Authors: Yale E. Cohen, Jung Hoon Lee, Joji Tsunada, Brian E. Russ
Categories: Article, auditory system, categorization, non-human primate, prefrontal cortex, vocalization
Source: Handbook of behavioral neuroscience
Authors: Yale E. Cohen, Jung Hoon Lee, Joji Tsunada, Brian E. Russ
Communication is one of the fundamental components of both human and non-human animal behavior. Whereas the benefits of language in human evolution are obvious, other communication systems have also evolved to convey information that is critical for survival. This chapter focuses on auditory communication signals, specifically species-specific vocalizations and the underlying neural processes that may support their use in guiding goal-directed behavior. We first highlight the fundamental role that species-specific vocalizations play in the socioecology of several species of non-human primates, with a focus on rhesus monkeys (Macaca mulatta). Finally, we discuss the role that the ventrolateral prefrontal cortex may play in the categorization of species-specific vocalizations.
Categories provide a mechanism to mentally reason, manipulate and respond to objects in our environment. If goal-directed behavior can be thought of as the processes that transform sensory signals into actions, the computations that form the intermediate steps of this transformation involve categorical representations (Ashby and Berretty, 1997; Grinband et al., 2006). These computations are further modulated by context, motivation and other factors that permit behavior to be both flexible and adaptive. In the wild, consider what happens when a lion sees a gazelle. If the lion is hungry, it may categorize the gazelle as prey and form a motor plan that allows it to capture the gazelle. But if the lion is not hungry, the lion may not categorize the gazelle as prey and may disregard the gazelle entirely. In this chapter, we highlight categorization and the factors contributing to categorization with an emphasis on auditory categorization in human and non-human primates. In particular, we highlight the categorization of higher-order features of vocalizations.
Any given stimulus (object) can be a member of several different categories, depending on what rules process the stimulus. These different categorical levels are often thought of as hierarchical; categories at the top of the hierarchy are the most general (superordinate categories) whereas those at the bottom (subordinate categories) are the most specific. Basic-level subordinate categories are the terms that people commonly use when encountering an object. For example, Lassie can be categorized as an animal, a dog, or a specific collie. The superordinate category would be “animal,” the basic-level category would be “dog,” and the subordinate category would be “collie.” However, these different categories are not equipotent, in the sense that they require different amounts of neural processing; we categorize objects into basic level categories faster and more accurately than we categorize objects into superordinate or subordinate categories (Rosch et al., 1976). Also, the level at which an object is categorized depends on previous experience and knowledge; a dog expert might classify Lassie, at the basic level, as a “collie” or more specifically a “rough collie” (Gauthier and Logothetis, 2000; Marschark et al., 2004).
Different hypothetical frameworks can be used to describe the relationship between an object’s membership in basic and more superordinate categories. One framework posits that superordinate categories contain a set of features that belong to all of the members of the more basic category (Smith et al., 1974; Rosch et al., 1976; Damasio, 1989; Devlin et al., 1998; Martin et al., 2002). For instance, the basic-level of category of “dog” might contain descriptors like “has fur,” “has wet nose,” “has four legs,” “breathes,” “is mobile,” “can reproduce on its own,” etc. On the other hand, the superordinate category of “animal” contains descriptors “breathes,” “is mobile,” “can reproduce on its own,” etc. An alternative view is that the properties of a basic-level category are not omitted from the superordinate category, but are represented as more abstract variable values in this higher-order category (MacNamara, 1982; Macnamara, 1999; Prasada, 2000).
Perceptual similarity is one of the key elements that determine a stimulus’ categorical membership (Liberman et al., 1967; Eimas et al., 1971; Kuhl and Miller, 1975; Lasky et al., 1975; Miyawaki et al., 1975; Streeter, 1976; Sandell et al., 1979; Kuhl and Padden, 1982, 1983; Boyton and Olson, 1987, 1990; Wyttenbach et al., 1996; Doupe and Kuhl, 1999). Perceptual categories are based on physical similarities or dissimilarities between auditory objects. For example, we can categorize male and female voices by listening to the pitch of their voices, with female voices characteristically having a higher pitch than males. In another example, listeners can perceive different speech signals as belonging to the same phonemic category, independent of pitch or timbre differences, or viewers can perceive different visual signals as being members of the same color category (Bornstein et al., 1976; Sandell et al., 1979; Boyton and Olson, 1987, 1990).
One prominent feature of perceptual categories is that they are often accompanied by categorical perception. In categorical perception, a subject’s perception of an object does not vary smoothly with changes in the physical properties of the object (Liberman et al., 1967; Ashby and Berretty, 1997; Miller et al., 2003). That is, objects on one side of the categorical boundary are treated similarly, despite potentially large differences between their physical properties. At locations near the category boundary, small changes in an object’s properties can lead to large changes in perception.
