Abstract

Theories about the neural bases of semantic knowledge tend between two poles, one proposing that distinct brain regions are innately dedicated to different conceptual domains and the other suggesting that all concepts are encoded within a single network. Category-sensitive functional activations in the fusiform cortex of the congenitally blind have been taken to support the former view but also raise several puzzles. We use neural network models to assess a hypothesis that spans the two poles: The interesting functional activation patterns reflect the base connectivity of a domain-general semantic network. Both similarities and differences between sighted and congenitally blind groups can emerge through learning in a neural network, but only in architectures adopting real anatomical constraints. Surprisingly, the same constraints suggest a novel account of a quite different phenomenon: the dyspraxia observed in patients with semantic impairments from anterior temporal pathology. From this work, we suggest that the cortical semantic network is wired not to encode knowledge of distinct conceptual domains but to promote learning about both conceptual and affordance structure in the environment.

INTRODUCTION

It is well known that knowledge of different conceptual categories (e.g., animals, tools, faces, landmarks) can be differentially impaired with brain damage (e.g., Lambon Ralph, Patterson, Garrard, & Hodges, 2003; Hillis & Caramazza, 1991; Warrington & Shallice, 1984) and can elicit different cortical responses in healthy adults (Hwang, Palmer, Basho, Zadra, & Müller, 2009; Martin & Chao, 2001; Chao, Haxby, & Martin, 1999; Martin, Wiggs, Ungerleider, & Haxby, 1996), but the importance of such findings for theories about the neural bases of conceptual knowledge remains unclear. For some, category sensitivity provides evidence of dissociable and innately constrained domain-specific knowledge systems in the mind and brain (Kanwisher, 2010; New, Cosmides, & Tooby, 2007; Caramazza & Shelton, 1998). Yet items from different conceptual domains vary in many of their sensory and motor characteristics, raising the alternative possibility that category sensitivity arises through learning about sensory–motor structure within a single domain-general knowledge system (for a review, see Chen & Rogers, 2014). The controversy reflects two fundamentally different views of the mechanisms that support everyday human knowledge.

fMRI studies of the congenitally blind have recently offered new evidence important to the issue (He et al., 2013; Bedny, Caramazza, Pascual-Leone, & Saxe, 2012; Mahon, Schwarzbach, & Caramazza, 2010). In a seminal study, Mahon, Anzellotti, Schwarzbach, Zampini, and Caramazza (2009) compared activations in the posterior fusiform gyrus (pFG) of sighted and congenitally blind adults while they made semantic judgments from spoken names of animals or artifacts. Both groups showed an activation gradient across the lateral/medial axis, with manmade objects, relative to animals, evoking more activation medially and less laterally in the pFG, consistent with the well-documented pattern in sighted populations for pictorial stimuli (Chouinard & Goodale, 2010; Mahon et al., 2007; Chao et al., 1999) and, less frequently, for spoken and written words (Mahon et al., 2009; Devlin, Rushworth, & Matthews, 2005). Thus, category-sensitive activation patterns can arise in the ventral visual stream in the absence of visual experience. This result appears to seriously challenge domain-general views of semantics that have attributed posterior fusiform organization to the influence of visual experience (e.g., Rogers, Hocking, Mechelli, Patterson, & Price, 2005).

To explain the pattern, Mahon et al. (2007, 2009) proposed that functional specialization in ventral temporal cortex is constrained, not solely by patterns of visual input, but also by innate connectivity between ventral temporal cortex and other areas. Blind and sighted participants show similar category effects because they have the same initial connectivity, which is configured to support distinct category-specific knowledge networks. In this sense, the data are offered as support for the hypothesis that evolutionary pressures have led to independent domain-specific knowledge systems in the brain (Caramazza & Shelton, 1998).

This distributed domain-specific (DDS) view can be viewed as advancing two separate proposals: first, that functional specificity in cortex is partly shaped by native patterns of connectivity, and second, that such connectivity produces a domain-specific organization of the cortical semantic network. In several elegant follow-up studies, Mahon and colleagues have provided substantial evidence for the former proposal (He et al., 2013; Bedny et al., 2012; Almeida, Mahon, & Caramazza, 2010), and this work is of a piece with a broader movement in the literature to understand how cortical network connectivity shapes functional specialization across many cognitive domains (Bouhali et al., 2014; Plaut & Behrmann, 2011; Martin, 2006). We believe it represents a significant advance in the effort to understand the organization of the cortical semantic network.

The second proposal of the DDS hypothesis, however, raises some nontrivial puzzles. First, Mahon et al. (2009) observed both similarities and differences between sighted and blind participants; the DDS hypothesis clearly explains the similarities but not the differences. Second, the mechanisms that give rise to the proposed domain-specific organization of cortex remain underspecified, beyond the general hypothesis that connectivity is important. The authors describe connectivity as being “concentrated around” distinct conceptual domains (Mahon & Caramazza, 2009) or “at the grain of a limited number of evolutionarily relevant domains of knowledge” (Caramazza & Mahon, 2003), but it is not clear what specific commitments these statements entail. Third, it is difficult to see how the DDS hypothesis might explain phenomena that challenge any domain-specific approach to semantic representation. One such phenomenon is the existence of domain-general semantic impairments resulting from circumscribed cortical pathology, such as that observed in semantic dementia (SD)—a progressive dementing illness that gradually erodes semantic knowledge for all conceptual domains and all modalities of reception and expression (Lambon Ralph, Lowe, & Rogers, 2007; Adlam et al., 2006; Hodges, Patterson, Oxbury, & Funnell, 1992). These deficits are caused by cortical thinning and hypometabolism circumscribed within ventrolateral anterior temporal cortex (Guo et al., 2013). It is difficult to see how localized pathology of this kind would produce domain-general knowledge impairments if semantic knowledge is supported by multiple independent domain-specific networks. Relatedly, although items from distinct semantic domains certainly differ in many of their properties, they can also share important characteristics. Thus, animals and tools are physical objects, subject in similar ways to forces of gravity, momentum, and inertia; hawks and gliders can hang suspended in the air for similar reasons; cows are living animals but are used for food and to make leather. If animals, foods, and manmade objects are supported by distinct knowledge systems, how might knowledge about a living cow generalize to the butchered cow? If semantic knowledge is only encoded in separate domain-specific systems, it is not clear how cross-domain knowledge generalization occurs (see Rogers & McClelland, 2004, 2008; Rogers & Plaut, 2002, for discussion).

