Perhaps some of the most interesting findings reported in the neuropsychological literature are patients who can successfully recognize some categories of objects but not others. For example, JBR, a patient who suffered temporal lobe damage following herpes simplex encephalitis, was able to give fairly precise descriptions of nonliving artefacts, but was only able to give very impoverished responses for living things (Warrington & Shallice, 1984). JBR described a compass as "tools for telling direction you are going" and a briefcase as "a small case used by students to carry papers", but when asked to describe a parrot he said "don’t know" and he described a snail as "an insect animal". Category specific recognition impairments for living things have been documented on numerous occasions in the neuropsychology literature (Basso, Capitani & Laiacona, 1988; De Renzi & Lucchelli, 1994; Forde, Francis, Riddoch, Rumiati & Humphreys, 1997; Sartori & Job, 1984; Sheridan & Humphreys, 1993; Silveri & Gainotti, 1988; Warrington & Shallice, 1984), and there are also a few reports of patients with impairments for nonliving things (Cappa, Frugoni, Pasquali, Perani & Zorat, 1998; Hillis & Caramazza, 1990; Sacchett & Humphreys, 1992; Warrington & McCarthy, 1983, 1987, 1994). Recent summaries of relevant cases are provided by Caramazza (1998), Forde (1999), Forde and Humphreys (1999) and Saffran and Schwartz (1994). These cases raise general and important questions about the nature of our knowledge about objects, and its neural implementation. For example, does the brain represent knowledge of specific categories in discrete areas? Alternatively, does knowledge about categories take a distributed form, with particular forms of information ‘weighted’ for the recognition of some but not other categories? Might even distributed knowledge be structured in some way, with (for example) item-specific perceptual knowledge being differentiated from forms of conceptual and contextual knowledge concerned with the relations between objects? Might inter-object contextual knowledge be separate from knowledge of the actions performed with objects? How might such structural differences in knowledge representation impact on the recognition of particular classes of object? What might be the implications for understanding what is often termed ‘semantic memory’? This paper is concerned with these questions, addressed in the light of work on category-specific deficits for living and non-livings.

1.1: Category-specific confoundings

However, before we embark on discussing the category-specific deficits that emerge following brain damage, we need to ensure that any effects are indeed real and not an artefact arising from some other confounding variables. For instance, one possibility is that category-specific deficits simply reflect the background of the person pre-morbidly. People may have difficulty identifying animals because they knew little about them in the first place and consequently these items suffer most after brain damage; other people who knew little about tools may present with a category-specific impairment for nonliving things etc. A more elaborated account along these lines is that the deficits reflect differences in familiarity between objects. Living things are often less familiar (as a category) than nonliving things which we see and use everyday, and so an apparent impairment in accessing knowledge about living things may be due to a more general impairment in retrieving information about low familiarity items (Funnell & Sheriden, 1992). Another possibility is that object recognition is more difficult for more visually complex stimuli, taxing the limited visual processing resources of some patients (Stewart, Parkin & Hunkin, 1992). Since living things are typically more visually complex than nonliving things, patients may present with what appears to be a category-specific impairment for living things.

Investigators should rightly be concerned that any effects they observe are not confounded by stimulus familiarity or complexity, and it is the case that many of the early studies did not control for such factors. Nevertheless, it seems unlikely that these factors can account for all the dissociations that have been observed. For example, some patients show category-specific losses for objects that they are particularly familiar with. We have observed a deficit in naming fruits and vegetables in a patient who was a food expert and wine connoisseur (Humphreys, Francis & Samson, in preparation), and Michelangelo, a patient with impoverished knowledge for living things, was formerly an active member of the World Wildlife Fund and could identify large numbers of animals before his brain damage (Sartori & Job, 1988; Sartori, Miozzo & Job, 1993). Also, the deficits apparent in patients can occur with objects that are highly familiar to average members of the population (e.g., apples, dogs, cats) so it is difficult to attribute impairments with these objects to a lack of familiarity. In addition, a simple effect of one variable (like familiarity or visual complexity) cannot account for the double dissociation between losses of knowledge for living and nonliving things, and such dissociations still occur even when items are matched for familiarity and complexity across categories or when statistical measures are taken to rule out such effects (Farah, Meyer & McMullen, 1996; Forde et al., 1997; Kurbat, 1997; Kurbat & Farah, 1998; Sartori et al., 1993). We conclude that not all such deficits are due to confounding factors, and thus the study of these category-specific impairments can help inform us about the nature of our stored knowledge and the way in which different objects are recognised.

Perhaps the most obvious account of category specific deficits is that they reflect the categorical organization of our underlying knowledge about the world, which is partitioned according to whether stimuli are living or nonliving. We will use this as the ‘default’ account of the deficit, to be held at the back of the reader’s mind while other ‘non-categorical’ accounts are reviewed, though we will reconsider it following these reviews. Prior to this, we will evaluate whether the deficits reflect the loss of particular forms of knowledge that are not categorical in nature (Section 2) or whether they reflect the interaction between perceptual processes and the kinds of knowledge required to differentiate between objects for different tasks (Section 3). In Section 4 we return to the idea that stored knowledge might be categorically organized. In Section 5, we consider all the above arguments in relation to function imaging studies that have shown selective activation of neural areas as a function of both the object being presented and the task. In Section 6, we outline a framework for understanding both the psychological and the neuro-anatomical data which we term the Hierarchical Interactive Theory (HIT) of object recognition and naming,. The model is related to several accounts that have previously been considered (e.g., Damasio, 1989, 1990; Warrington & McCarthy, 1987), but differs in emphasising that object recognition proceeds through a hierarchical series of stages and that re-entrant activation plays a differential role in particular tasks (e.g., for naming rather than for object recognition). We discuss the relations between the model and others in the literature.

1.2: Defining semantic memory

Category-specific impairments have frequently been interpreted in terms of deficits in accessing semantic memory for objects. Consequently, such deficits have been used to argue about the nature of semantic memory (e.g., Warrington & Shallice, 1984). But what is semantic memory? In Warrington’s (1975) seminal paper, which was the first systematic investigation of patients with selective impairments of long-term stored knowledge about objects and their properties, the term ‘semantic memory’ was defined as "that system which processes, stores and retrieves information about the meaning of words, concepts and facts". Consistent with this, semantic memory is still considered by many to hold the information that allows us to give meaning to the objects we see and the words we read and hear. However, despite the general use of the term, or perhaps because of it, there has been little attempt to provide a more rigorous definition (though, see Caramazza, Hillis, Rapp & Romani, 1990). It is perhaps symbolic that semantic memory is represented as an under-specified ‘cloud’ in many standard models of cognition (e.g., see Morton & Patterson, 1980, Fig.’s 4.1-4.3). Now our semantic representation of a dog would generally be considered to include multiple facts such as: that it is an animal, it has 4 legs and a tail, it barks, it likes to chase cats, it is a ‘man’s best friend’, it was once featured in a sentimental song by Elvis Presley and so on. Thus, on this basis, semantic knowledge includes information about the general category of the item, visual information about it’s shape and parts, other sensory information (e.g., about the sound it makes, what it feels like), the relationship between it and other items, and general contextual knowledge abstracted even from any sensory information of the object itself (e.g., that the ballad ‘Old Shep’ refers to a dog). Given the diversity of the information considered to be ‘semantic’, and the fact that there are fundamental differences in the nature of the knowledge involved (i.e., some pertains to the sensory properties associated with an object [e.g., it is brown and has floppy ears] and some reflects how 2 or more objects relate to each other [e.g., dogs chase cats] etc.), it is perhaps surprising that all of this information is considered to be represented in one homogenous ‘store’. Indeed, even when attempts have been made to define subsets of semantic knowledge, this has typically led to dichotomous distinctions being made, such as that between visual and verbal semantics (see Riddoch, Humphreys, Coltheart & Funnell, 1988, for one review). One aim of this paper, through the review of category-specific deficits, is to argue that the concept of a ‘semantic system’, in any unitary sense, may be one of the victims of an attempt to define the nature of our stored knowledge in more detail. If the ‘semantic system’ retracts to no more than the form of knowledge recruited to perform a particular task – with this knowledge differing across tasks, then the ‘system’ becomes a fiction. For instance, there may be little more in common between different forms of semantic knowledge than there is between, say, different forms of sensory knowledge (visual, auditory). It may be more fruitful to specify the different forms of knowledge than to seek out some unifying principle across what turn out to be separable knowledge stores. However, since the notion of a unified semantic system has been at the heart of much of the research, we set out by using this term when we discuss the work. In this initial usage, semantic memory may be defined as the central knowledge store, for all input and output modalities, that contains information about the meaning of objects – much like a multi-modal dictionary that can both be accessed and that can express itself in a variety of different ways (from print, speech, and visual images as input, to print, speech, action as output).

2: The loss of particular forms of knowledge: senory and functional knowledge.

2.1: The sensory/functional distinction

Warrington and Shallice (1984) reported four case studies of patients who had particular problems in identifying living things. For example, the patient we described in the Introduction, JBR, was able to give precise definitions of nonliving things (compass and briefcase) but very poor definitions of living things (parrot and snail). In addition to his poor definitions of living things, JBR was only able to produce descriptions indicating identification for 6% of pictures of living things, but 90% of nonliving things. A second patient, SBY, named no pictures of living things but could produce correct descriptions for 75% of the pictures of nonliving things. Warrington and Shallice suggested that since the patients were impaired at accessing information from both words and pictures, the locus of their impairment must be within the semantic system, rather than a lower level visual recognition problem. They also noted that their patients were poor at defining some nonliving things, such as cloths and precious stones. An associated problem with these particular nonliving things is difficult to understand if the patients’ problems are confined to living things. Instead, Warrington and Shallice proposed that the deficits reflect loss of some but not all forms of semantic knowledge about objects - notably loss of sensory semantic knowledge. They argued that, in order to identify living things (fruit, vegetables or animals), retrieval of fine-grained sensory information was necessary. For example, they suggested that to distinguish between a raspberry and a strawberry, detailed information about color, size, shape and texture was necessary. In contrast, they suggested that recognition of nonliving things ‘depends crucially on determination of its functional significance’ (page 849). Quite what is meant here by functional significance is not clear. Our knowledge about the function of an object could include information about how an object is acted upon (e.g., by turning the wrist, for a spanner) or about how the object itself operates (e.g., a car operates by consuming petrol). As we hope to show, authors have used the term ‘functional’ in a variety of ways when referring to our stored knowledge of objects. The different forms of functional information, however, need not be equivalent, and may themselves be represented in contrasting ways.

Warrington and Shallice proposed that two independent systems may have evolved: one storing ‘functional’ information important for identifying nonliving objects, and the other storing sensory information important for identifying living things. They suggested that patients like JBR and SBY have an impairment to the visual/perceptual semantic system which leads to particular problems naming living things (see Figure 1). However, since the source of the problem is in a store specifying sensory rather than category-specific knowledge, the problem can also generalize to nonliving things that also depend on the retrieval of sensory knowledge for their identification - cloths and precious stones perhaps being two examples.

