Categorization: a fundamental principle of cognitive economy.

Many animals can detect basic characteristics of objects or events and store them as generalized classes beyond specific details of sensory inputs. Hence, crickets sharply divide a range of sound frequency into either “mate” (<16 kilohertz) or “bat” (>16 kilohertz) - a predator - even though the tones vary along a continuum. While such categorical perception is known primarily from human speech perception the cricket-example suggests that forming perceptual categories is a rather widespread feature of sensory systems (Wyttenbach et al., 1996). The ability to classify similar but not identical things as equivalent (by reacting to them in the same manner) is, in fact, an optimal solution to minimize processing requirements, which is advantageous especially for small information-processing systems (Cook, Wright & Kendrick, 1990; Delius, Siemann & Jitsumori; 2000). The inclination to compress the amount of information to be retained is observed in many animals although the strategy used to attain this “cognitive economy” (Huber, 2004) varies between (and within) species: Categorization can be done at different levels of abstraction, requiring different degrees of instruction.

The seminal work by Herrnstein & Loveland (1964), for instance, showed that pigeon could be trained to classify photographs on the base of whether or not they contained a human being. This classification skill does not simply apply to the trained pictures but generalized to novel pictures not used in the training phase. By relying on low-level visual features pigeons can even learn to discriminate color slides of paintings by Picasso from those of Monet and generalize this ability to novel paintings by other cubists (e.g., Braque) and impressionists (e.g. Cezanne) (Watanabe et al. 1995). More developed animals such as monkeys, by contrast, can learn to take multiple stimulus dimensions into account in order to make sharp distinctions between categories such as “animals” or “trees”. (see Miller et al., 2003). In contrast to pigeons this ability is not simply based on grouping of the stimuli according to physical similarity because monkeys can be trained to categorize computer generated visual stimuli along a morphing continuum of various blends of two different objects (Freedman et al. 2001). Hence, by merging different amounts of a “cat” and a “dog” image, monkeys can be trained to develop discrete category boundary between the two classes of animals even as the similarity between two stimuli of one category (e.g. two different types of cats) is lower than the similarity between stimuli belonging to the other category (e.g., a cat and a cat-like dog). The sharp boundaries between categories is even evident at the neural level as single neurons in the (inferior temporal cortex and the prefrontal) cortex of the trained animals react selectively to a member of a given category with little differentiation between various examples of the category but sharp transition in neural activity between categories (Freedman et al., 2001; Vogels, 1999).

In humans (and possibly in other animals) the evolutionary pressure to minimize processing requirement seem to have lead to the development of specialized and functionally dissociable neural circuits for the processing of objects for which rapid identification could have had survival advantage (Caramazza & Shelton, 1998). Hence, brain damage in humans can lead to impairments at recognizing or identifying one category of objects more than another. For instance, while some patients have difficulties in correctly distinguishing cucumbers from green peppers without having problems in identifying animals, other patients do fine with vegetables and fruits but identify a giraffe as kangaroo or claim that the natural color of an elephant is orange. Post traumatic impairments at recognizing or identifying one category of objects more than another have been reported for semantic categories as broad as living vs. non-living things to more narrow categories such as fruits/vegetables, animals, and tools. Functional brain imaging studies with healthy subjects confirm these observations by revealing distinct cortical networks activated by living and non-living things. Non-living things activating regions in the parietal and frontal lobes of the language dominant left hemisphere, while living things activate a network of cortical regions mainly in the occipital lobes and along the temporo-parietal junctions (see Figure 1).

Figure 1: A 2D map of the two hemispheres (Geometric atlas; Toro & Burnod, 03) with the main lobes indicated by different colors (pale red= frontal lobe; pale green= parietal lobe; pale yellow = temporal lobe; grey = occipital lobe. The central circle represents the insula). The blue and violet shaded zones in the circles specify cortical activity observed across a series of experiments with healthy subjects while they were observing pictures of living things or non-living things, respectively.

A priori, the development of segregated neural circuits for specific categories could emerge entirely as the result of experience and learning (see the monkey-example above). Accordingly, artificial neural networks can develop separate neural representations for arbitrary categories provided that they are adequately trained (Polk & Farah, 1995). In humans, however, the organization of conceptual knowledge about living and non-living things seems to be specified in the genome, as lifelong selective deficits for living things (with normal knowledge about non-living things) can be observed after brain injuries that occurred few days post natal. That is, prior to any experience the cortical substrate for acquiring semantic memory for nonliving-things can develop in the absence of the development of function of semantic memory for living-things (Farah & Rabinowitz, 03).

In conclusion, a strong common bias to reduce processing loads (through the search of commonalities between objects) may have led to the emergence of categorization skills in different species. The specific solution to achieve this goal, however, varies between kinds. It is yet unknown when categorization appeared in the phylogenetic scale and whether it appeared only once. However, as such, categorization does not require language skills. Does language rely on categorization (Herrnstein, 1984)?