This article considers whether one important contribution of the DDS hypothesis—the proposal that ventral temporal representations are partly shaped by native connectivity—can explain the interesting imaging data in sighted and blind populations, within a domain-general semantic knowledge system. Our approach builds on the “semantic hub” theory that was developed in part to explain the domain-general knowledge impairments observed in SD (Lambon Ralph et al., 2007; Patterson, Nestor, & Rogers, 2007; Rogers et al., 2004). In common with many others, this view proposes that conceptual knowledge arises from learned associations among sensory, motor, and linguistic representations localized in different brain regions and communicating via pathways in a distributed cortical network. Although many pathways exist in the network, the theory further proposes that there exists a single cross-modal “hub” located in the anterior temporal lobes (ATL) that is important for mediating interactions among many different surface modalities for all domains of knowledge. By virtue of such mediation, the hub comes to encode representations that express conceptual similarity structure and so support conceptual generalization (Rogers et al., 2004). A key assumption of the approach is that the same “hub” neurons contribute to the representation of all conceptual domains, so that they can express both differences and commonalities across domains. Consequently, knowledge of all conceptual domains is disrupted with neuropathology to the hub, as observed in SD (Lambon Ralph et al., 2007; Noppeney et al., 2007).

We here elaborate this view by adopting a central thesis of the DDS hypothesis, also presaged in earlier work by Plaut (2002). Specifically, we propose that different pathways in the network vary in the strength or effectiveness of their connectivity. Regions that are highly interconnected will exert strong mutual influences and so will respond similarly to given inputs. The activation patterns generated within such network will thus reflect not only the effects of learning in a structured environment but also the base connectivity of the network. Specifically, category-sensitive activations in pFG arise because the medial aspect of this region is more effectively connected to parietal regions that support object-associated manipulation knowledge (Mahon, Kumar, & Almeida, 2013). In contrast to the DDS proposal, however, we further hypothesize that such patterns can arise within a system that remains domain general insofar as critical elements contribute to semantic processing for all conceptual domains. Such an account would thus reconcile the challenges faced by both domain-general and domain-specific approaches to semantic knowledge.

Although the hypothesis itself is straightforward, its assessment is not. Existing neurocomputational models of semantic knowledge have been connected only loosely (Lambon Ralph et al., 2007; Rogers et al., 2004) or not at all (Cree, McNorgan, & McRae, 2006; Tyler, Moss, Durrant-Peatfield, & Levy, 2000; Farah & McClelland, 1991) to the anatomy of the cortical semantic network. Although useful for understanding patterns of behavior in healthy and disordered populations, such models cannot shed light on the anatomical distribution of functional activation patterns evoked by semantic tasks; indeed, no prior computational model has been applied to the understanding of fMRI data in semantic tasks. Second, our working hypothesis requires us to specify how visual, verbal, and praxic representations interact within the semantic network, yet prior semantic models have mainly focused on interactions between visual and verbal representations. Third, the introduction of praxic representations must allow for an account of the dyspraxic patterns commonly observed in patients with semantic knowledge impairments (e.g., Hodges, Bozeat, Lambon Ralph, Patterson, & Spatt, 2000), yet no such model has previously been advanced.

We addressed these challenges by investigating three different extrapolations of the ATL hub model of semantic processing (Lambon Ralph et al., 2007; Patterson et al., 2007; Rogers et al., 2004). The hub model aligns coarsely with the anatomy of the temporal lobe insofar as phonological representations of words are supported by the superior temporal gyrus (Graves, Desai, Humphries, Seidenberg, & Binder, 2010; Davis, Di Betta, Macdonald, & Gaskell, 2009; Catani, Jones, & Ffytche, 2005), visual representations of objects are supported by posterior inferotemporal cortex (Rogers et al., 2005; Goodale & Milner, 1992), and these regions are reciprocally connected with anterior temporal cortex where different sensory streams converge (Binney, Parker, & Lambon Ralph, 2012). The three models presented here, however, differ in their proposals about how praxic representations encoded in left lateral parietal (LP) cortex (Binkofski & Buxbaum, 2013; Buxbaum, Kyle, Grossman, & Coslett, 2007) interact with the rest of the network, with two motivating factors.

The first is neuropsychological: along with other modalities of reception and expression, knowledge about object praxis is seriously degraded in patients with SD. In particular, SD patients show impairments acting on common objects or matching them for shared praxis or function but spared knowledge of action affordances in their use of novel tools and in components of praxis directed toward everyday objects (Bozeat, Lambon Ralph, Patterson, & Hodges, 2002; Hodges et al., 2000). The preservation of systematic information (such as visuo-praxic affordances) with degradation of more idiosyncratic information (such as unaffordanced elements of praxis) is a hallmark of SD in visual and verbal tasks (Patterson et al., 2006); consequently prior articulations of the “hub” theory have proposed that praxic representations interact directly with the ATL hub so that praxic knowledge degrades in the same way as visual and verbally expressed knowledge with ATL pathology (Rogers & Patterson, 2007).