The classic example of categorical perception is the categorization of speech units into phonetic categories (Liberman et al., 1967; Eimas et al., 1971; Mann, 1980; Kuhl and Padden, 1982, 1983; Lotto et al., 1998; Holt, 2006). In a seminal study, Liberman et al. (1967) created morphed versions of two different phonemes and asked subjects to report the phoneme that they heard. Liberman and colleagues found that when subjects were presented with a morphed stimulus that contained more than 50% of a phoneme prototype, the subjects reliably perceived that stimulus as the prototype. That is, even though the presented stimuli varied smoothly in their physical features, subjects perceived the presented stimuli as being either one of the two phoneme prototypes.
Interestingly, the perceptual categorization of phonemes is not strictly a human behavior (Liberman et al., 1967; Kuhl and Padden, 1982, 1983; Kluender et al., 1987; Lotto et al., 1997; Russ et al., 2008). Rhesus macaques, chinchillas and Japanese quail perceive human phonemes in a manner comparable to that of humans. Since the manner in which perceptual categories, at least phonemes, are coded appears to be similar across a wide variety of animal species, it is hypothesized that the mechanisms underlying this perceptual categorization may be a fundamental component of vertebrate auditory processing from which human speech was bootstrapped.
Categories are not only formed based on the perceptual (physical) features of stimuli. Categories can also be based on more abstract information. An abstract category is one in which a group of arbitrary stimuli are linked together as a category based on some shared feature, a functional characteristic, or acquired knowledge. For instance, despite vast physical differences, “hammer,” “band saw,” and “pliers” are all members of the “tool” category. Similarly, a combination of physical characteristics and knowledge about their reproductive processes allow us to categorize “dogs,” “cats,” and “killer whales” in the category of “mammals.” However, if we use different criteria to form a category of “pets,” “dogs” and “cats” would be members of the “pet” category but “killer whales” would not.
Non-human primates can also categorize stimuli into abstract categories. Monkeys can be trained to categorize objects as being animals or non-animals (Fabre-Thorpe et al., 1998) or as trees or non-trees (Vogels, 1999). The capacity to represent even more abstract categories, such as ordinal numbers (Orlov et al., 2000; Nieder et al., 2002), is also present. Although these studies provide important insight into how abstract categories are built, their generalization to more ethological, natural conditions is limited by their use of non-ethological stimuli.
Behavioral studies that have used ethological stimuli have shown that non-human primates may form categories “spontaneously.” That is, they form categories in the absence of laboratory-based operant training. A classic example is the categorization of food-related species-specific vocalizations by rhesus monkeys (Hauser and Marler, 1993a,b; Hauser, 1998; Gifford III et al., 2003). In rhesus monkeys, information about the discovery of rare, high-quality food is transmitted by two different by a “harmonic arch;” and by a different vocalization called a “warble.” Importantly, whereas both harmonic arches and warbles transmit the same type of information, they have distinct spectrotemporal properties (i.e., they sound different). In contrast, “grunts” transmit a different type of information (the discovery of common, low-quality food) and are acoustically distinct from harmonic arches and warbles.
Despite these acoustic differences, rhesus monkeys categorize these food-related calls based on the transmitted information and not their acoustic features. Monkeys do not discriminate between vocalizations that transmit the same referential information (i.e., harmonic arches and warbles) even though these vocalizations have different acoustic features. In contrast, they do discriminate between vocalizations that transmit different types of information (i.e., grunts versus warbles/harmonic arches). That is, rhesus monkeys perceive harmonic arches and warbles as if they belong to a single functional category (based on referential information and not based on acoustics) and treat grunts as a second, distinct category.
The formation of abstract acoustic categories is seen in other monkey species. Female Diana monkeys categorize and respond similarly to a male Diana monkey who is eliciting a leopard-alarm call or to a crested guinea fowl that is eliciting its unique species-specific leopard-alarm call (Züberbuhler and Seyfarth, 1997; Züberbuhler, 2000a,b,c). Diana monkeys also form cross-species categories with putty-nose monkeys, based on the ability of putty-nose monkeys to provide vocal warnings of eagles (Eckardt and Züberbuhler, 2004). These observations suggest that Diana monkeys form abstract categorical representations of vocalizations independent of acoustics and the species generating the signal. Finally, an example of a non-communicative multimodal category is “looming” stimuli; rhesus monkeys treat approaching auditory or visual stimuli in a comparable manner (Schiff et al., 1962; Ghazanfar et al., 2002; Maier et al., 2004).