Categorization in preverbal infants

Studies with preverbal children have shown that as soon as infants are able to parse words from the speech stream, words have a powerful influence on conceptual organization (Waxman, 2004). Using a novelty-preference task Waxman and colleagues have shown that children as young as 9 months old form category-based commonalities simply by establishing links between novel words and objects. In the novelty-preference task children are familiarized to a series of objects from a given category (e.g. animals), and are then tested with two new objects of which one is a member of the familiarized category (e.g., another animal) and the other a member of a new category (e.g. a fruit). If infants notice the category-based commonalities among the familiarization objects, in the test phase they should show a preference for the fruit. During the familiarization phase the presentation of the member of the given category is either accompanied by a neutral sentence such as: “look at this”, or a sentence that introduced a novel word: “this is a blicket” (Waxman and Markow, 1995; 1998). While children in the neutral condition show no preference for one or the other object in a subsequent test phase, infants in the novel-word condition show a clear preference for the object that belongs to the new category (i.e. the fruit). The presence of words seems even to allow infants to generalize beyond shared shape similarity to nonvisible properties of objects (Graham et al., 2004). According to Waxman and colleagues, these experiments show that infants begin the task of word learning equipped with an (universally) available expectation linking novel words to a range of commonalities between objects. In the authors’ terms, “words serve as invitation to form categories” (Waxman and Markow, 1995). Slightly turning the argument we could therefore say that as infants are strongly predisposed to seek for optimal solutions to minimize processing requirements, words provide a perfect clue allowing them to rapidly form categories.

Although nothing in the here described issue allows speculations about the emergence of language, an innate drive to minimize information processing could at least foster its evolution.

Why words?  A series of speculative hypotheses

Interestingly, the successful categorization of objects by young human infants’ as described above hinges on the use of verbal stimuli and does not occur for equally attention-engaging non-verbal auditory stimuli (Balaban & Waxman, 1997; 2002). That is, although both verbal and non-verbal auditory stimuli capture infants’ attention, only words support the establishment of categories. And this, even before children understand their meaning. What is so special about words that they allow the quick establishment of categories?

Though highly speculative, a possible reason for the advantage of verbal over non-verbal sounds in the formation of categories could lie in the role of (vocal) imitation/simulation. Human infants are predisposed to produce verbal sounds and by this the status of a spoken word might be different from the status of a non-verbal (non-reproducible) sound. In fact, studies with human adults have demonstrated that observing and executing a motor action activate partly overlapping cortical regions in several areas of the brain (Iacoboni et al., 1999). This observation has been taken to assume that observing actions (that can be executed by the observer) involves a “resonance” mechanism that allows to directly mapping a perceived action onto an internal motor representation of the same action. This internal simulation, in turn, is the base of imitation. Coherent with this assumption, distinct patterns of neural activity are observed when adult humans are asked to report whether or not people (imitable) or dogs (not imitable) can perform a common set of behaviors. While people and dogs are capable of performing many of the same actions (e.g. run, sit, bite), the representation of this knowledge seem to be associated with distinct patterns of neural activity (Mason et al., 2004).

The “resonance” or “mirror neuron” system has initially been observed through analysis of the activity of single neurons in rhesus monkeys (Rizzolatti et al., 1999). Yet, while many species can copy the models’ choice of object (goal of the action), imitation of the perceived motor action of a model is rare in non-human animals (Premack, 2004). In human infants, by contrast, automatic imitation of elementary actions such as facial and manual gestures is already present at only few days of age (e.g., Meltzoff & Moore, 1977). Hence, if this “resonance” mechanism truly exists, it has gained sophistication in humans compared to monkeys and might provide an important step towards the understanding of human evolution in general (Ramachandran, 01).

By way of such a system verbal stimuli could thus gain their special status (independently of the fact that it is language!) over the equally attention-engaging but non-“resonating”, non-verbal auditory stimuli in infants’ categorization. Moreover, at the same time as linking words to commonalities between objects satisfy the hypothesized bias to reduce processing loads (by forming categories), it also provides infants with a mean to establish a rudimentary lexicon that can be fine-tuned subsequently. Finally, although categorization by help of words may initially rely on physical aspect of the stimuli (e.g. visual similarity between objects), once the infant understood the principle, words could help to form abstract categories beyond simple grouping-by-similarity and may eventually lead to the development of concepts.

Lastly, the rapid formation of categories by help of verbal stimuli in young infants requires that someone tell the infant: “this is an X”. In other words, there is a need for a “teacher”. Like imitation, teaching seems to be strictly human and does not occur in other animals, not even in chimpanzees (Inoue-Nakamura et al., 1997). Teaching reverses the flow of information found in imitation by providing feedback to the imitator (Premack, 04) and therefore requires some degree of intentional communication. Teaching entails the ability to represent other minds. In the frame of the “resonance” mechanism hypothesis, miming and teaching rely on much the same principle. Anytime you watch someone else doing something (or even starting to do something as the system is also activated by still pictures that only imply motion, Nishitani and Hari, 2002), the resonance mechanism allows the "reading" and understanding of other's intentions, and thus to develop a sophisticated "theory of minds."  