The second motivation stems from studies of connectivity between pFG and LP regions that support “online” object-directed action (dorso-dorsal stream) and long-term object use representations (ventro-dorsal stream; Binkofski & Buxbaum, 2013; Rizzolatti & Matelli, 2003). Contra the proposal from neuropsychology, human diffusion-weighted tractography does not reveal direct anatomical connectivity between ATL and LP (Binney et al., 2012; Caspers et al., 2011; Wakana, Jiang, Nagae-Poetscher, van Zijl, & Mori, 2004), but there is considerable evidence for an indirect pathway projecting through the medial pFG. Mahon and colleagues (Garcea & Mahon, 2014; Mahon et al., 2007, 2013) have reported significant functional connectivity between medial pFG and parietal areas including the supramarginal gyrus and a more superior region (Mahon et al., 2013). Tractography in nonhuman primates shows long projections from LP cortex into the ventral temporal lobe (Zhong & Rockland, 2003; Seltzer & Pandya, 1994), and human white matter tractography suggests that the dorsal aspect of the inferior fronto-occipetal fasciculus, which runs medially along the temporal lobe, branches dorsally at a point near pFG to terminate in superior parietal radiations (Wakana et al., 2004). These observations are all consistent with the view that medial pFG connects with parietal regions that support object-directed action. Moreover, both medial and lateral FG send white matter projections to the ATL (Binney et al., 2012).

We consider three variants based on the original hub model (see Figure 1) that vary in the architecture through which associations among visual, verbal and praxic representations are encoded. The hub-only model incorporates a direct pathway between praxic representations and the ATL hub, the fully connected model includes this pathway as well as a path connecting visual and praxic regions directly, and the anatomically constrained model includes the visual–praxic pathway suggested by connectivity studies without direct connection between praxic and hub regions. To capture lateral/medial differences in pFG connectivity, visuo-praxic connections in the fully connected and anatomically constrained models were anatomically graded so that “medial” visual units interacted more effectively with praxic units than did “lateral” visual units.

Figure 1. 

Illustrations for the “hub” model of semantic processing. (A) A hypothetical cortical semantic network for object recognition, use, and naming tasks. This simplified network involves ATL, LP cortex, pFG, and the perisylvian language system (mostly in superior temporal gyrus). Solid lines delineate identified anatomical pathways (Binney et al., 2012). The dark blue dashed lines and red dotted line delineate two hypotheses about how praxic representations interact with the network. (B) Model architectures for three hypotheses about network connectivity. The colored ovals are hidden layers which are analogous to critical brain regions illustrated in A and used to simulate brain activation patterns. The dark blue dashed lines depict the lateral/medial gradient in connectivity effectiveness between visual and praxic representations. This structure differentiated the hub-only from fully connected and anatomically constrained models. The red dotted line illustrates the direct connection between praxic representations and the ATL hub, included in hub-only and fully connected models.

Figure 1. 

Illustrations for the “hub” model of semantic processing. (A) A hypothetical cortical semantic network for object recognition, use, and naming tasks. This simplified network involves ATL, LP cortex, pFG, and the perisylvian language system (mostly in superior temporal gyrus). Solid lines delineate identified anatomical pathways (Binney et al., 2012). The dark blue dashed lines and red dotted line delineate two hypotheses about how praxic representations interact with the network. (B) Model architectures for three hypotheses about network connectivity. The colored ovals are hidden layers which are analogous to critical brain regions illustrated in A and used to simulate brain activation patterns. The dark blue dashed lines depict the lateral/medial gradient in connectivity effectiveness between visual and praxic representations. This structure differentiated the hub-only from fully connected and anatomically constrained models. The red dotted line illustrates the direct connection between praxic representations and the ATL hub, included in hub-only and fully connected models.

In what follows, we first show that the similarities and differences in functional activations evoked in sighted and blind individuals emerge in just one of the architectures. We then consider how well each architecture explains dyspraxic patterns arising from ATL atrophy, with results that validate conclusions drawn from the simulation of imaging results. The simulations illustrate a mechanism by which category-sensitive patterns can arise from network connectivity as suggested by the DDS hypothesis but also suggest a somewhat different explanation as to why the system has come to exhibit a particular pattern of connectivity in the first place.

SIMULATION 1: CATEGORY SENSITIVITY IN HEALTHY AND CONGENITALLY BLIND GROUPS

The goal of the first experiment was to assess the ability of each hypothesis to account for the category-sensitive patterns of functional activation in the pFG of sighted and congenitally blind populations.

Methods

Model Architectures

The three model architectures shown in Figure 1 were implemented using the light efficient network simulator (Rohde, 1999). Each model included three visible layers that directly encoded, as distributed patterns of activation over units, prespecified representations corresponding to the visual appearance of an object, the object's name and verbal description, and its associated praxis. Each visible layer was reciprocally connected with its own modality-specific hidden layer, providing model analogs to the pFG (visual hidden units), superior temporal gyrus (verbal hidden units), and LP cortex (praxic hidden units). An additional hidden layer served as the ATL hub (cross-modal units). All units employed a sigmoidal activation function and were given a fixed bias of −3 so that, in the absence of input from other units, they tended to adopt a low activation state.

