons. J. Exp. Psychol. Anim. Behav. Process. 1994, 20, 264–277. [CrossRef] Huber, L.; Lenz, R.A. Test of the Linear Feature Model of Polymorphous Concept Discrimination with Pigeons. Q. J. Exp. Psychol. 1993, 46, 1–18. [CrossRef] Jitsumori, M. A Prototype Effect and Categorization of Artificial Polymorphous Stimuli in Pigeons. J. Exp. Psychol. Anim. Behav. Process. 1996, 22, 405–419. [CrossRef] [PubMed] Von Fersen, L.; Lea, S.E. Category Discrimination by Pigeons Using Five Polymorphous Features. J. Exp. Anal. Behav. 1990, 54, 69–84. [CrossRef] [PubMed] White, K.G.; Alsop, B.; Williams, L. Prototype Identification and Categorization of Incomplete Figures by Pigeons. Behav. Process. 1993, 30, 253–258. [CrossRef] Smith, J.D.; Chapman, W.P.; Redford, J.S. Stages of Category Learning in Monkeys (Macaca mulatta) and Humans (Homo sapiens). J. Exp. Psychol. Anim. Behav. Process. 2010, 36, 39–53. [CrossRef] [PubMed]

Behav. Sci. 2016, 6, 12

41. 42. 43.

44. 45. 46. 47.

48. 49. 50. 51. 52. 53. 54. 55. 56.

57. 58.

59. 60. 61. 62. 63. 64. 65.

23 of 24

Blair, M.; Homa, D. Expanding the Search for a Linear Separability Constraint on Category Learning. Mem. Cognit. 2001, 29, 1153–1164. [CrossRef] [PubMed] Murphy, G.L. Ecological Validity and the Study of Concepts. In The Psychology of Learning and Motivation; Ross, B.H., Ed.; Academic Press: San Diego, CA, USA, 2003; Volume 43, pp. 1–41. Murphy, G.L. The Study of Concepts Inside and Outside the Laboratory: Medin versus Medin. In Categorization Inside and Outside of the Laboratory: Essays in Honor of Douglas L. Medin; Ahn, W., Goldstone, R.L., Love, B.C., Markman, A.B, Wolff, P., Eds.; APA: Washington, DC, USA, 2005; pp. 179–195. Cook, R.; Smith, J.D. Stages of Abstraction and Exemplar Memorization in Pigeons’ Category Learning. Psychol. Sci. 2006, 17, 1059–1067. [CrossRef] [PubMed] Shepard, R.N. Toward a Universal Law of Generalization for Psychological Science. Science 1987, 237, 1317–1323. [CrossRef] [PubMed] Shepard, R.N. Perceptual-Cognitive Universals as Reflections of the World. Behav. Brain Sci. 2001, 24, 581–601. [CrossRef] [PubMed] Smith, J.D.; Coutinho, M.V.C.; Couchman, J.J. The Learning of Exclusive-or Categories by Monkeys (Macaca mulatta) and Humans (Homo sapiens). J. Exp. Psychol. Anim. Behav. Process. 2011, 37, 20–29. [CrossRef] [PubMed] Malt, B.C. Category Coherence in Cross-Cultural Perspective. Cognit. Psychol. 1995, 29, 85–148. [CrossRef] Smith, J.D.; Boomer, J.; Zakrzewski, A.C.; Roeder, J.L.; Church, B.A.; Ashby, F.G. Deferred Feedback Sharply Dissociates Implicit and Explicit Category Learning. Psychol. Sci. 2014, 25, 447–457. [CrossRef] [PubMed] Newell, B.R.; Dunn, J.C.; Kalish, M. The Dimensionality of Perceptual Category Learning: A State-Trace Analysis. Mem. Cognit. 2010, 38, 563–581. [CrossRef] [PubMed] Nosofsky, R.M.; Johansen, M.K. Exemplar-Based Accounts of “Multiple-System” Phenomena in Perceptual Categorization. Psychon. Bull. Rev. 2000, 7, 375–402. [PubMed] Nosofsky, R.M.; Stanton, R.D.; Zaki, S.R. Procedural Interference in Perceptual Classification: Implicit Learning or Cognitive Complexity? Mem. Cognit. 2005, 33, 1256–1271. [CrossRef] [PubMed] Minda, J.P.; Smith, J.D. Prototypes in Category Learning: The Effects of Category Size, Category Structure, and Stimulus Complexity. J. Exp. Psychol. Learn. Mem. Cognit. 2001, 27, 775–799. [CrossRef] Antzoulatos, E.G.; Miller, E.K. Differences between Neural Activity in Prefrontal Cortex and Striatum during Learning of Novel Abstract Categories. Neuron 2011, 71, 243–249. [CrossRef] [PubMed] Garner, W. The Processing of Information and Structure; Wiley: New York, NY, USA, 1974. Smith, J.D.; Beran, M.J.; Crossley, M.J.; Boomer, J.; Ashby, F.G. Implicit and Explicit Category Learning by Macaques (Macaca mulatta) and Humans (Homo sapiens). J. Exp. Psychol. Anim. Behav. Process. 2010, 36, 54–65. [CrossRef] [PubMed] Ashby, F.G.; Maddox, W.T. Human Category Learning. Annu. Rev. Psychol. 2005, 56, 149–178. [CrossRef] [PubMed] Smith, J.D.; Berg, M.E.; Cook, R.G.; Murphy, M.S.; Crossley, M.J.; Boomer, J.; Spiering, B.; Beran, M.J.; Church, B.A.; Ashby, F.G.; et al. Implicit and Explicit Categorization: A Tale of Four Species. Neurosci. Biobehav. Rev. 2012, 36, 2355–2369. [CrossRef] [PubMed] Smith, J.D.; Ashby, F.G.; Berg, M.E.; Murphy, M.S.; Spiering, B.; Cook, R.G.; Grace, R.C. Pigeons’ Categorization may be Exclusively Nonanalytic. Psychon. Bull. Rev. 2011, 18, 414–421. [CrossRef] [PubMed] Pearce, J.M.; Esber, G.R.; George, D.N.; Haselgrove, M. The Nature of Discrimination Learning in Pigeons. Learn. Behav. 2008, 36, 188–199. [CrossRef] [PubMed] Foard, C.F.; Kemler, D.G. Holistic and Analytic Modes of Processing: The Multiple Determinants of Perceptual Analysis. J. Exp. Psychol. Gen. 1984, 113, 94–111. [CrossRef] [PubMed] Smith, L.B.; Kemler, D.G. Levels of Experienced Dimensionality in Children and Adults. Cognit. Psychol. 1978, 10, 502–532. [CrossRef] Rosch, E.; Mervis, C.B. Family Resemblances: Studies in the Internal Structure of Categories. Cognit. Psychol. 1975, 7, 573–605. [CrossRef] Smith, J.D.; Crossley, M.J.; Boomer, J.; Church, B.A.; Beran, M.J.; Ashby, F.G. Implicit and Explicit Category Learning by Capuchin Monkeys (Cebus apella). J. Comp. Psychol. 2012, 126, 294–304. [CrossRef] [PubMed] Semendeferi, K.; Lu, A.; Schenker, N.; Damásio, H. Humans and Great Apes Share a Large Frontal Cortex. Nat. Neurosci. 2002, 5, 272–276. [CrossRef] [PubMed]