Warrington and Shallice also noted that their patients tended to be either consistently correct or incorrect when trying to identify stimuli and suggested that such a consistent deficit reflected degenerate semantic knowledge. Interestingly, the items that the patients were consistent on in one modality (say with pictures) were not necessarily the same as those for which they were consistent in other modalities (say when defining a word). From this they concluded that semantic knowledge is partitioned not only into independent modules for sensory and functional knowledge, but also for input modality. Thus there may exist sensory semantics for visual objects (pictures and real objects) and sensory semantics for verbal input (written and spoken words). Since these semantic systems are distinguished by input modality as well as by the information represented, there may well be duplication of sensory and functional knowledge within the modality-specific systems (e.g., the fact that a bee has yellow and black stripes will be represented in both the visual and verbal semantic systems specifying the sensory properties of objects) (see Figure 1; also McCarthy & Warrington, 1994; Warrington & McCarthy, 1994).

We return to this issue of modality in the General Discussion.

Figure 1

Figure 1. A model which distinguishes knowledge stores for sensory and functional knowledge, separately for visual and verbal input (after Warrington & Shallice, 1984). Note that, according to this account, patient JBR (Warrington & Shallice, 1984) – who has problems with living things when given both visual and verbal input - has functionally independent impairments in the visual/sensory and verbal/sensory semantic systems.

Warrington and McCarthy (1987) presented a rather different view of this sensory/functional dichotomy, which they discussed in terms of a distributed model of semantic memory. They drew upon ideas of how objects might be recognized put forward originally by Lissauer (1890). Lissauer suggested that "the recognition of an object can only occur when at the time of its perception a number of ideas are evoked which relate to that object. These bring into consciousness those characteristics which the mind has learned to associate with it and those conditions in which it has been previously experienced … memories laid down through different sensory modalities contribute to these associations but it is only when they are brought into awareness and linked with the percept that the recognition of an object becomes complete". Warrington and McCarthy (1987) suggested that these ‘memories laid down through different sensory modalities’ would not be stored in a homogenous semantic store or ‘module’ but in modality congruent ‘channels’. For example, visual information would be stored in a visual channel, ‘functional’ information in a motor channel, information about the sound an object makes in an auditory channel, and so on. Warrington and McCarthy suggested that these channels could be relatively fine-grained so that visual information might actually be stored in a number of sub-channels (e.g. for colour, size, shape). Furthermore, different channels of sensory or motor information would have different degrees of importance for different items. For example, as initially proposed by Warrington and Shallice (1984), perceptual/sensory information would be important for the identification of many living things whereas ‘functional’ information, defined to include motor actions, may be crucial for the identification of nonliving things. The specialization of different parts of the system for particular objects would lead to a quasi-categorically organised knowledge base.

This account predicts that quite fine-grained category-specific impairments may occur because different categories within the living and nonliving groups (e.g. animals, tools, fruit, clothing) would have different patterns of weighting across the channels. For example, they suggested that accessing colour knowledge might be important for recognizing fruit (e.g. distinguishing between a raspberry and a blackberry) but accessing shape information may be relatively more important for distinguishing between two flowers (e.g. a daffodil and a tulip). Thus patients may have deficits with subsets of living or nonliving things, for example, with fruit and vegetables but not animals, or with tools but not clothing. These more selective patterns of deficit have been observed (e.g., Hart, Berndt & Caramazza, 1985, for a deficit with fruit and vegetables; Caramazza & Shelton, 1998, and Hart & Gordon, 1992, for a deficit with animals; McCarthy & Warrington, 1987, for a deficit with tools). In addition, patterns of association are expected between the loss of certain forms of knowledge and deficits with particular objects. For example, if a patient had a specific impairment in naming fruit this ought to be accompanied by an impairment in retrieving the colour of objects from memory, since colour knowledge is likely to be important for distinguishing between different fruit (see Price & Humphreys, 1989, for evidence with normal subjects). However, Luzzatti and Davidoff (1994) reported two case studies of patients who had a marked impairment at retrieving the color of objects, but no particular problem in naming fruit and vegetables. Luzzatti and Davidoff argued that an impairment in retrieving object-colour knowledge did not necessarily impair naming performance for categories of living things, such as fruit and vegetables.

An attempt to capture some of the properties of the sensory-functional distinction in a more formal model was made by Farah and McClelland (1991). They simulated semantic representations using a distributed associative memory system (cf. McClelland & Rumelhart, 1985). Items were represented in terms of patterns of activation across processing units corresponding to either the perceptual-sensory or functional properties of objects. The number of sensory units assigned a non-zero value in the coding of a stimulus, relative to the number of functional units, differed for living and nonliving things. For living things about seven times more sensory than functional features were active; for nonliving things, there were equal proportions of active sensory to functional units. This differential weighting of the representations of living things was based on the number of properties generated by subjects when asked to mark the visual and functional attributes in dictionary definitions, and far more visual attributes than functional attributes were marked. In contrast, roughly equal numbers of visual and functional attributes were marked for nonliving things. The model was then trained to associate the sensory and functional properties of objects with representations of their visual attributes and names. Performance on object naming was tested by giving the visual attributes as input and seeing whether the correct name was generated as output. The model was also lesioned, with sensory or functional units being differentially affected. Lesions affecting the sensory units lead to marked impairments in identifying living things; lesions affecting the functional units affected non-living things more than living things. In addition, because of the distributed nature of the semantic system, lesions of the sensory units also produced some loss in retrieving the functional properties of stimuli, with living things being affected more severely.

Farah and McClelland’s simulation provides an existence proof that a model of this form, with distributed sensory and functional knowledge about objects, can produce apparent category-specific deficits when lesioned. It also shows that a form of double dissociation can be generated, with either living or nonliving things being affected - depending on which form of knowledge is impaired. We consider this argument further when we discuss category-specific deficits for nonliving things. However, such an existence proof of course does not demonstrate that the same underlying architecture exists for the stored knowledge of objects in humans. Also, Farah and McClelland’s procedure for estimating the sensory and functional properties of objects has been criticised (see Caramazza & Shelton, 1998). Farah and McClelland asked subjects to mark the functional attributes of objects by asking the questions; "what does the item do or what is it used for?" Caramazza and Shelton suggest that this question is biased to nonliving things and leads to underestimates of what might be termed the functional properties of living things (here Caramazza and Shelton refer to non-sensory properties such as ‘lives in a desert’, and ‘carnivore’). When all non-sensory properties are noted in dictionary definitions then the bias for more sensory properties for living things is greatly reduced (Caramazza & Shelton, 1998; Moss, Tyler & Devlin, in press).

In the distributed associative memory framework used by Farah and McClelland, category-specific deficits for living things should generalise to include poor retrieval of the non-sensory properties of these items (although this problem should be less severe than the impairment in retrieving sensory properties). This runs into two difficulties. Firstly, there are patients whose problems in the retrieval of non-sensory attributes of living things is at least as severe as their problems in retrieving sensory properties (Caramazza & Shelton, 1998; Laiacona et al., 1993; Sheridan & Humphreys, 1993). Indeed, this pattern of an equal deficit for sensory and non-sensory knowledge of living things was observed in one of the original Warrington and Shallice cases, when re-examined by Funnell and de Mornay-Davies (1996). Secondly, there are some patients who show extremely good (normal) retrieval of the non-sensory properties of living things whilst being impaired at retrieving their sensory properties (Forde et al., 1997; Hart & Gordon, 1992; Humphreys, Riddoch & Price, 1997; Riddoch & Humphreys, 1993). This pattern, in which non-sensory knowledge can be affected to as severe a degree, to a less severe degree or not at all, suggests that the sensory and non-sensory (‘functional’) properties of living things are not as tightly coupled as in Farah and McClelland’s architecture. The results also have implications for accounts of category-specific deficits in terms of correlated sensory and functional features (Section 2:2) and in terms of a semantic system that is truly categorical in nature (Caramazza & Shelton, 1998) (Section 4).

2.2: Correlated sensory and functional features

We have suggested that a simple distributed memory account, which separates sensory from non-sensory knowledge fails to capture the full pattern of deficits for these two forms in patients. Other investigators have suggested that the difference between living and nonliving categories lies not in the relative importance of perceptual or functional attributes for identification per se, but in the links between them. For example, for nonliving things the connection between shape and action is not arbitrary, since the shape of the item is typically constructed in a way that will best perform the action intended. De Renzi and Lucchelli (1994) argued that this link between visual and functional (in the sense of action) properties for nonliving things makes these items less vulnerable when stored sensory information is degraded. They outlined a case study of a patient who had difficulty in recognising living things and in performing object decisions, drawing and describing the perceptual differences between living things. Interestingly, their patient also had problems with nonliving things when the tasks involved retrieving the colour of items and naming objects from sound. De Renzi and Lucchelli proposed that the naming impairment for living things resulted from a general failure to retrieve the perceptual features of objects from every category (and not just the category of living things). However, for nonliving things this deficit may be compensated for by the close links between visual attributes and function, which provide an alternative route to accessing the representation of the object. De Renzi and Lucchelli acknowledged that "it remains to be explained how the function of an object can be inferred from its visual appearance, if this has not been recognized or retrieved from memory" (page 19). However, they suggested that for nonliving things, a sensory and a functional semantic store could interact. Thus, in visual object identification, functional cues could "help specify hypotheses on the nature of the stimulus that were left undefined by visual processing" (page 20). For living things they proposed that this route would be unavailable since, for these items, they assumed there to be few links between perceptual and functional attributes. In addition, they suggested that the retrieval of colour knowledge and identification via sound would be impaired for all categories because these perceptual properties have no direct functional associations.

Another account that stresses the importance of the interaction between sensory and functional properties of objects is the OUCH model of semantic memory (Caramazza et al., 1990). This model does not differentiate between different types of stored knowledge; both perceptual and functional (action-related) properties are said to be stored within a single semantic system. OUCH states that, during visual object processing, salient parts of objects directly activate corresponding perceptual and functional attributes in semantic memory. For example, salient parts of a fork (e.g. the prongs and handle) directly activate corresponding semantic knowledge (i.e. about prongs and handles in general) and this pattern of activation could then be used to generate hypothesis about what the object might be. This direct activation of action-related information from the parts of the object leads to privileged access to semantic memory for objects relative to words (cf. Potter & Faulconer, 1975). According to this account, damage to semantic memory may lead to a more severe impairment for living relative to nonliving things, if nonliving things can benefit from a higher degree of correlation between their sensory and functional features.