The models varied in how the hidden layers connected. For the hub-only and fully connected models, all modality-specific hidden layers were reciprocally connected with the ATL units. For the fully connected and anatomically constrained models, pFG units were reciprocally connected with LP units. To capture the proposed gradation in connectivity effectiveness along this pathway, units in the pFG layer were situated along an anatomical lateral-to-medial axis. The error derivatives (and hence the strength of influence on weight changes) on visuo-praxic connections was scaled according to the visual unit's position along this axis, with medial-most units having a larger scaling factor than other connections in the model and the rate diminishing for increasingly lateral units according to a sigmoid function:
formula
where S(i) is the scaling parameter applied on error derivatives, n is the number of units in visual layer, and i (the unit location) ranges from 1 to 20 on the medial-lateral axis. Across units, the mean learning rate was equal to that on other connections in the model.

Training Environment

A model environment was constructed containing visual, verbal, and praxic representations for 24 different animal items and 24 different tool items, with each domain organized into four basic categories, each containing six exemplars. Visual representations for each item were generated stochastically in accordance with the constraints identified by Rogers et al. (2004) with animals sharing more visual properties overall than tools, Manimal = 7.58, SD = 1.38 and Mtool = 4.17, SD = 1.20 (McRae, de Sa, & Seidenberg, 1997). Thus, (1) items in different domains shared few properties, (2) items within the same category shared many properties, (3) animals from different categories shared more properties than did artifacts from different categories, and (4) animals had more visual properties than did artifacts. Each item was also given a unique name represented by a single unit and a set of 10 features as verbal descriptors, which were designed to capture the linguistic environment for describing animal and tool objects. The number of verbal features was largely comparable between categories (see Table 1).

Table 1. 

Mean (SD) Number of Features Per Item for Visual, Praxic, Verbal Representations for Animal and Tool Items in Training Environment

Visual (n = 30)Praxic (n = 10)Verbal (n = 10)Name (n = 48)Total (n = 98)
Animal 7.58 (1.38) 0.00 (0.00) 5.04 (1.16) 1.00 (0.00) 13.63 (1.66) 
Tool 4.17 (1.20) 4.38 (1.38) 4.21 (0.83) 1.00 (0.00) 13.75 (2.36) 
Visual (n = 30)Praxic (n = 10)Verbal (n = 10)Name (n = 48)Total (n = 98)
Animal 7.58 (1.38) 0.00 (0.00) 5.04 (1.16) 1.00 (0.00) 13.63 (1.66) 
Tool 4.17 (1.20) 4.38 (1.38) 4.21 (0.83) 1.00 (0.00) 13.75 (2.36) 

Praxic representations were also constructed for each item, taking the form of distributed patterns over the 10 visible units in the visible praxic layer. For all animal items, all praxic units were simply turned off, capturing the intuition that most animals are not associated with rich praxic information. For artifacts, praxic representations were generated for each tool by copying the item's visual properties and flipping the state of each unit with a small probability (p = .2), producing a model analog of partial visual affordance: Each visual tool feature reliably predicted a particular praxic feature across items, but this prediction was violated for some items. Consequently, the visual and praxic features of tool category had a correlation of .65 overall.

Tasks and Training Procedures

Given partial information about an item as input, the models were trained to activate the item's associated properties, including its name, verbal description, and visual and praxic features. Model inputs could be single names, verbal descriptions, visual images, or praxic features. Inputs were applied to visible units by providing these with direct excitatory input, and units throughout the network were updated successively over time in random order. Time was discretized with an integration constant of 30, so that each unit updated its state 30 times in one cycle. After eight update cycles, target values were applied to visible units and weights were updated using backpropagation with a learning rate of 0.01 (scaled as noted above for visuo-praxic connections) and weight decay of 0.0005 without momentum (Rumelhart, Hinton, & Williams, 1988). Although fully sighted model variants were trained with all varieties of input and outputs, congenitally blind model variants were trained with the same parameters on the same patterns, but without visual experience: Visual inputs were never applied to the model, and visual units were never given targets.

The response to a given input was counted as correct if, across all visible units, target properties were active above 0.5 and nontarget below 0.5 (for a similar approach, see Harm & Seidenberg, 2004).1 All models, both sighted and blind, were trained for 100k epochs at which point they generated correct output by this criterion for the great majority of inputs. Therefore, the sighted and blind variants of this network were equated on the total amount of nonvisual experience.2

Simulation of Functional Brain Imaging Results

The brain imaging studies involved tasks in which participants made a semantic judgment from either a picture or the spoken name of a familiar item. To simulate these tasks, input corresponding to either an item's visual appearance (for picture comprehension) or its name (for word comprehension) was applied to the model, which then cycled until unit activations stopped changing. Given that each unit was associated with a large negative bias and its initial activation was rather low, unit activations then were largely driven by positive input weights and so provide a model analog to synaptic activity in the network dynamics. After settling, unit activations were recorded and taken as an analog of the BOLD response at a single voxel. This value was then distorted with Gaussian noise (μ = 0, δ2 = 0.1) to reflect the error in signal estimation intrinsic to brain imaging methods.

For each architecture and model population (sighted/blind), 15 different model training runs were conducted, each initialized with a different random set of weights, to simulate 15 different subjects. Our analysis then followed Mahon et al.'s (2009) procedure. The response of each unit was averaged across items in each condition (tool/animal) and then spatially smoothed with a Gaussian kernel (μ = 0, δ2 = 1) encompassing two adjacent units. A group-level contrast was performed to find the peak activation for both animal and tool concepts using the averaged data across the 15 model subjects. An ROI analysis was then performed on the activation value of the peak unit averaged with two neighboring units on either side. This procedure was used to generate the results shown in Figure 2.3

Figure 2. 