 Balaban MT, Waxman SR. Do words facilitate object categorization in 9-month-old infants? J Exp Child Psychol 1997; 64: 3-26.

Caramazza A, Shelton JR. Domain-specific knowledge systems in the brain the animate-inanimate distinction. J Cogn Neurosci 1998; 10: 1-34.

Cook, R. G., Wright, A. A. & Kendrick, D. F. (1990). Visual categorization by pigeons. In M. L. Commons, R. J. Herrnstein, S. Kosslyn & D. Mumford (Eds.), Quantitative analysis of behavio: Vol. 8. Behavioral approaches to pattern recognition and concept formation (pp. 187-214). Hillsdale, NJ: Lawrence Erlbaum.

Delius, J. D., Jitsumori, M., & Siemann, M. (2000). Stimulus equivalencies through discrimination reversals. In C. Heyes & L. Huber (Eds.), The evolution of cognition (pp. 103-122). Cambridge, MA: MIT Press.

Farah, M Rabinowitz, (2003). Genetic and environmental influence on the organization of semantic memory in the brain: Is "living things" an innate category? Cognitive neuropsychology, 20, 401-408

Freedman DJ, Riesenhuber M, Poggio T, Miller EK. Categorical representation of visual stimuli in the primate prefrontal cortex. Science. 2001 Jan 12;291(5502):312-6.

Graham SA, Kilbreath CS, Welder AN. Thirteen-month-olds rely on shared labels and shape similarity for inductive inferences. Child Dev. 2004 Mar-Apr;75(2):409-27.

Herrnstein, R. J. (1984). Objects, categories, and discriminative stimuli. In H. L. Roitblat, T. G. Bever, & H. S. Terrace (Eds.), Animal cognition (pp. 233-261). Hillsdale, NJ: Lawrence Erlbaum Associates. 

Herrnstein RJ, Loveland DH. Complex Visual Concept in the Pigeon. Science 1964; 146: 549-51.

Huber, L. (2004) Visual Categorization in Pigeons. In Robert G. Cook (Ed). Avian visual cognition. Comparative Cognition Press.

Iacoboni M, Woods RP, Brass M, Bekkering H, Mazziotta JC, Rizzolatti G. Cortical mechanisms of human imitation. Science. 1999 Dec 24;286(5449):2526-8.

Inoue-Nakamura N, Matsuzawa T. Development of stone tool use by wild chimpanzees (Pan troglodytes). J Comp Psychol 1997; 111: 159-73.

Mason MF, Banfield JF, Macrae CN.  Thinking about actions: the neural substrates of person knowledge. Cereb Cortex. 2004 Feb;14(2):209-14.

Meltzoff AN, Moore MK. Imitation of facial and manual gestures by human neonates. Science 1977; 198: 74-8.

Miller EK, Nieder A, Freedman DJ, Wallis JD. Neural correlates of categories and concepts. Curr Opin Neurobiol. 2003 Apr;13(2):198-203.

Nishitani N, Hari R. Viewing lip forms: cortical dynamics. Neuron. 2002 Dec 19;36(6):1211-20.        

Polk TA, Farah MJ. Brain localization for arbitrary stimulus categories: a simple account based on Hebbian learning. Proc Natl Acad Sci U S A 1995; 92: 12370-3.

Premack D. Psychology. Is language the key to human intelligence? Science 2004; 303: 318-20.

Rizzolatti G, Fadiga L, Fogassi L, Gallese V. Resonance behaviors and mirror neurons.

Arch Ital Biol. 1999 May;137(2-3):85-100.

Toro R, Burnod Y. Geometric atlas: modeling the cortex as an organized surface. Neuroimage 2003; 20: 1468-84.

Vogels R. Categorization of complex visual images by rhesus monkeys. Part 2: single-cell study. Eur J Neurosci. 1999 Apr;11(4):1239-55

Watanabe, S., Sakamoto, J., & Wakita, M. (1995). Pigeons' discrimination of paintings by Monet and Picasso. Journal of the Experimental Analysis of Behavior, 63, 165-174. 

Waxman, Sandra R. (in press). Everything had a name, and each name gave birth to a new thought: Links between early word-learning and conceptual organization. In D. G. Hall & S. R. Waxman (Eds.), From many strands: Weaving a lexicon. Cambridge: MIT Press.

Waxman SR, Markow DB. Words as invitations to form categories: evidence from 12- to 13-month-old infants. Cognit Psychol 1995; 29: 257-302.

Waxman SR, Markow DB. Object properties and object kind: twenty-one-month-old infants' extension of novel adjectives. Child Dev 1998; 69: 1313-29.

Wyttenbach RA, May ML, Hoy RR. Categorical perception of sound frequency by crickets. Science 1996; 273: 1542-4.