Both De Renzi and Lucchelli’s account and the OUCH model emphasise the importance of correlations between sensory and functional properties of objects. Other authors, however, stress that the correlations within sets of sensory and/or sets of functional features may also play a predictive role in retrieving information from semantic memory. McRae, de Sa and Seidenberg (1997) had subjects list the features of objects from categories of both living and nonliving things. They found that, whilst sensory properties were listed with roughly equal frequency for living and nonliving things, there were significantly more non-sensory features (e.g. used for carpentry, worn by women) for nonliving things (supporting Warrington and Shallice’s (1984) model of semantic memory). Furthermore, living things tended to have more correlated features than nonliving things (11% of feature pairs relative to only 6% of feature pairs for nonliving things). Living things were more densely represented across the correlated feature pairs - so that a smaller set of common features captured more of the properties of living things. They suggested that correlated features play a differential role in recovering information about living and nonliving things; for living things, correlated features could lead to robust recovery of a common set of core attributes (e.g., is animate, eats, breathes) but greater difficulties in individuating objects. This proposal can account for some patterns of degenerative performance found in patients with Alzheimer’s disease (AD). Gonnerman et al. (1997) reported that in the early stages of AD, living things could be identified better than nonliving things, but as the disease progresses the pattern reverses and living things are relatively more difficult to identify (see also Devlin et al., 1998). The greater number of correlated features for living things may help to protect the identification of these items from small amounts of generalised brain atrophy (as in the early stages of AD). With more widespread damage though, features are lost and performance with living things may decrease catastrophically, as the lost features contribute to many exemplars within the category. Nonliving things, having relatively fewer correlated features, are vulnerable to damage on a more individual basis, but then show a less catastrophic loss as the disease progresses. This differential pattern of decline was simulated by Devlin et al. using a Hopfield network with interconnected semantic units that tended to ‘push’ an input pattern into a stable activation state. Interestingly, in this model, category specific deficits emerged even with random non-selective damage - as might be assumed to occur in AD (though see Perry, 1999, for a discussion of the limitations of the model). In models such as that of Farah and McClelland (1991) lesions had to affect either the sensory or functional features differentially to generate a category-specific effect.

However, the data on patients with AD are not clear cut. For example, Silveri, Daniele, Giustolisi and Gainotti (1991) found that AD patients with moderate deficits were worse with living than with nonliving things (see also Garrard, Patterson, Watson & Hodges, 1998; Giustolisi et al., 1993) - the opposite result to that reported by Devlin et al. (1998) and Gonnerman et al. (1997). Gonnerman et al. in fact found a consistent deficit for living things in one patient even when identification performance for nonliving things remained at a high level, which is not consistent with their group study. It is not clear that accounts in terms of correlated features will be able to provide a framework for these disparate results. Garrard et al. (1998) also query how, in neural terms, inter-correlations between representations might protect features from degenerative decay. They suggest that the differences between the majority of degenerative patients who have problems with living things, and the minority who can be found with deficits for nonliving things, related to differences in the initial area of neo-cortical involvement. In the majority of cases there is transfer of the disease from the transentorhinal region to temporal neocortex. In the minority there may be bi-parietal involvement. The contrast between the two sets of patients may reflect the storage of perceptual features in temporal cortex and more action-based (functional) features in fronto-parietal regions (see also Gainotti, Silveri, Daniele & Giustolisi,1995).

An extension to the proposal that correlated features are important has been made by Moss, Tyler and colleagues (Durrant-Peatfield, Tyler, Moss & Levy, 1997; Moss, Tyler, Durrant-Peatfield & Bunn, 1998; Moss, Tyler & Jennings, 1997; Tyler & Moss, 1997). They point out that, for living things, many of the correlated perceptual features are associated with common biological functions (such as breathing, eating and reproducing). The distinctive perceptual features of living things, however, are not strongly correlated with this kind of functional information (e.g. a tiger’s stripes; see Keil, 1992). It is in this last respect that living and nonliving things differ. Nonliving things have distinctive perceptual features that are correlated with their function in terms of action (e.g. the serrated edge of a saw, the prongs of a fork). This last point is similar to the argument made by De Renzi and Lucchelli and the OUCH model (see above), the difference being that Moss, Tyler and colleagues highlight that, for living things, common (inter-correlated) perceptual features are associated with functional properties whilst for nonliving things, functional attributes are associated with distinctive perceptual properties. After brain damage the features strongly associated to the functional properties of objects may be better preserved than those with weak associations. This will have different consequences for living and nonliving things. For living things, information about biological function will be recovered from the linked, common perceptual features. However, this will not help the identification of individual stimuli, for which distinctive features are important. For nonliving things, distinctive visual features are correlated with function; identification of individual items is thus better. Durrant-Peatfield et al. (1997) report simulations of these patterns in a feedforward connectionist model trained to associate an input pattern corresponding to ‘perceptual’ and ‘functional’ properties of objects to a matching output pattern (‘auto-associative’ learning). The training set was varied so that, for living things, common perceptual features co-occurred with common functional properties; for nonliving things, distinctive perceptual features co-occurred with distinctive functional properties. When connections between input and output units were randomly disconnected, the identification of living things tended to be more affected, though recovery of their shared functional properties was relatively well preserved.

The account put forward by Moss, Tyler and colleagues is able to explain why information about biological function can be preserved in patients with poor identification of living things (Moss et al., 1997; Tyler and Moss, 1997). For example, in one case reported by Moss et al. (1998) a patient with impaired naming of living things was nevertheless able to group these items according to their shared properties (does it have legs? Does it lay eggs?). He was poor at retrieving distinctive properties of living things, whether visual or functional (e.g., properties concerned with survival). Retrieval of shared category properties for nonliving things was, if anything, worse than retrieval of similar properties for living things, perhaps because of an inherent advantage for living things in such tasks (due to the shared functional properties being based on shared perceptual properties, for living things; see section 3.1). Indeed common information for living things, whether concerned with biological function or general category, is accessed rapidly also by normal subjects (Humphreys et al., 1997; Tyler & Moss, 1997).

We conclude that differences between shared and distinctive features, and the degree to which these features correlate with the function of the object, are likely to be important contributing factors in category specific impairments. However, as we hope to demonstrate, models need to be elaborated further in order to account for the full pattern of dissociations that have been documented. In particular, models need to have a more articulated structure, specifying different forms of stored knowledge and different stages of object identification. HIT, which differentiates both the different forms of knowledge representation and the contrasting stages of object identification, provides a framework that can allow for a fuller account of the different patients in the liaterature. This is described in Section 6.

3: Interactions between perceptual processes and knowledge for particular tasks.

3.1: The Cascade model

When normal subjects are asked to list the parts of objects, living things tend to be listed as having proportionately more shared parts than nonliving things. Similarly, when the outline contours of standardised drawings of objects are compared across category exemplars, living things tend to have higher levels of contour overlap than do nonliving things (see Humphreys et al., 1988). These different indices provide an approximate measure of the similarities of the perceptual structure of objects within their categories, with living things having more similar structures than nonliving things. Humphreys et al. (1988) termed this within-category property ‘structural similarity’. Differences in structural similarity between living and nonliving things may contribute to the differences in identification that can be observed between these categories of object. To illustrate, consider the ‘Cascade’ model of visual object recognition outlined by Humphreys et al. (1988). This is composed of several stages including visual recognition of an object’s structure (access to stored structural descriptions), access to semantic information and access to the object’s name. Stored structural descriptions are held to represent information about the shape of objects but not to include other information such as an object’s use or its association with other objects. In this case, the term 'semantic memory' was reserved to apply only to these latter forms of (non-perceptual) knowledge. If activation can be passed onto one stage before processing at an earlier stage is completed (i.e., if processing operates in cascade), then differences in structural similarity will directly affect semantic access and name retrieval when objects are presented visually. Structurally similar objects will activate the structural representations of perceptual neighbours across their category. As a consequence, functional and associative information common to the category is derived quickly, but there is then increased competition between category exemplars for individual identification. Structurally dissimilar objects will activate fewer perceptual neighbours, so that activation of functional and associative information will be slower and less widespread. Nevertheless, individual identification should be more efficient (e.g., in a naming task), since competition from perceptually and functionally similar neighbours will be reduced (see Humphreys et al., 1988, 1997). Note that this account of rapid access to common functional and associative information, along with slowed access to identity information due to within-category structural similarity for living things, is in many ways similar to the proposals concerning common and distinctive features made by Moss, Tyler and colleagues. In the Cascade model, though, these ideas are tied to an architecture specifying the different stages involved in object identification. We will argue that this provides important explanatory power in accounting for the variety of category specific deficits in patients, and it remains a feature of the HIT approach which we outline in the final section.

The Cascade model predicts differences in performance for living and nonliving things even in normal subjects. Consistent with this, Humphreys et al. (1988) found that normal subjects named pictures of living things more slowly than pictures of nonliving things matched for familiarity and name frequency (see also Lloyd-Jones & Humphreys, 1997; Snodgrass & Yuditsky, 1996). When access to stored structural descriptions is measured using object decision, the benefit for nonliving things remains but is reduced (Lloyd-Jones & Humphreys, 1997). According to the model, the larger difference between living and nonliving things in naming, relative to object decision, is a result of the small differences in the efficiency of accessing structural descriptions being exacerbated by competition accruing from common (overlapping) functional and associative representations being activated. However, category decisions are faster for living things compared to nonliving things (Humphreys, Price & Riddoch, 1999; Riddoch & Humphreys, 1987a). This advantage for living things is larger when stimuli are presented as pictures rather than words, though it still exists with words (Job, Rumiati & Lotto, 1992). Since the effect is larger with pictures, we suggest that it is not simply due to category information being more closely linked to living things; rather there is privileged access to common semantic (functional and associative) information from the visual properties of living relative to nonliving objects. Advantages for identifying nonliving things over living things in fact occurs not only with normal human subjects but also with monkeys! Gaffan and Heywood (1993) trained monkeys to make discrimination responses to pictures of living and nonliving things. They found that the monkeys took longer to learn the responses associated with living things, with the effect increasing as the number of stimuli in the set increased. This work provides converging evidence for the idea that living things have high levels of perceptual overlap.

A simulation of the results on human identification was reported by Humphreys, Lamonte and Lloyd-Jones (1995). They used an interactive activation and competition framework with pools of units representing structural, semantic (functional and associative) and name information about objects (see Figure 2). Name units were further divided to represent either specific or superordinate category names. Input activations, given to the structural descriptions, were based on the rated perceptual similarity between items within their categories; for example, a ‘dog’ as input would maximally activate its own structural description (activation 1), but it would also activate the structural descriptions of other, similar items (e.g., activating descriptions for a fox, a cat, a sheep with values of .6, .5, .3 etc.). Living things were rated as perceptually more similar than nonliving things, and so shared more activation values between their category members. The net effect of this difference in shared activation values was that there was rapid access to superordinate names for living things but slowed access to individual names (due to increased competition from multiple activations at a semantic level). This model was also ‘lesioned’ by having noise added either to activations at a structural description level or to the weights connecting structural to semantic, or semantic to name representations. After lesioning, the identification of living things was selectively disrupted relative to the identification of nonliving things; essentially this occurred because of the increased structural and functional/associative overlap generated during the visual processing of living things. In contrast to the effects on identification there was relatively little effect of lesioning on access to superordinate information; indeed, if anything the disruption affected performance with nonliving things more. These data from the lesioned model match the results from many of the patients with category-specific deficits with living things.