Mean activation in lateral and medial ROIs of the visual layer for three models. (A–C) Hub-only models. (D–F) Fully connected models. (G–I) Anatomically constrained models. Top: Sighted variants performing picture viewing task (inputs from visual modality); middle: sighted variants performing name comprehension task (inputs from verbal modality); bottom: blind variants performing name comprehension task (inputs from verbal modality). Error bars are standard errors.

Figure 2. 

Mean activation in lateral and medial ROIs of the visual layer for three models. (A–C) Hub-only models. (D–F) Fully connected models. (G–I) Anatomically constrained models. Top: Sighted variants performing picture viewing task (inputs from visual modality); middle: sighted variants performing name comprehension task (inputs from verbal modality); bottom: blind variants performing name comprehension task (inputs from verbal modality). Error bars are standard errors.

To assess the gradient of the category effect along the lateral/medial axis of the model analog to posterior fusiform, we again replicated the approach of Mahon et al. (2009). At each unit and for each model subject, we computed the t statistic contrasting activation for animals versus tools. Across model subjects, we then computed confidence limits on this t statistic at each unit location along the axis. This analysis was used to generate Figure 3.

Figure 3. 

Contrast-weighted t values of animal versus tool from the lateral to the medial aspect in model simulations and neuroimaging data. Positive t values indicate greater activation for animals (red), and negative t values indicate greater activation for tools (blue). Numbers between panels indicate proportion of shared variance (r2) between datasets (*p < .05, **p < .01). (A–C) Anatomically constrained model; (D–F) empirical data generated by averaging the t values reported in Mahon et al. (2009) across left and right hemispheres; (G–I) fully connected model. Top: Visual viewing (picture comprehension) in sighted individuals/models; middle: word comprehension in sighted individuals/models; bottom: word comprehension in blind individuals/models.

Figure 3. 

Contrast-weighted t values of animal versus tool from the lateral to the medial aspect in model simulations and neuroimaging data. Positive t values indicate greater activation for animals (red), and negative t values indicate greater activation for tools (blue). Numbers between panels indicate proportion of shared variance (r2) between datasets (*p < .05, **p < .01). (A–C) Anatomically constrained model; (D–F) empirical data generated by averaging the t values reported in Mahon et al. (2009) across left and right hemispheres; (G–I) fully connected model. Top: Visual viewing (picture comprehension) in sighted individuals/models; middle: word comprehension in sighted individuals/models; bottom: word comprehension in blind individuals/models.

Results and Discussion

On model analogs of the tasks of interest, both sighted and blind models for all three architectures generated the correct pattern of activation over output units with greater than 90% accuracy across all model training runs. We then assessed whether the category-sensitive activation patterns reported in pFG from prior work arise in the visual hidden layer of each model, using the approach of Mahon et al. (2009) described earlier. Results are shown in Figure 2 and Table 2.

Table 2. 

Descriptive and Statistical Data for the Simulations of Category-specific Activation in Neuroimaging Data

Cat × ROIs InteractionHub-only ModelFully Connected ModelAnatomically Constrained Model
F(1, 14) = 7.395*F(1, 14) = 49.640***F(1, 14) = 72.211***
Lateral Mean (SD)Medial Mean (SD)Lateral Mean (SD)Medial Mean (SD)Lateral Mean (SD)Medial Mean (SD)
Visual viewing (sighted) Animal 0.275 (0.077) 0.377 (0.087) 0.497 (0.101) 0.227 (0.079) 0.459 (0.101) 0.201 (0.069) 
Artifact 0.315 (0.106) 0.266 (0.062) 0.315 (0.111) 0.444 (0.066) 0.324 (0.083) 0.442 (0.067) 
t value −0.946 3.572** 3.578** −7.793*** 3.997** −8.388*** 
 
 Cat × ROIs Interaction F(1, 14) = 8.804** F(1, 14) = 48.711*** F(1, 14) = 65.730*** 
Lateral Mean (SD) Medial Mean (SD) Lateral Mean (SD) Medial Mean (SD) Lateral Mean (SD) Medial Mean (SD) 
Name comprehension (sighted) Animal 0.273 (0.078) 0.381 (0.088) 0.497 (0.103) 0.997 (0.077) 0.463 (0.101) 0.206 (0.071) 
Artifact 0.381 (0.097) 0.261 (0.066) 0.314 (0.108) 0.972 (0.065) 0.321 (0.080) 0.431 (0.068) 
t value −1.063 3.679** 3.625** −7.687*** 4.281** −8.030*** 
 
 Cat × ROIs Interaction F(1, 14) = 0.384 F(1, 14) = 3.286 F(1, 14) = 39.547*** 
Lateral Mean (SD) Medial Mean (SD) Lateral Mean (SD) Medial Mean (SD) Lateral Mean (SD) Medial Mean (SD) 
Name comprehension (blind) Animal 0.017 (0.012) 0.020 (0.014) 0.023 (0.012) 0.126 (0.094) 0.199 (0.062) 0.127 (0.090) 
Artifact 0.016 (0.012) 0.016 (0.010) 0.023 (0.012) 0.147 (0.108) 0.203 (0.061) 0.308 (0.085) 
t value 0.619 1.146 −0.163 −1.786 −.314 −8.154*** 
Cat × ROIs InteractionHub-only ModelFully Connected ModelAnatomically Constrained Model
F(1, 14) = 7.395*F(1, 14) = 49.640***F(1, 14) = 72.211***
Lateral Mean (SD)Medial Mean (SD)Lateral Mean (SD)Medial Mean (SD)Lateral Mean (SD)Medial Mean (SD)
Visual viewing (sighted) Animal 0.275 (0.077) 0.377 (0.087) 0.497 (0.101) 0.227 (0.079) 0.459 (0.101) 0.201 (0.069) 
Artifact 0.315 (0.106) 0.266 (0.062) 0.315 (0.111) 0.444 (0.066) 0.324 (0.083) 0.442 (0.067) 
t value −0.946 3.572** 3.578** −7.793*** 3.997** −8.388*** 
 