Traditionally, the auditory cortex has been thought to be involved in feature extraction. However, more recent work has shown that the auditory cortex, particularly the primary auditory cortex, plays a substantive role in more advanced stages of auditory processing, such as auditory-object analysis (Sutter et al., 2000; Fishman et al., 2001, 2004; Miller et al., 2001; Nelken et al., 2003; Petkov et al., 2003; Poremba et al., 2003; Micheyl et al., 2005; Griffiths et al., 2007; Petkov et al., 2008). Consistent with these studies, other work suggests that the auditory cortex may also be involved in the computations underlying category processing in both human and non-human primates (Steinschneider et al., 1995; Guenther et al., 2004; Poeppel et al., 2004; Selezneva et al., 2006). For example, Brosch and colleagues (Selezneva et al., 2006) have shown that auditory-cortex neurons respond categorically to sequences of tone pips that are either increasing or decreasing in frequency. In a related study, the primary auditory cortex has been shown to contain a distributed representation of the voice-onset time of human phonemes (Steinschneider et al., 1995). It is thought that this representation and related representations form the neurophysiological bases for the perceptual categorization of phonemes. However, unlike the Brosch study, this work by Steinschneider emphasizes that categories are represented in population activity and not at the level of the single cell. Finally, visual and multimodal stimuli, such as faces and bimodal looming representations, also appear to have categorical-like representations in different regions of the auditory cortex (Hoffman et al., 2008; Maier et al., 2008).
What types of categorical processing then occur in subsequent areas of the cortical hierarchy? We suggest that there are two major classes of computational processing. First, neurons become increasingly sensitive to more abstract categories. For example, recent work from our group has suggested a role for the ventrolateral prefrontal cortex (vPFC) in categorizing the referential information that a vocalization transmits, as opposed to the vocalization’s acoustic properties (Gifford III et al., 2005; Cohen et al., 2006). Using an oddball paradigm (Näätänen and Tiitinen, 1996), we found that the activity of vPFC neurons was not modulated by transitions between presentations of food vocalizations that transmitted the same information (high-quality food), even though these vocalizations had significantly different acoustic structures. The vPFC activity, however, was modulated by transitions between presentations of food vocalizations that transmitted different types of information (low-quality versus high-quality food). These data suggested that, on average, vPFC neurons are modulated preferentially by transitions between presentations of food vocalizations that belong to functionally meaningful and different categories.
Second, the categorical representations in more central areas such as the prefrontal cortex are used to flexibly guide an animal’s behavior (Miller, 2000; Miller et al., 2002). That is, categorical information in the prefrontal cortex is critical for both the selection and retrieval of task-relevant information as it relates to the rules of an ongoing task (Asaad et al., 2000; Ashby and Spiering, 2004; Bunge, 2004; Badre et al., 2005; Bunge et al., 2005). Indeed, recent work from our laboratory (Russ et al., 2008) has demonstrated that when monkeys are asked to categorize pairs of the spoken words, vPFC neurons do not reflect the perceptual differences between the spoken words. Instead, vPFC activity reflects the monkeys’ behavioral reports (decisions) as to whether they perceive that the pair of spoken words is the same or whether they are different. That is, vPFC activity reflects how the monkeys actually respond and does not reflect how they should respond (which is based on the actual sensory/perceptual differences between the spoken words). Moreover, this activity seems to play a causal role in the decision-making process; disruption of the vPFC through transcranial magnetic stimulation significantly alters the time it takes the monkeys to report their decision (Russ and Cohen, unpublished observations).
This chapter has emphasized the relationship between categorical perception and the neural correlates underlying these representations. The focus has been on how vocalizations and other communication signals are represented in the auditory pathway. This chapter raises a number of questions. How are perceptual categories transformed into more abstract representations and what are the neural computations underlying this transformation? On a related note, what specific cortical regions are involved in categorization? Are categories processed in distinct cortical hierarchies or do cortical regions code all levels of categories simultaneously? Finally, this chapter emphasized unimodal categorization. However, communication signals are inherently multimodal (Sumby and Pollack, 1954; McGurk and MacDonald, 1976; Stein and Meredith, 1993; Calvert et al., 1997; Partan and Marler, 1999). Thus, a fundamental function of neural processing may be to integrate auditory and visual stimuli that provide complementary information (Hinde and Rowell, 1962; van Hooff, 1962; Maestripieri, 1997; Partan and Marler, 1999; Hauser and Akre, 2001; Partan, 2002; Ghazanfar and Logothetis, 2003; Ghazanfar et al., 2005). Are similar processes and areas involved in the categorization of visual and auditory stimuli or are new processes and cortical areas engaged to process these multimodal stimuli (Ghazanfar and Schroeder, 2006)? Since our world is full of multimodal sensory information, it will be important to investigate how we combine information from multiple domains into a single coherent signal.