Figure 2

Figure 2. The interactive activation and competition model of object naming proposed by Humphreys et al. (1995). -à indicates excitatory links and ---o inhibitory links. This model incorporates two of the basic principles of the HIT framework: functionally isolatable subsystems (e.g., separating perceptual knowledge about objects [in the structural description system] from other forms of stored knowledge), and interactive processing (top-down as well as bottom-up). In a more detailed account, the semantic representations as specified here are further divided and represented in a distributed form.

One of the interesting results in these studies of simulated lesions was that the identification of living things was impaired even when noise was only added to the connections between the semantic and name units (i.e., at the ‘output’ end of the model). In this last case, the impairment still arose even when access to structural and semantic information was intact (though there was still the natural increase in competition for access to the names of living things). This result is of interest because it matches the pattern reported in a few patients with category specific impairments for living things. For example, Farah and Wallace (1992) and Hart, Berndt and Caramazza (1985) have both reported problems in patients that seem confined to name retrieval. In both cases, the patients could retrieve information about living things once they were given their names, and Farah and Wallace report that their patient could categorise fruit and vegetables even though the naming of these items was selectively impaired (see also Forde et al., 1997; Humphreys et al., 1997). Such an apparent selective problem in name retrieval is difficult to account for in models that do not have a distinct (and dissociable) stage of access to name information following access to semantics. Nevertheless, despite the success of the Cascade model in accounting for such naming disorders, we shall reconsider the evidence on this point in Section 6, when we introduce the HIT model of object identification.

3.2: Depth of processing within the structural description system

The Cascade model predicts that category specific impairments for living things can emerge following damage to the structural description system, though this description system itself is not categorically organised. According to the model, damage to structural representations may lead to problems in differentiating between living things because these stimuli belong to categories with large numbers of structurally similar exemplars. Category- specific impairments at the level of the structural description do not reflect any functional distinction based on biological category but emerge because of the different processing demands of living and nonliving things. However, Sartori and Job (1988) offered an alternative suggestion and proposed that the structural description system might be categorically organised. They presented a case study of a patient, Michelangelo, who was significantly worse at naming living compared to nonliving things. In addition, Michelangelo performed poorly on object decision for living things, though performance with nonliving things was within normal limits. Object decision is typically used to test the integrity of the structural description system, since it only requires subjects to assess whether or not they have seen a particular ‘visual pattern’ before and does not require them to access any further (functional or phonological information) information. Consequently, Sartori and Job (1988) suggested that Michelangelo’s category specific impairment resulted from damage to the structural description system. They suggested further that the particular problem with living things could arise because (1) the system was categorically organised or (2) living things require more detailed processing at this stage within the object recognition system (as suggested in the Cascade model, Humphreys et al., 1988). To distinguish between these two alternatives they asked their patient to specify the important differences between 12 pairs of animals, 12 pairs of vegetables and 12 pairs of objects. Michelangelo scored at ceiling for objects but outside the control range for vegetables and animals. Sartori and Job suggested that if Michelangelo had a general impairment in accessing perceptual attributes from stored structural descriptions he should have problems generating the important perceptual differences between objects, regardless of category. Since he only had problems with living things, they argued that these must be represented in a functionally independent compartment. However, if living and nonliving things are represented in terms of contrasting amounts of overlapping activation within the structural description system, then we see no reason why differences should not emerge between the categories even when the system is addressed from the spoken names. Increased overlap for living things should still make their perceptual properties relatively difficult to retrieve. Furthermore, both Michelangelo and controls sometimes gave perceptual attributes and sometimes gave functional attributes for items from all categories, when contrasting the differences between stimuli. Controls tended to give functional attributes for objects and perceptual attributes for animals and vegetables. Consequently, if Michelangelo had intact functional information but poor perceptual information for all categories, he would be able to score more highly with objects compared to the other two categories, as Sartori and Job observed. Stronger evidence for the idea that the structural description system is categorically organised would be reports of patients who have impairments to the structural description system that selectively affect nonliving things; this pattern remains to be documented.

More recently, Sartori, Job, Miozzo, Zago and Marchiori (1993) have used Marr’s (1982) model of visual object recognition to account for category specific deficits for living things. According to this model, structural descriptions of objects are represented in a hierarchical form. At the top of the hierarchy descriptions correspond to general category exemplars, with the descriptions becoming increasingly detailed further down the hierarchy eventually distinguishing items at a subordinate level. Sartori et al. suggested that patients with category specific impairments for living things may have problems accessing the lowest (most detailed) levels of these structural hierarchies, and argued that living things were most affected because they have ‘deeper’ representations than nonliving things. Note that this idea is similar to the view outlined by Humphreys et al. (1988), who also argued that living things require more fine-grained or ‘deeper’ processing at the level of structural descriptions. For Humphreys et al. (1988), ‘deeper’ processing is an emergent consequence of overlap between the structural descriptions of stimuli.

3.3: Shape processing differences across categories

The Cascade model emphasizes the importance of perceptual overlap between exemplars within a category as a factor that can lead to processing differences between living and nonliving things. However, the approach suffers from a failure to specify in detail the nature of the perceptual overlap that may be critical. Attempts to overcome this was made by Arguin, Bub and colleagues (e.g., Arguin, Bub & Dudek, 1996; Dixon, Bub & Arguin, 1997). They used computer generated stimuli, derived from variations in elongation, tapering and bending, to resemble real objects. For example, a banana can be described as having positive values on elongation and bending, but a zero value on tapering; an orange would have zero values on all of these dimensions. Arguin et al. reported data from a patient, ELM, who showed a category specific deficit for living things following two separate strokes. One task involved presentation of four shapes, one in each quadrant of a display, followed by presentation of one of the shapes at the centre. ELM had to point to which of the four locations had been occupied by the subsequent target. When the items on a trial differed along a single dimension the patient performed significantly better (29% errors) than when the items varied along two dimensions simultaneously in the conjunction condition (57% errors). Arguin et al. proposed that ELM failed to extract information from two visual dimensions simultaneously, and tended instead to attend to just one dimension (so performing better when stimuli only varied along a single dimension).

Arguin et al. also went on to show that this deficit in extracting multiple visual dimensions combined with effects of semantic similarity between items. ELM was required to label the same shapes either as particular fruit and vegetables or (in another block) as nonliving things. With the semantically close living things, ELM again performed worse in a conjunction condition relative to a single dimension condition. However, with a set of unrelated nonliving things, there was no difference between performance with conjunction and single feature sets. The deficit for learning conjunction stimuli paired with semantically close living items also generalized to faces (see Dixon et al., 1997). Arguin et al. argued that visual processing deficits, concerned with whether patients can extract several visual dimensions simultaneously, can combine with semantic similarity between items to create problems in identifying particular objects. To link these problems to a deficit with previously known living things, we must assume that living things, more than nonliving things, share both semantic and visual features; in addition, the visual features of living things may also vary along more than one dimension simultaneously. Support for this first assumption comes from data on normal picture naming under deadline conditions. Normal subjects make a greater range of errors that are both semantically and visually related to targets for living things than for nonliving things (Vitkovitch, Humphreys & Lloyd-Jones, 1993). Greater overlap in access to semantic as well as visual representations is also supposed by accounts that point out the importance of correlated perceptual features for living things (Gonnerman et al., 1997; Moss et al., 1998).

These results on learning feature and conjunction stimuli indicate that differential performance can be evoked with the same set of shapes, according to whether the shapes map onto semantically close or distant representations. It is also possible that contrasts in visual similarity within a set of items also contribute to performance differences, since the semantically close sets used are also very often perceptually similar (e.g., Dixon et al., 1998, showed a deficit for ELM when learning an association between conjunction stimuli and labels for different makes of car – nonliving things within a perceptually close sub-ordinate category). However it remains an open question whether data derived from a small set of shapes, and tasks that require new learning rather than the retrieval of previously learned information, can be generalised to account for deficits across the broader class of living things. It may be that, within this broad class, perceptual overlap (as well as covariance along multiple dimensions) combines with semantic similarity to create particular problems in identification.

3.4: Visual processing deficits without loss of stored knowledge

As we will document in Sections 4 and 6, many patients with category specific deficits seem to have impairments of stored knowledge, with there being impaired access to associative and functional (Section 4) or perceptual knowledge (Section 6) for the affected items. However, category specific deficits have also been reported in patients for whom there is no apparent loss of stored knowledge. For example, HJA is an agnosic patient with a severe problem in recognising many visually presented objects. The deficit is worse for living than for nonliving things, even when these items are matched for name frequency, familiarity and visual complexity (Riddoch & Humphreys, 1987b; Riddoch, Humphreys, Gannon, Blott & Jones, 1999). Nevertheless, when initially tested HJA showed good stored knowledge for living things, being able to provide detailed definitions and drawings from memory. The only clear problem in stored knowledge occurred when he was asked to retrieve colour information, though this deficit occurred for both living and nonliving things. In his case, it is difficult to attribute the category specific recognition deficit to loss of stored perceptual or functional features; rather it appears to reflect a problem in differentiating between items that have many close perceptual neighbours (i.e., living things relative to nonliving things). This hypothesis fits with the variety of other visual perceptual impairments in this patient (Riddoch & Humphreys, 1987b).

Interestingly, when retested some 16 years after the stroke that generated the recognition problems, HJA showed some deterioration in his stored knowledge for the perceptual properties of objects. His drawings from memory were more difficult for independent observers to identify and his definitions listed fewer visual attributes; the decrease in the number of visual attributes listed was more marked for living than for nonliving things (Riddoch et al., 1999). However, there was not a general decrease in HJA’s performance; he produced more non-visual attributes in his definitions when retested. These data suggest that on-line visual processes interact with memory processes to maintain stored representations of the visual features of objects. Over the longer term, a visual processing deficit leads to some degradation of these memory representations. This may affect representations of living things more because representations of visual attributes need to remain finely tuned to differentiate between these (perceptually similar) items and/or because visual attributes are strongly weighted in our stored representations of living things (Section 2).

Changes due to learning over time may also explain discrepancies in the performance of one of the first patients reported with category-specific deficits by Warrington and Shallice (1984). In the original study, patient JBR did not show a consistent pattern of deficit across modalities for items probed with pictures and with words, though he was consistently impaired on particular items within each modality. This result has implications for the issue of modality-specific representation of stored knowledge, which we return to in section 6.6. Funnell and de Mornay Davies (1996), however, found that JBR did show consistency across modalities when re-tested some years later. Here learning may have helped him re-establish links for certain items, though there was not a general learning effect as he remained consistently impaired on some items across modalities.