 Cat × ROIs Interaction F(1, 14) = 8.804** F(1, 14) = 48.711*** F(1, 14) = 65.730*** 
Lateral Mean (SD) Medial Mean (SD) Lateral Mean (SD) Medial Mean (SD) Lateral Mean (SD) Medial Mean (SD) 
Name comprehension (sighted) Animal 0.273 (0.078) 0.381 (0.088) 0.497 (0.103) 0.997 (0.077) 0.463 (0.101) 0.206 (0.071) 
Artifact 0.381 (0.097) 0.261 (0.066) 0.314 (0.108) 0.972 (0.065) 0.321 (0.080) 0.431 (0.068) 
t value −1.063 3.679** 3.625** −7.687*** 4.281** −8.030*** 
 
 Cat × ROIs Interaction F(1, 14) = 0.384 F(1, 14) = 3.286 F(1, 14) = 39.547*** 
Lateral Mean (SD) Medial Mean (SD) Lateral Mean (SD) Medial Mean (SD) Lateral Mean (SD) Medial Mean (SD) 
Name comprehension (blind) Animal 0.017 (0.012) 0.020 (0.014) 0.023 (0.012) 0.126 (0.094) 0.199 (0.062) 0.127 (0.090) 
Artifact 0.016 (0.012) 0.016 (0.010) 0.023 (0.012) 0.147 (0.108) 0.203 (0.061) 0.308 (0.085) 
t value 0.619 1.146 −0.163 −1.786 −.314 −8.154*** 

*p < .05.

**p < .01.

***p < .001.

In the hub-only architecture, unsurprisingly, no category effect was observed for fully sighted models in the lateral peak for either task. The medial peak showed reliably greater activation for animals than for artifacts in these simulations, likely reflecting the greater number of visual features contained in the visual patterns representing animals. In the blind models, pFG activation was low for both categories. Overall, the pattern was very different from that observed in the imaging data.

Both fully connected and anatomically constrained models replicated the category-sensitive pattern for words and images in fully sighted participants, but the fully connected model did not replicate the pattern observed in the congenitally blind: Greater activation arose in medial relative to lateral units, with no category effect. In contrast, the anatomically constrained model replicated the pattern well: both lateral and medial peaks were moderately active, and the category effect was observed in medial but not lateral peaks.

We next considered how category-sensitive activation patterns varied across the lateral-to-medial axis of the pFG layer. As shown in Figure 3, Mahon et al. (2009) found that the contrast of animal to tool activation shifted gradually along this axis. In sighted individuals, greater activation for animals gives way to the reverse pattern as one moves from lateral toward medial pFG voxels (Figure 3D for pictorial stimuli and E for word stimuli). The congenitally blind participants showed a similar gradient with an important qualitative difference: No category effect was observed in lateral pFG, but this null effect gradually gave way to a reliable effect for tools in medial pFG (Figure 3F).

We computed analogous curves for the anatomically constrained and fully connected model variants (left and right columns of Figure 3). Qualitatively similar patterns were observed for both sighted models, and the gradients in each correlated strongly with those in the imaging study (r2 between .75 and .89 for all comparisons). In the blind models, however, the fully connected architecture failed to capture the key pattern: only a very small effect was observed in the medial-most extent of the pFG layer, and the model gradient correlated only weakly with the observed gradients (r2 = .249). In contrast, the anatomically constrained model again provided a good match to the data: No category effect was observed in the lateral-most units, a strong effect was observed in the medial-most units, and the model gradient correlated strongly (r2 = .697) with the observed gradient.

Note that the ATL hub plays the same functional role in all three networks: It mediates interactions between visual and verbal representations, with all units contributing to the representation of all items. That is, it serves as a domain-general cross-modal hub. The simulations thus illustrate that category-sensitive patterns in pFG can emerge within an interactive neural network that remains domain-general, depending upon the base connectivity of the network.

The contrasting patterns arise from differences in the way activation propagates in the three networks. The blind models must learn associations among names of objects, their verbal descriptions, and their praxis. When all surface representations communicate directly with the ATL units, the hub can directly encode all of these associations. The pFG units never get input and never drive outputs and so can remain largely silent without affecting system behavior. When praxic representations do not connect directly to the ATL hub, the pFG units indirectly mediate these interactions and so become active. The same forces that cause category-sensitive patterns in the sighted models then come into play: Greater connectivity between LP and medial pFG units, coupled with richer praxis for artifacts, causes the medial pFG units to participate more strongly in the representations of artifacts (for similar thoughts, see Almeida, Fintzi, & Mahon, 2013; Mahon et al., 2007).

The model also, however, explains why lateral pFG regions show greater activation for animals in sighted individuals but not in the congenitally blind. In the sighted models, medial pFG units contribute more to the representation of artifacts, so the remaining units must “pick up the slack” in communicating visual information about animals to the ATL hub. Consequently, the lateral units come to more strongly activate for animals, although lateral and medial pFG units are all equally well connected with the ATL hub. In the blind models, there is no visual information about animals to communicate to the hub—the pFG units only serve to communicate interactions between the hub and the praxic representations. Thus, pFG units that are strongly connected to the ATL hub but not to the praxic representations—the lateral units—come to respond equally strongly to items from both domains.