Behav. Sci. 2016, 6, 12

66. 67. 68.

69. 70.

71. 72.

73. 74. 75. 76. 77. 78. 79. 80.

24 of 24

Roberts, A.C. Comparison of Cognitive Function in Human and Non-human Primates. Cognit. Brain Res. 1996, 3, 319–327. [CrossRef] Washburn, D.A. Stroop-like Effects for Monkeys and Humans: Processing Speed or Strength of Association? Psychol. Sci. 1994, 5, 375–379. [CrossRef] [PubMed] Smith, J.D.; Zakrzewski, A.C.; Johnston, J.J.; Roeder, J.L.; Boomer, J.; Ashby, F.G.; Church, B.A. Generalization of Category Knowledge and Dimensional Categorization in Humans (Homo sapiens) and Nonhuman Primates (Macaca mulatta). J. Exp. Psychol. Anim. Learn. Cognit. 2015, 41, 322–335. [CrossRef] [PubMed] Casale, M.B.; Roeder, J.L.; Ashby, F.G. Analogical Transfer in Perceptual Categorization. Mem. Cognit. 2012, 40, 434–449. [CrossRef] [PubMed] Smith, J.D.; Beran, M.J.; Redford, J.S.; Washburn, D.A. Dissociating Uncertainty States and Reinforcement Signals in the Comparative Study of Metacognition. J. Exp. Psychol. Gen. 2006, 135, 282–297. [CrossRef] [PubMed] Zakrzewski, A.C.; Church, B.A.; Smith, J.D. The Transfer of Category Knowledge: The Affordances of Explicit Rules. Unpublished work, 2016. Zakrzewski, A.C.; Church, B.A.; Johnson, J.M.; Boomer, J.; Ashby, F.G.; Smith, J.D. Can Rhesus Macaques (Macaca mulatta) Transfer Category Knowledge When Reinforcement Is Deferred? Poster session presented at the 56th Annual Meeting of the Psychonomic Society, Chicago, IL, USA, 19–22 November 2015. Medin, D.L.; Schwanenflugel, P.J. Linear Separability in Classification Learning. J. Exp. Psychol. Hum. Learn. 1981, 7, 355–368. [CrossRef] Medin, D.L.; Dewey, G.I.; Murphy, T.D. Relationships between Item and Category Learning: Evidence that Abstraction is not Automatic. J. Exp. Psychol. Learn. Mem. Cognit. 1983, 9, 607–625. [CrossRef] Anderson, J.R. The Adaptive Nature of Human Categorization. Psychol. Rev. 1991, 98, 409–429. [CrossRef] Feldman, J. Tuning your Priors to the World. Top. Cognit. Sci. 2013, 5, 13–34. [CrossRef] [PubMed] Ashby, F.G. Multidimensional Models of Categorization. In Multidimensional Models of Perception and Cognition; Ashby, F.G., Ed.; Erlbaum: Hillsdale, NJ, USA, 1992; pp. 449–483. Fried, L.S.; Holyoak, K.J. Induction of Category Distributions: A Framework for Classification Learning. J. Exp. Psychol. Learn. Mem. Cognit. 1984, 10, 234–257. [CrossRef] Rosch, E. Cognitive Representations of Semantic Categories. J. Exp. Psychol. Gen. 1975, 104, 192–233. [CrossRef] Blough, D.S. Attention Shifts in a Maintained Discrimination. Science 1969, 166, 125–126. [CrossRef] [PubMed] © 2016 by the authors; licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC-BY) license (http://creativecommons.org/licenses/by/4.0/).