Several other studies, though, cast doubt on the necessary relationship between impaired visual and/or structural processing and category-specific deficits. For example, Humphreys and Rumiati (1998) reported data from a patient with suspected Alzheimer’s Disease, whose problem seemed to reside in poor perceptual knowledge for objects. Though this patient performed well at perceptual match tasks, she was impaired at object decision and on tasks requiring that objects be matched associatively. She also made visual naming errors when objects were misidentified. Despite this, there was no evidence of a category-specific deficit between living and non-living things. Lambon Ralph, Howard, Nightingale and Ellis (1998) also documented a patient with poor ability to match objects to definitions stressing visual as opposed to functional properties of objects; in addition there was impoverished production of perceptual features in her own definitions. Lambon Ralph et al. found trends for these problems with perceptual information to be worse for nonliving than for living things. Naming also tended to be worse for artefacts. This patient did show good performance on object decision, though, suggesting that there was a problem in interrogating perceptual knowledge from verbal input rather than there being an impairment of perceptual knowledge per se. Hart and Gordon (1992) too reported a case with poor retrieval of perceptual knowledge from names, along with good access from vision, though here the problem was more pronounced for living things.

The contrast between patients who show a deficit for living things due to impaired perceptual/structural knowledge, and those who do not, may be due to differences in the information the patients can draw upon to support performance. The patient of Humphreys and Rumiati, for instance, showed a priming effect on object naming when presented with multiple stimuli from the same category. This is consistent with there being partial activation of associative knowledge which could boost impaired visual identification when stimuli are presented in an appropriate context. Humphreys and Rumiati proposed that structural descriptions were activated below threshold level unless activation was increased top-down from members of the same category. For patients able to use such top-down activation, living things may benefit due to their inter-correlated common features; for some patients there may even be relatively strong category activation when single objects are presented reducing any advantage for nonliving things. Could this also lead to a reverse effect, with nonliving items being the more difficult? Lambon Ralph et al. in fact reported a perceptual deficit along with poor matching and production of definitions for nonliving things. In evaluating this, though, it should be borne in mind that, in studies measuring control performance, perceptual definitions of nonliving things are often harder to identify than perceptual definitions of living things (see Humphreys et al., 1997). This may be expected if perceptual information is weighted less strongly than functional (e.g., action-related) information in the stored representations of such stimuli. A patient may perform relatively worse on tests of this kind with nonliving things, then, due to a general rather than a category-specific decrease in their performance. A test such as drawing from memory, where detailed perceptual information must be retrieved, may enable an impairment with living things to be detected (as indeed was reported by Hart & Gordon, 1992).

One other possibility, not tested to date, is that differences between the recognition of living and nonliving things could reflect the ease with which patients can use particular perceptual representations. For example, the agnosic patient HJA has a perceptual deficit which still allows him to derive global shape descriptions, but these representations are not elaborated with local perceptual detail (e.g., see Boucart & Humphreys, 1992; Humphreys, Riddoch & Quinlan, 1985). Global shape information overlaps across many living things, so that a patient using unelaborated global representations may tend to find these objects especially difficult to identify. Other patients, however, may extract more local detail about the parts of objects and not about the global shape. The parts of nonliving things can have a functional role in their own right (e.g., a wheel of a bicycle), and a patient using local details with such stimuli may sometimes identify such parts as if they are the whole object: a problem we have observed in patients with simultanagnosia. This tendency to identify parts as wholes may be reduced for living things, because either their parts are not functional in their own right or they are diagnostic of the whole (e.g., an animal’s head). Patients who ‘weight’ parts more than wholes may find identification more difficult for nonliving than for living things.

As a final point here we note that the prediction, that impaired structural knowledge should necessarily disrupt living things more than nonliving things, turns out not to be straight forward – even without taking into account the distinction between global and local object coding. In simulations of the HIT model, we have found that the form of lesioning itself is critical for predicting whether a category-specific deficit emerges for living things after damage to ‘early’ stages of the model. We return to this point in section 6.

3.5: Conclusions

The research reported in this section highlights the importance of visual factors, such as perceptual overlap and covariance in multiple visual dimensions, for at least some aspects of category specific identification impairments. These visual factors may interact with stored representations of visual features (e.g., overlap within a structural description system) and with the semantic retrieval process, to make identification difficult for living things. According to the Cascade model, the combined visual and semantic effects can lead to identification deficits when lesions affect different levels in the object recognition system (access to structural descriptions, access to associative and functional knowledge, and access to object names; Humphreys et al., 1995). Visual deficits seem sufficient to produce ‘category specific’ impairments, but, at the same time, they may not be necessary.

4: Category specific knowledge

In Sections 2 and 3, we have discussed evidence suggesting that category specific deficits reflect factors other than the categorical nature of our stored knowledge (such as the importance of visual or functional information for representing different categories, or differences in perceptual and semantic overlap between categories). However, the most straightforward account of such deficits is that they arise because our stored knowledge is differentiated according to category of object. We have reviewed Sartori and Job’s (1988) claim that stored visual/perceptual knowledge within the structural description system may be categorically organised, and, although we concluded that there was (currently) little empirical support for this idea, we have not discussed the possibility that other types of knowledge are categorically organised. According to the Cascade model visual/perceptual knowledge in the structural description system is functionally independent from other types of semantic information (e.g., functional information or inter-object associations) and it is possible that this nonperceptual knowledge is categorically organised. We now review this hypothesis in more detail, asking whether this proposal is necessary to account for at least some forms of category specific disorders.

4.1: Category specific impairments for nonliving things

One argument for semantic knowledge being categorically organised is that some patients can have deficits for nonliving things rather than living things. Nielsen (1946) first reported this in a patient, CHC, who was able to recognise living things, such as faces and flowers, but not nonliving things, such as a car, a hat or a telephone. More recently, Warrington and McCarthy (1983, 1987, 1994) outlined three case studies of patients who had particular problems recognizing nonliving things. These patients performed significantly better with living things compared to nonliving things on matching to sample tasks. For example, in spoken word/picture matching, patient VER was better with flowers (93%) and animals (86%) compared to nonliving objects (63%) (Warrington & McCarthy, 1983); patient YOT was better with animals (86%) and flowers (86%) compared to nonliving objects (67%) (Warrington & McCarthy, 1987); patient DRS was better with animals (95%) compared to nonliving objects (74%) (Warrington & McCarthy, 1994). Sacchett and Humphreys (1992) further demonstrated that category specific impairments for nonliving things can remain when confounding variables such as frequency and familiarity are controlled. Their patient, CW, was significantly better at naming line drawings and at performing picture-word matching tasks for living things (despite the fact that they were less familiar, more visually complex and matched on frequency to the nonliving things). Quite similar data were reported by Hillis and Caramazza (1991), who showed a double dissociation between naming living and nonliving things using the same set of items with two patients. Silveri et al. (1997) also report data from a patient impaired at naming nonliving things using the same stimuli as had been used to elicit a deficit for living things in patient Michelangelo, reported by Sartori and Job (1988). Such double dissociations are consistent with a fractionation of semantic knowledge between that required for the identification of living things and that for the identification of nonliving things.

However, other accounts can be offered for the deficits with nonliving things. Warrington and McCarthy (1987), for example, suggested that impairments could occur at a finer-grained level than the living/nonliving dichotomy. The patient documented in their paper, YOT, was particularly poor at spoken word/written word matching for "small manipulable objects" (e.g. office utensils). They suggested that this was consistent with the view that particular forms of knowledge are needed for (or are differentially ‘weighted’ for) the identification of different objects (see Section 2). For instance, small manipulable objects might be adversely affected if these items depend on the retrieval of motor memories for their identification and if brain regions supporting these memories were damaged. To account for finer-grained deficits of this sort, we may need to make more than a dichotomous distinction between sensory and ‘functional’ knowledge (cf. Farah & McClelland, 1991; Warrington & Shallice, 1984). Different forms of both sensory and functional knowledge need to be separated. Furthermore, the ‘functional’ knowledge used in the recognition of living things (e.g., as actors) may differ from that used in the recognition of nonliving things (as instruments).

Now, although it seems intuitively plausible that functional information plays an important role in the identification of nonliving things, there has been surprisingly little behavioural data on the topic (though see section section 5.1 for correlatory evidence from functional imaging studies). Some direct empirical support comes from Humphreys and Riddoch (1999). They examined a child, JS, with learning difficulties due to a genetic abnormality and birth trauma. JS was presented with pictures of artefacts along with either a contextual scene or with an action (mimed by the experimenter). In the scene condition JS was told that ‘this is a ‘n’ and you find it in a ‘m’’ (e.g., ‘this is a glass and you find it on a table’, where the glass was shown alongside a pictured table with a place setting). In the action condition she was told ‘this is a ‘n’ and you do this with it’ (‘this’ being shown by the mime). JS was asked to repeat the object’s name, and then the next learning trial began. Subsequently she was presented with the individual objects in isolation and asked to recall the names. Performance was better after learning with the actions than the contexts. The finding is of some interest since it suggests that pairing an artefact with an action is useful in establishing a stable link between the object and its name. Interestingly JS was better at naming nonliving things relative to living things. Humphreys and Riddoch proposed that nonliving things might benefit from having object-specific action pairings, which can help name retrieval. In contrast, the actions performed by living things will tend to be more similar and hence less useful in supporting naming of the individual stimulus. Note also that the actions performed by living things are also different in kind from the actions performed on nonliving things. This too may be important for learning in such a child.

Most studies of patients with deficits for nonliving things have not clearly defined the stage in the recognition and naming process at which the deficit might occur, using tasks designed to ‘tap’ structural, functional/associative (semantic) or name knowledge. The patient reported by Cappa et al. (1998), however, showed no effect of category when answering probe questions, and they suggest that the problem was in name retrieval for nonliving things. Other patients, though, have shown a deficit in discriminating between nonliving targets and semantic distractors (e.g., Sacchett & Humphreys, 1992), which is more suggestive of a deficit in accessing associative knowledge. It may be that, as with deficits for living things, we will need to account for impairments arising at different levels of the object processing system.