SIMULATION 2: DYSPRAXIA FROM ATL PATHOLOGY

Although results from Simulation 1 provide an account of the brain imaging data, they also raise a puzzle. The direct connection between praxis and the ATL hub was motivated by the dyspraxic pattern arising from ATL pathology in SD (Bozeat et al., 2002; Hodges et al., 2000). If praxic representations do not communicate directly with the ATL hub and do receive direct connections from pFG, what explains this phenomenon? Answering this question was the goal of Simulation 2.

The key data stem from studies of performance on two different tasks in SD requiring their praxic knowledge. In the object use task, participants demonstrate the use of an everyday object, and features of the elicited action are coded as correct or incorrect based on comparison to a control group. Praxic features are considered affordanced if, across many objects, they reliably covary with a particular visual property, and as unaffordanced otherwise, and separate accuracy scores are tabulated for these property types (Bozeat et al., 2002). The central findings are that patients with SD show increasing impairment generating the correct praxis (Figure 4A), but that unaffordanced features are much more vulnerable than affordanced features (Figure 5A).

Figure 4. 

Effect of lesioning hub units on the activation of praxis in sighted models. (A) Accuracy on familiar tool use as a function of the severity of semantic deficit measured by word-picture matching (WPM) in patients with SD (data from Hodges et al., 2000; Figure 3). (B–D) Overall accuracy generating correct praxis for familiar items as a function of the proportion of deleted hub units in the hub-only model (B), the fully connected model (C), and the anatomically constrained model (D).

Figure 4. 

Effect of lesioning hub units on the activation of praxis in sighted models. (A) Accuracy on familiar tool use as a function of the severity of semantic deficit measured by word-picture matching (WPM) in patients with SD (data from Hodges et al., 2000; Figure 3). (B–D) Overall accuracy generating correct praxis for familiar items as a function of the proportion of deleted hub units in the hub-only model (B), the fully connected model (C), and the anatomically constrained model (D).

Figure 5. 

The effect of lesioning hub units on the activation of affordanced (gray) or unaffordanced (black) praxic properties from visual inputs for familiar tool items. Linear regression lines are also shown in the corresponding colors. (A) Patient data from Bozeat et al. (2002); (B) the hub-only model; (C) the fully connected model; (D) the anatomically constrained model.

Figure 5. 

The effect of lesioning hub units on the activation of affordanced (gray) or unaffordanced (black) praxic properties from visual inputs for familiar tool items. Linear regression lines are also shown in the corresponding colors. (A) Patient data from Bozeat et al. (2002); (B) the hub-only model; (C) the fully connected model; (D) the anatomically constrained model.

The second task requires participants to solve simple mechanical puzzles by selecting the correct novel tool from among four options and using it to extract a block from its container. Participants are scored on whether they (1) select the correct novel tool and (2) use it successfully to extract the block. Despite their impaired praxic knowledge for everyday objects, patients with SD perform normally at this task (Figure 7).

Thus, the goal of Simulation 2 was to understand why ATL pathology in SD causes (1) an overall loss of praxic knowledge for everyday objects, but with (2) sparing of affordanced elements of praxis and (3) a spared ability to use novel objects in mechanical puzzle tasks. As noted earlier, previous work has suggested that dyspraxia in SD arises because praxic representations interact directly with the ATL hub, yet Simulation 1 suggested that the functional imaging results of interest will arise only when praxic representations communicate indirectly with the hub via the medial pFG—a hypothesis that is also consistent with recent studies of functional and structural connectivity in cortex. We therefore assessed the ability of each architecture to explain dyspraxia in SD by simulating the effects of the disorder on model analogs of the object use and mechanical puzzle tasks.

Methods

The sighted variants of the same trained models from Simulation 1 were used in this investigation.

Simulating the Object Use Task

To simulate the object use task, the visual patterns for familiar items were applied across visual input units and the pattern generated across the praxis layer was inspected. Overall performance for each item was scored as correct if all praxic features were on the correct side of 0.5 and incorrect otherwise Additionally, the praxic features associated with each item were classified as either affordanced or unaffordanced, according to the criteria of Bozeat et al. (2002). Specifically, praxic features that, in the model training environment, corresponded to an item's visual properties were deemed “affordanced,” whereas praxic properties without presence of corresponding visual properties were deemed “unaffordanced.” We then computed, across all familiar items and at each level of damage, the proportion of affordanced and unaffordanced features correctly activated above a threshold of 0.5.

Simulating the Mechanical Puzzle Task

To simulate the visual appearance of novel tools, we generated new visual input patterns over features common to artifacts. Target praxic patterns were also created for each tool, consisting of the praxic features that covaried with the item's visual features in the training environment. One praxic feature for each novel item was deemed important for solving each puzzle. To simulate tool selection in a particular puzzle, the visual inputs for four novel tools, only one possessing the critical praxic feature, were applied to the model in succession. The model was considered to have chosen whichever tool most strongly activated the critical feature. Use of the selected item was subsequently scored as correct if the model successfully activated all target features and no nontarget features above a threshold of 0.5 in the praxic output layer. Fourteen trials were constructed using the 20 novel tools by arranging them in different groups of 4.

Simulation of ATL Pathology in SD

To simulate ATL pathology observed in SD, the models were subjected to increasingly severe “lesions” created by randomly removing an increasing proportion of units in the ATL hub layer. We investigated removal of 5–30% of units in 5% increments, producing six levels of impairments to the cross-modal representations (Devlin, Gonnerman, Andersen, & Seidenberg, 1998). In each administration of damage to each model, we assessed model performance on analogs of the object use and mechanical puzzle tasks.