4.2: Impaired functional knowledge for living things

Accounts that stress the importance of forms of functional knowledge only for nonliving things have some difficulty with the finding that some patients show impaired retrieval of functional as well as visual/perceptual information about living things. In contrast, the same patients may demonstrate reasonably good functional knowledge about nonliving things (Caramazza & Shelton, 1998; Funnell & de Mornay Davies, 1996; Laiacona et al., 1993; Samson, Pillon & de Wilde, 1998; Sheriden & Humphreys, 1993). If functional knowledge is differentially important for identifying artifacts, then loss of this knowledge ought to lead to poor identification of these items. Clearly this is not necessarily the case. The data can be accommodated, though, by the idea of categorically organised stored knowledge. For example, Caramazza and Shelton interpreted the above pattern of data in terms of there being separate semantic representations for living and nonliving things, so that damage to the representations for living things leads to poor retrieval of both stored perceptual and functional knowledge for these stimuli. Caramazza and Shelton suggested that this separation of knowledge for living and nonliving things reflects evolutionary pressures that highlight the importance of the categories animal, plant life and artifact. They argued that, because of evolutionary gains in distinguishing between these three types of object, categories for animals, plants and artifacts form the basis for the organisation of conceptual knowledge. "The evolutionary adaptations for recognising animals and plant life would provide the skeletal neural structures around which to organise the rich perceptual, conceptual and linguistic knowledge modern humans have of these categories". This view, that stored knowledge in organised by these three basic categories, can account for some of the finer-grained dissociations found between patients showing category specific deficits. For instance, whilst some patients seem particularly poor with fruits and vegetables (Farah & Wallace, 1992; Forde et al., 1997; Hart et al., 1985), others have been reported with deficits only for animals and not for fruit and vegetables (Caramazza & Shelton, 1998; Hart & Gordon, 1992). This would be expected if knowledge about animals and plant life is functionally (and anatomically) separate. For accounts that do not assume that stored knowledge is categorically organised, such finer-grained impairments could be attributed to the use of particular forms of knowledge to identify particular objects; perhaps shape information is more important for animals and colour or texture for fruit and vegetables. Loss of information about shape or colour knowledge might selectively impair recognition of animals or fruit/vegetables, respectively (see Section 4.1 above).

Consistent with Caramazza and Shelton’s hypothesis, Keil (1987) has argued that the underlying conceptual structures for living and nonliving things are qualitatively different, and will be affected in qualitatively different ways by the transformations that can be applied to them. Keil used the example of a chair that has it’s back sawn off and an extra leg glued on – most people would agree that it has become a stool rather than a chair. However, dying the color of a dog’s fur to red, making its tail bushier and allowing it to live in the wild, will not change it from a dog to a fox. By changing the salient perceptual and action-related features of nonliving things we can alter identities; the same does not apply for living things. When asked to justify why, for example, the dog did not become a fox, participants stated that it was because the internal organs (or DNA) would remain the same (although they could not state how the internal organs of foxes and dogs differ). Keil proposed that ‘rather it is a belief in biological essence that seems to grow out of a naïve theory of natural kinds that is driving their intuitions’.

The argument for knowledge being organised along categorical lines has been given further support by studies of semantic development in children. S. Gelman (1988) reported that 4 year-olds can explicitly state whether objects are made by people or not (see also S. Gelman & Kremer, 1991), and a number of studies have demonstrated that preschoolers (again around 4 years old) have a relatively sophisticated understanding of the differences between living and nonliving things. For example, R. Gelman and Meck (R. Gelman, 1990) asked children to describe what was on the inside and outside of animate and inanimate objects. The insides and outsides of animate things were described in different ways, whilst the insides and outsides of inanimate objects were described in the same way. R.Gelman (1990) argued that these results highlight children’s tendencies to generalise properties across examplars of animate objects, but not across the animate-inanimate boundary, and proposed that knowledge about animate and inanimate objects was domain-specific and governed by contrasting underlying rules.

A similar argument comes from the work of Massey and R. Gelman (1988). They presented children with colored photographs of novel examples of mammals, non-mammalian animals, rigid complex artifacts, wheeled objects and statues that had animal-like parts, and asked them which items were capable of going up a hill by themselves. Three and four years olds chose the animals, and rejected the artifacts with animal-like parts. This indicates an ability to use the appropriate visual-perceptual properties present in the novel objects to decide whether or not the object was capable of self-initiated movement. R. Gelman (1990) argued that this ability to distinguish between objects that can and cannot move on their own is the basis for the development of the conceptual distinction between animate and inanimate objects.

However a number of other studies have indicated that the animate/inanimate distinction is acquired even earlier in infancy (Mandler & McDonough, 1996; Poulin-Dubois, Lepage & Ferland, 1995; Smith, 1989). Smith (1989) presented 12 month old infants with pictures of nonobjects that were made from combining parts of animals (to make’living nonobjects’) or artefacts (to make ‘nonliving nonobjects’), and, using an habituation paardigm, demonstrated that the children could distinguish between the ‘living’ and ‘nonliving’ things. Mandler, Bauser and McDonough (1991) further showed that 18 month old infants could make distinctions even at a more fine-grained level, separating animals, plants and nonliving things. Indeed, even within the domain of nonliving things the infants could apparently distinguish furniture from kitchen utensils, though they could not separate tools and musical instruments. These studies suggest that infants are able to make judgements about global categories very early, perhaps even before they can categorise objects at a basic level (e.g., as chair as opposed to furniture). Mandler et al. (1991) posited that infants do not use physical similarity when forming categories but rely instead on more abstract properties, such as movement patterns and origin of movement (self generated or extrinsic), which are associated with each category (see also R. Gelman, 1990; Mandler, 1992). Consistent with this, Mandler and McDonough (1996) found that 14 month old infants generalised their responses to actions between living things, but not to nonliving things (and vice versa). When shown a dog drinking the infants were willing to make a rabbit drink but not a motorcycle. When shown a car being started with a key they were willing to start a truck with a key but not a fish.

These findings suggest that even young infants can distinguish between different categories, lending support to the idea that there is a predisposition to differentiate between living and nonliving things. Furthermore, if living and nonliving things do come to belong to different knowledge domains, perhaps represented in different neural areas, then we would expect category-specific patterns of impairment after brain damage.

4.3: Some problems

The proposal that our semantic knowledge is categorically organised has primarily been driven by reports of patients who have category specific impairments accessing visual and functional information about living things, with no concomitant problems with nonliving things. However, these patients are only problematic for accounts stressing the importance of visual/perceptual knowledge for living things, and functional knowledge for nonliving things, if the functional knowledge evaluated is the same for both categories. But, as we have already noted, the term ‘functional’ knowledge has been used in neuropsychological studies to refer to all kinds of non-sensory information and does not always refer strictly to the function (or use) of the object. Typically ‘functional’ knowledge for nonliving things refers to object usage and motor activity on the part of the actor; in contrast, ‘functional’ knowledge for living things includes biological functions (e.g., eating), the context in which animals are found, the sounds animals make etc.. Hence the ‘functional’ knowledge that can be impaired in patients with impairments for living things is not the same as the ‘functional’ knowledge about nonliving things that can be spared. Until there have been reports of patients with category specific problems in accessing visual and functional knowledge when the type of information is matched across categories, there is no empirical reason to abandon an account which stresses the importance of different types of information for different categories. Furthermore, we suggest that accounts stressing the importance of different types of knowledge for different categories can in fact account for patients who have category specific impairments accessing all kinds of semantic knowledge (i.e., ‘functional’ as well as visual/perceptual). Let us suppose that living things are represented primarily in terms of visual and sensory features. Now, if these important (perhaps even defining) visual/sensory features are impaired, then patients may not have enough information to differentiate one exemplar from another. For example, if a patient does not know that a giraffe has a long neck, he/she does not really know what a giraffe is, and therefore could not answer a question tapping ‘functional’ knowledge, such as "does a giraffe eat meat or leaves?". This hypothesis has been supported by recent neuroimaging studies, which show that modality-specific areas associated with the processing of form are activated when subjects answer verbally presented questions about the visual and functional/categorical properties of living things (Chao, Haxby & Martin, 1999; Thompson-Schill, Aguirre, D’Esposito & Farah, 1999, see Section 5). It appears that perceptual knowledge of form is fundamentally important for our stored representations of living things. ‘Knowing’ what a living is, and being able to answer all kinds of questions about it, depends heavily on being able to access this stored perceptual information. In contrast, activation in cortical areas associated with form processing is not enhanced when subjects answer questions about the functional/categorical properties of nonliving things. This is consistent with the idea that the important contrasts, or defining attributes, are visual/perceptual for living things but not necessarily for nonliving things. Chao et al. (1999) in fact report increased activation in the middle temporal gyrus for tools (relative to animals) when categorical information must be retrieved. They point out that this region borders areas known to be specialised for processing motion (area V5), suggesting that information about motion when tools are used plays an important role in their categorisation.

We also suggest that the developmental research is consistent with the idea that different types of knowledge are important for different categories. Indeed, R. Gelmen (1990) argued that children learn to differentiate between animals and non-animals on the basis of the different movement patterns shown by these categories. Furthermore, the categorical account of stored knowledge does not distinguish between different types of learned representation (visual/perceptual and semantic/functional etc.), and consequently, has problems explaining why visual processing disturbances can be linked to poor identification and loss of stored knowledge for living things (Section 3). Also, the account has difficulties in accommodating cases where functional knowledge for living things appears to be intact and the deficit is only for the visual properties of these objects (Hart & Gordon, 1992; Riddoch & Humphreys, 1993). In Section 6 we outline the HIT account, that allows these different lines of evidence to be integrated.

Section 5: Anatomical considerations

5.1: Brain imaging studies

A number of recent studies have attempted to assess whether living and nonliving things are stored in anatomically separate brain regions, using functional imaging techniques. Martin, Wiggs, Ungerleider and Haxby (1996) used positron emission topography (PET) to compare the brain regions which were active when subjects were naming living (animals) and nonliving (tools) things. When they subtracted the regions activated when subjects were naming tools from the areas activated when subjects were naming animals, they found that naming animals selectively activated the left medial occipital lobe and inferior temporal regions. When they reversed the subtraction (i.e. areas active when naming tools minus areas active when naming animals) they found that naming tools selectively activated the left middle temporal gyrus and the left premotor region. Interestingly, the areas selectively activated when subjects were naming tools were very similar to those activated when subjects named actions associated with objects (Martin et al., 1995) or when they imagined grasping objects (Decety et al., 1994). This is consistent with the idea that activating information about object use and the associated motor action is important for naming tools. In contrast, the activation of stored visual information, within the left medial occipital and inferior temporal cortex, may be more prolonged and important for the identification of living things. Again this idea is supported by converging PET data indicating that similar areas are activated when subjects retrieve stored colours associated with objects (Martin et al., 1995). The results support the notion that different brain regions mediate the identification of living and nonliving things. Futhermore, since converging evidence can indicate the functional roles of the different brain regions, the results highlight the importance of particular domains of knowledge for certain objects (respectively visual and action-related, functional knowledge for living and nonliving things).

Quite similar results to those reported by Martin et al. (1996) were found by Perani et al. (1995), using a task in which pairs of objects had to be matched according to whether they had the same base-level names (e.g. a tennet and a hacksaw). Damasio et al. (1996), like Martin et al. (1996), examined functional activation in an object naming task, but focused on activation within regions of the temporal lobe. They found that naming animals resulted in activation of the left inferotemporal region (more anterior to regions highlighted by Martin et al., 1996) whilst the naming of tools led to activation of posterior middle and inferior temporal gyri.

It is not clear whether some of the reported anatomical differences reflect variations in scanning procedures or in the task requirements; nevertheless, the data do highlight the importance of the inferior temporal lobe for naming animals and the left, inferior frontal and posterior middle temporal regions for naming tools. There may, in addition, be selective involvement of the left medial occipital area in animal identification.

One problem with these PET studies is that the stimuli have tended not to be matched across the categories. Differences in the activations found with animals and tools, then, may be due to the contrasting types of knowledge invoked by these stimuli or differences in stimulus complexity or familiarity. Stronger activation of posterior visual processing areas (e.g. in medial occipital cortex) may be found with animals because they are visually more complex. Moore and Price (1999) attempted to assess this by comparing PET activations in naming and picture-word matching tasks with visually complex animals, visually simpler fruit/vegetables, visually complex (multi-component) nonliving things (vehicles, appliances) and visually simpler (single-component) nonliving things (tools, utensils). They found increased activation in the right medial extra-striate and occipito-temporal areas for multi-component relative to single-component items, and particularly for animals relative to the other categories they examined. This suggests that these posterior brain regions are implicated in the processing of more complex visual configurations, with animals being particularly dependent on activation here. In addition, living things (animals and fruits and vegetables) were associated with activation of the left anterior temporal cortex and nonliving things with activation of the left posterior middle temporal cortex (see also Damasio et al., 1996; Martin et al., 1996). These results suggest that, instead of thinking of there being one area specialised for processing living things, and one for nonliving things, there are rather multiple areas, implicated in processing and retrieving different forms of knowledge for the particular stimuli. For example, for living things, areas supporting both visual processing and the retrieval of associative knowledge may be important. Thus extra-striate areas are sensitive to visual complexity, and they are also implicated in the processing of animals, consistent with there being a strong mediating role of visual knowledge in the identification of these stimuli. Activation of the anterior temporal cortex for living things, similar to that reported by Damasio et al. (1996), may instead link visual processing to the retrieval of stored associative knowledge. This fits with lesion data from a patient reported by Breedin, Saffran and Coslett (1994) who suffered profound atrophy of the anterior, inferior temporal cortices whilst leaving more posterior areas of temporal cortex intact. This patient had only a mild deficit on object decision tests designed to assess access to stored structural knowledge, and he showed strong perceptual priming (Srinivas, Breedin, Coslett & Saffran, 1997), along with poor judgements based on semantic relatedness between objects. His performance was also worse with living than nonliving things. For accounts such as the Cascade model (section 3.1), living things should generate stronger associative as well as perceptual competition between stimuli, requiring prolonged differentiation at both perceptual and associative levels, as indicated by different sites of activation in functional imaging studies and effects of brain lesions at contrasting sites.

To link data on functional imaging to models in a more precise way, we need to establish the brain sites implicated at different stages of object processing. Price, Moore, Humphreys, Frackowiak and Friston (1996), for example, assessed whether there were particular brain regions linked to object naming as opposed to object recognition (referring here to access to stored structural knowledge). They had subjects perform four tasks: (1) name (coloured) objects, (2) name the colour of nonobjects (matched to the object for complexity), (3) say "yes" to the coloured objects (the object baseline), and (4) say "yes" to the coloured nonobjects (the nonobject baseline). They assumed that objects gain automatic access to structural information (cf. Glaser, 1994), and argued that the contrast between the two baseline conditions (with objects and nonobjects) can reveal areas involved in object recognition (activated by objects but not by equally complex nonobjects). The brain regions mediating name retrieval may be indicated by the contrasts between the two naming conditions (conditions 1 and 2) relative to the two baselines (since both the naming tasks required retrieval of a stored phonological label associated with the visual stimulus). Of most relevance here, though, is the interaction given by the contrast between the object naming condition and its baseline (conditions 1 minus 3) relative to the contrast between the nonobject colour naming condition and its baseline (conditions 2 minus 4). This interaction indicates areas associated specifically with retrieving the names of known objects from vision. Price et al. found that the interaction was linked to selective activation of the left inferior and posterior temporal lobe. This study suggests that this region is particularly involved in name retrieval for known objects, and it is more activated in name retrieval than in object recognition alone. The same area has also been found to be activated more by animals than by tools (Martin et al., 1996; Perani et al., 1995). If such inferior and posterior temporal regions are linked to the involvement of stored visual knowledge in processing, the data suggest that this knowledge is recruited particularly when name retrieval is involved. The conclusion that follows from this is that, even in on-line tasks, knowledge sources may be activated to different degrees, in a top-down manner (e.g., for naming, but not for recognition). We return to this in our formulation of the HIT account.

Studies using fMRI have added extra information, by differentiating further between regions within the posterior temporal cortex and by contrasting activation from objects with that found when the same knowledge is addressed from verbal questions. Thompson-Schill et al. (1999) proposed that stored visual knowledge was not only activated when subjects were asked to name living things, and to answer questions concerning their visual attributes, but also when asked to answer questions about non-visual information (e.g. are pandas found in China? are snails edible?). In contrast, they argued that it was not necessary to access stored visual information when functional questions were asked about nonliving things. In their study, subjects were asked yes-no questions about the visual and functional properties of living and nonliving things, and fMRI was used to assess whether the left ventral occipito-temporal cortex (particularly the fusiform gyrus) was activated in each condition, relative to a baseline condition in which nonsense auditory stimuli were presented. The left ventral occipito-temporal cortex was chosen as the area of interest because previous studies have shown that it is involved in representing the visual properties of objects (D’Esposito, Detre, Aguire, Aslop, Tippett & Farah, 1997; Martin et al., 1995). Thompson-Schill et al. found that there was a significant interaction between category (living vs. nonliving) and type of question (visual vs functional) in the pattern of activity in the fusiform gyrus (assumed to be involved in representing modality-specific visual/perceptual information). In particular, they found that, while this area was only activated by the visual questions for nonliving things, it was activated by both the visual and functional questions for living things. Chao et al. (1999) likewise report a remarkable overlap of areas in the lateral fusiform gyrus activated by animals both in an object naming task and in a task requiring both naming and the retrieval of meaning from words (e.g., to answer the question, ‘forest animal?’ To the word ‘deer’). They found similar overlap across tasks for tools, but in different areas - the middle fusiform gyrus and middle temporal gyrus. These results point to common regions being activated from different modalities. The regions implicated in the study of Chao et al. have, in other experiments, been linked to the processing of form (in the lateral fusiform gyrus) and the processing of motion (in middle fusiform and temporal areas) (see Bonda, Petrides, Ostry & Evans, 1996; Ungerleider & Haxby, 1994). Hence the findings are consistent with specific forms of knowledge being drawn on to different degrees when accessing stored knowledge about living and nonliving things, and that these forms of knowledge are organised respectively around sensory and motor processes. It also fits with this argument that areas of the left inferior frontal cortex, when scanned, have been shown to be activated more by nonliving than living things (and typically the nonliving things have been tools). From other imaging studies it appears that left inferior frontal cortex is implicated in the retrieval of knowledge about actions, and it is activated when subjects must name actions associated with tools as well as when they name the tools themselves (Grabowski, Damasio & Damasio, 1998; Grafton, Fadiga, Arbib & Rizzolatti, 1997; Martin et al., 1995). Chao et al. add that additional differences are apparent within the categories of living and nonliving items (e.g., between houses, tools and chairs), reflecting the kinds of knowledge important in the representation of each type of stimulus (e.g., houses, though nonliving, are not associated with activation of the middle temporal gyrus). We return to the implications of these data in Section 6.

In summary, neuroimaging data indicate that living and nonliving things activate different areas of cortex in identification tasks. These different areas of cortex can also be linked to contrasting stages of identification: visual processing and access to visual knowledge (bilateral, medial extra-striate), contrasting with access to associative (anterior temporal), motion and action knowledge (in left medial temporal and fronto-parietal regions). Naming, as opposed to recognition, also seems to involve activation of additional visual processes, consistent with there being top-down recruitment of visual knowledge for individuation of a target object from its perceptual neighbours (particularly in left inferior occipito-temporal regions).

These conclusions are also consistent with evidence on category differences in event-related potentials (ERPs). ERP studies are particularly useful for providing information about the time course over which representations are activated. An ERP study by Keifer (in press) suggests that living and nonliving stimuli generate both contrasting and overlapping ERPs, at different times. Subjects had to judge whether a super-ordinate category probe (presented verbally or by means of two pictures from the chosen category) was appropriate to a target stimulus (presented either as a name or a picture). Keifer found that early ERP differences emerged for living and nonliving targets (after about 160-200ms), which were restricted to when pictorial presentations were used. These early effects are consistent with there being more elaborated perceptual processing of pictures of living things than nonliving things. There were also differences between the categories on later ERPs, with living and nonliving things respectively producing a reduced N400 component over occipito-temporal (bilateral) and fronto-central (left hemisphere) sites. Other ERP studies have demonstrated differences in fronto-central sites for N400 components associated with action verbs vs. concrete nouns (Dehaene, 1995; Pulvermuller, Preissl, Lutzenberger & Birbaumer, 1996), suggesting that the fronto-central changes are linked to the retrieval of associated action knowledge. These differences in the later ERP component in Keifer’s study were found with both words and pictures as stimuli. The late effects are consistent with living and nonliving things drawing on contrasting forms of associative knowledge, stored in different brain regions, for both verbal and pictorial stimuli alike.

5.2: Lesion studies

In addition to their PET data, Damasio et al. (1996) reported results from group studies of patients with deficits in naming living or nonliving things (see also Tranel, Damasio & Damasio, 1997). They found that impairments in naming animals were associated with damage of the left inferotemporal region, and impairments in naming tools were associated with damage to the junction of the temporal, occipital and parietal cortices. Damasio et al. (1996) argued that their data reflected naming rather than recognition problems, since patients were classified according to whether independent raters could identify the object from the description produced by patients when the object could not be specifically named. In contrast, Tranel et al. (1997) reported patients classified by recognition impairments (after unilateral right as well as left hemisphere damage). They reported that recognition problems for tools were associated with damage to the left occipito-temporal-parietal junction, whilst recognition problems for animals were associated with both left and right medial occipito-temporal lesions. The lesion sites linked to recognition problems for animals in this last study were more posterior to those linked to naming problems in the earlier study, though conclusions need to be cautious. Correct identification of object descriptions by independent raters (as in Damasio et al., 1996) does not necessarily indicate that the patients had intact access to semantic knowledge. Also in these studies patients were classed as showing deficits if they performed at a level of 2 standard deviations below that shown by control subjects; however, this does not necessarily mean that the patients were more impaired on one category of object than another, when the categories were compared directly (see Caramazza & Shelton, 1998). Nevertheless we note that the recognition deficits associated with posterior right as well as left hemisphere medial occipito-temporal damage link to Moore and Price’s (1999) finding that multi-component objects (and particularly animals) strongly activate this brain region. It may be that additional visual processing is needed in this area to differentiate within sets of complex, structurally similar objects, generating both enhanced activation in normal subjects and deficits in recognizing animals in brain lesioned patients. This also fits with evidence on the category-specific deficits for living things found in the patients reported by Forde et al. (1997), Humphreys et al. (1997) and Riddoch et al. (1999). These patients had posterior occipito-temporal damage and all had particular problems in interrogating visual knowledge about objects. For instance, the patients were poor at object decision, at answering definitions that stressed the visual properties of objects and at drawing objects from memory. These problems were more pronounced for living things than for nonliving things. Associative and functional (action related) knowledge about both living and nonliving things was good.

More anterior temporal regions may be involved in access to associative knowledge, necessary for name retrieval for living things (see also Breedin et al., 1994). Consistent with this, deficits in some patients with poor identification of living things is linked to bilateral damage to antero-medial parts of the temporal lobes as well as to more inferior temporal regions (Gainotti et al., 1995). However, the degree to which the damage needs to be bilateral is unclear. Caramazza and Shelton (1998) for example report a patient with impaired knowledge of living things following unilateral damage confined to the left hemisphere (though to fronto-parietal regions in this case). The patients reported by Forde et al. (1997) and Humphreys et al. (1997) also suffered unilateral left hemisphere lesions (but, as noted above, to relatively medial occipito-temporal areas). For these patients, the deficits for living things were apparent in naming rather than recognition tests (e.g., fruits and vegetables could be categorised but not named). This suggests that left hemisphere representations are implicated particularly in naming tasks, with posterior occipito-temporal regions being interrogated to differentiate targets within sets of living things.

In contrast to the data on deficits with living things, Gainotti et al. (1995) conclude that deficits in recognising and naming nonliving things occur after damage to left parietal-frontal regions. However the data base for these patients is somewhat limited. Silveri et al. (1997) report one such case where an analysis using PET was conducted. They showed hypometabolism confined to the middle temporal gyrus, the hippocampal and parahippocampal regions and the inferior parietal lobe of the left hemisphere. In contrast, Cappa et al. (1998)’s patient, who apparently had problems particularly in naming nonliving things, had a unilateral lesion affecting the anterior left temporal lobe. Tipper, Glosser and Farah (1996) similarly reported problems in naming nonliving relative to living things in patients with resections of the left anterior temporal lobe to relieve epilepsy.

Overall, the neuropsychological data concur with the results from functional imaging studies in suggesting that different brain regions mediate the recognition and naming of living and nonliving things. Lesions affecting inferior occipito-temporal regions, extending anteriorly into the temporal lobe, seem particularly to disrupt processing for living things and lesions affecting left temporal-parietal and also parietal-frontal areas disrupt processing for nonliving things. There are some anomalies though, such as the left fronto-parietal lesion linked to a deficit in living things reported by Caramazza and Shelton (1998), and the data suggesting that the anterior left temporal area may be involved in naming nonliving things (Cappa et al., 1998; Tipper et al., 1996). These inconsistencies illustrate the importance of collecting further data, across larger sets of patients each assessed to provide a functional analysis of any deficit relative to a model of object naming.

Section 6: The Hierarchical Interactive Theory (HIT)

The data we have reviewed have demonstrated the reality of category-specific deficits in object naming as empirical phenomena, with effects occurring for both living and nonliving things. The majority of accounts that have been outlined to explain these deficits have emphasized single factors - the categorical nature of our stored knowledge (Section 4), the differential roles of perceptual vs. functional knowledge for living and nonliving things, respectively (Section 2), the importance of correlated perceptual features or of perceptual-functional relations (Section 2), the interaction of perceptual processes with other forms of knowledge needed for a task (Section 3). Our interpretation of the evidence is that, whilst particular deficits can be linked to each account (so that each is sufficient to explain particular patients), no single factor account can cover all of the patients (i.e., no one account is necessary). Instead, we propose a framework that is powerful enough to accommodate different patterns of deficit in contrasting patients.

6.1: Hierarchical knowledge and interactive processing

We begin by discussing the need for a hierarchical system for object processing by outlining a model of visual object recognition and naming, as this is the task most commonly used to diagnose category specific impairments. We then discuss the need for top-down (interactive) processes within such a model. These two ideas, of a hierarchical and interactive system, capture the essence of the HIT account. To account for object naming, HIT adopts the architecture of the Cascade model of object recognition, which had two defining characteristics: a hierarchy of stored representations and the assumption that partial activation could be transmitted between processing systems. The hierarchy of stored representations in the Cascade model contained three types of stored knowledge: (1) stored structural descriptions, (2) stored functional and inter-object associative information (‘semantic’ knowledge) and (3) name representations. According to the model, selective damage can occur to each form of representation, so that, for example, a patient can have a pronounced deficit in semantic knowledge (e.g., poor retrieval of associative or functional knowledge about an object) without necessarily having a deficit in stored structural knowledge. As we outline below, this is consistent with dissociations between the performance of different patients on tasks designed to ‘tap’ each stage of memory storage. Note, however, that although a hierarchy of memory storage is assumed, processing within models of this type does not depend on access to each memory stage being completed before the next is initiated. Thus, as far as processing is concerned, only a first pass of activation is hierarchical. Processing, after this first pass, is not necessarily hierarchical. For example, there may be a delay in accessing precise structural knowledge from an object, though access to categorical knowledge may be completed (see Humphreys et al., 1997).

Let us consider the neuropsychological evidence for a hierarchy of memory storage. Take first the distinction between stored structural descriptions and other types of knowledge. Evidence for this comes from patients who can perform object decision, but who still show impaired access to semantic information about an object. In object decision tasks patients can be asked to discriminate between stimuli (of equal perceptual complexity and perceptual ‘goodness’), some of which are real objects and others nonobjects created by interchanging the parts of real objects. Since objects and nonobjects cannot be distinguished on the basis of general perceptual information, their classification must depend on association with stored knowledge. Some patients can succeed at object decision whilst being impaired at matching associatively related objects or accessing any further information about the entity (Hillis & Caramazza, 1995; Humphreys & Riddoch, 1999; Riddoch & Humphreys, 1987c; Sheridan & Humphreys, 1993; see also Breedin et al., 1994). This suggests that such patients can access structural knowledge (to distinguish objects and nonobjects) even when access to associative knowledge is deficient.

Some patients with category specific deficits for living things are poor at performing object decisions, especially for items from the affected categories (e.g., Caramazza & Shelton, 1998; Sartori & Job, 1988). However, this is not universally the case, and some patients demonstrate intact object decision performance for affected categories (Laiacona, Capitani & Barbarotto, 1997; Sheridan & Humphreys, 1993). We propose that these contrasting patients have lesions affecting different functional stages of object processing. The former ‘category specific’ patients have deficits affecting visual access to structural knowledge (in the structural description system) in addition to any further deficits affecting functional/associative knowledge or name retrieval. The latter ‘category specific’ patients do not have a deficit affecting the visual activation of structural knowledge, and their deficit resides at a later processing stage.

There are other explanatory advantages in a model separating stored visual/perceptual knowledge, in the structural description system, from other types of knowledge. In particular, such a model can account for patients who show good knowledge of functional/associative knowledge of living things when given their spoken name, whilst still being impaired in naming them from vision (Forde et al., 1997; Riddoch & Humphreys, 1987c, 1993). Damage confined to the structural description system should leave other forms of knowledge intact, though visual naming (and access to visual knowledge in general) may be impaired. Such a model can also help explain a pattern of deficit in which a patient has good visual access to structural knowledge (e.g., object decision), good auditory access to functional/associative knowledge, but poor auditory access to structural knowledge (e.g., drawing from memory). This pattern was reported by Hart and Gordon (1992) and Riddoch and Humphreys (1987c). It can be accommodated if there is a selective impairment in mapping back from the semantic system to the structural description system, and if the structural description system is used when long-term knowledge of visual properties must be retrieved.

All of the above deficits can be conceptualised in terms of hierarchical memory representations that are accessed in a feed-forward manner, as proposed by the Cascade model (Humphreys et al., 1998) - providing there is also off-line interrogation of visual/perceptual knowledge in the structural description system. The latter process would be required when we are asked to retrieve perceptual knowledge when given the name of an object. We next consider a pattern of performance that suggests that, even for on-line naming, some form of re-interrogation of structural knowledge is required. This leads to our formulation of the HIT model.

As we discussed in Sections 3.1 and 5.2, some authors have reported patients whose category specific problems seem more pronounced on name retrieval rather than on tests of semantic knowledge (Cappa et al., 1998; Farah & Wallace, 1992; Forde et al., 1997; Hart et al., 1985; Humphreys et al., 1997). Nevertheless, in a detailed evaluation that has been conducted on two such patients with problems with living things, the fundamental problem appeared to be in activating perceptual knowledge about objects. As we have noted, Humphreys et al. (1997) reported two patients who were impaired at naming living things but, on forced-choice tests, showed apparently good access to associative and categorical knowledge about these objects (e.g., successfully categorising fruits and vegetables). Yet, despite this good performance on associative and categorical tasks, deficits in perceptual knowledge were revealed by tests of drawing from memory, object decision, and naming to perceptual definitions. The apparent inconsistency between, on the one hand, the relatively good access to associative and categorical information from vision and, on the other, the perceptual knowledge impairment, may be accounted for in at least two ways. It may reflect the ability of patients to access partial associative and categorical knowledge and to respond accurately using this information under forced-choice conditions. In addition, it may reflect the role of top-down (re-entrant) activation of perceptual knowledge in object naming, which we discuss below.

The idea here is that, in a first pass in object processing, there is activation of stored structural descriptions and partial activation of associative/functional knowledge. However, for naming to be achieved there needs to be further (top-down) interrogation of perceptual knowledge, which drives the process of differentiating a target object from its close neighbours. For living things, this may require further interrogation of form information for animals, colour and texture for fruits and vegetables and so forth. For nonliving things this may mean further interrogation of action-related functional knowledge, that distinguishes one artefact from another. In patients such as those described by Humphreys et al. (1997), with problems in naming living things, we propose the following. There is initial visual access to associative/functional knowledge, but the mild impairment in perceptual knowledge prevents successful re-entrant activation from being achieved. Consequently, patients cannot access enough information to differentiate between the target item and its close perceptual and semantic neighbours. The HIT model incorporates this idea of re-entrant processing, which can be conceptualized in terms of an interactive activation and competition framework, a