Results and Discussion

Model performance was compared to the patient data reported in both Bozeat et al. (2002) and Hodges et al. (2000) and is reported in Table 3. Figure 4 shows patient and model performance for each architecture on the object use task. Correct praxis always declined steadily with increasing disease severity, even in the anatomically constrained architecture where praxic representations are disconnected from the hub. Much larger effects were observed in the hub-only and fully connected architectures, however: Serious dyspraxia arose when only 5% of hub units were removed, contrasting with the patient data where comparably poor praxis is observed only in late stages of the disease.

Table 3. 

Averaged Accuracy Data for the Simulations of Degenerated Praxic Knowledge in SD

Level of Severity (Proportion of Hub Units Lesioned)Overall Accuracy
5%10%15%20%25%30%LesionControl
Familiar Tool Use 
Tool use (item) 
 hub-only 0.51 0.22 0.20 0.09 0.11 0.07 0.20 
 full-conn 0.56 0.39 0.31 0.25 0.19 0.16 0.31 
 anat-cons 0.84 0.75 0.50 0.58 0.46 0.47 0.60 
Affordanced praxic features 
 hub-only 0.89 0.81 0.74 0.66 0.71 0.64 0.74 
 full-conn 0.92 0.90 0.86 0.84 0.82 0.82 0.86 
 anat-cons 0.98 0.97 0.92 0.93 0.91 0.92 0.94 
Unaffordanced praxic features 
 hub-only 0.80 0.61 0.63 0.43 0.56 0.45 0.58 
 full-conn 0.79 0.71 0.66 0.61 0.61 0.63 0.67 
 anat-cons 0.93 0.91 0.78 0.77 0.72 0.74 0.81 
 
Mechanical Puzzle 
Novel tool selection 
 hub-only 0.73 0.65 0.57 0.49 0.53 0.45 0.57 0.82 
 full-conn 0.81 0.77 0.75 0.70 0.67 0.66 0.73 0.81 
 anat-cons 0.82 0.82 0.78 0.79 0.73 0.78 0.79 0.81 
Novel tool use 
 hub-only 0.80 0.73 0.72 0.60 0.76 0.65 0.71 0.79 
 full-conn 0.85 0.84 0.81 0.83 0.79 0.80 0.82 0.89 
 anat-cons 0.83 0.85 0.83 0.85 0.82 0.80 0.83 0.84 
Level of Severity (Proportion of Hub Units Lesioned)Overall Accuracy
5%10%15%20%25%30%LesionControl
Familiar Tool Use 
Tool use (item) 
 hub-only 0.51 0.22 0.20 0.09 0.11 0.07 0.20 
 full-conn 0.56 0.39 0.31 0.25 0.19 0.16 0.31 
 anat-cons 0.84 0.75 0.50 0.58 0.46 0.47 0.60 
Affordanced praxic features 
 hub-only 0.89 0.81 0.74 0.66 0.71 0.64 0.74 
 full-conn 0.92 0.90 0.86 0.84 0.82 0.82 0.86 
 anat-cons 0.98 0.97 0.92 0.93 0.91 0.92 0.94 
Unaffordanced praxic features 
 hub-only 0.80 0.61 0.63 0.43 0.56 0.45 0.58 
 full-conn 0.79 0.71 0.66 0.61 0.61 0.63 0.67 
 anat-cons 0.93 0.91 0.78 0.77 0.72 0.74 0.81 
 
Mechanical Puzzle 
Novel tool selection 
 hub-only 0.73 0.65 0.57 0.49 0.53 0.45 0.57 0.82 
 full-conn 0.81 0.77 0.75 0.70 0.67 0.66 0.73 0.81 
 anat-cons 0.82 0.82 0.78 0.79 0.73 0.78 0.79 0.81 
Novel tool use 
 hub-only 0.80 0.73 0.72 0.60 0.76 0.65 0.71 0.79 
 full-conn 0.85 0.84 0.81 0.83 0.79 0.80 0.82 0.89 
 anat-cons 0.83 0.85 0.83 0.85 0.82 0.80 0.83 0.84 

Hub-only for the hub-only model; full-conn for the fully connected model; anat-cons for the anatomically constrained model.

Figure 5 shows patient and model performance for affordanced versus unaffordanced praxic features plotted against overall severity. Affordanced features fared better than unaffordanced features in all cases, but with qualitative differences across architectures: Both property types declined equally rapidly in hub-only models, whereas unaffordanced properties were much more vulnerable in the anatomically constrained models. The fully connected model showed an intermediate pattern, but with significant loss for both feature types.

We quantified these observations by computing the difference in mean accuracy for affordanced versus unaffordanced properties at each level of damage and correlating this with disease severity. Bozeat et al. (2002) observed that the advantage for affordanced properties grew larger with increasingly severe semantic impairment, with a correlation of r2 = .86, p < .01 across patients. This result was closely matched in the anatomically constrained model (r2 = .88, p < .01), but not the other two architectures (both r2s < 5, ns; Figure 6).

Figure 6. 

Correlation between disease severity and retrieval of affordanced versus unaffordanced praxic features. Regression lines for each property type are shown with r2 and p values in the top right. (A) Patient data replotted from original data of Bozeat et al. (2002); (B) the hub-only model; (C) the fully connected model; (D) the anatomically constrained model.

Figure 6. 

Correlation between disease severity and retrieval of affordanced versus unaffordanced praxic features. Regression lines for each property type are shown with r2 and p values in the top right. (A) Patient data replotted from original data of Bozeat et al. (2002); (B) the hub-only model; (C) the fully connected model; (D) the anatomically constrained model.

Finally, Figure 7 shows patient and model data in the simulation of tool selection and tool use in the mechanical puzzle task. Patients performed as well as controls on average in both aspects of this task. The same was true of the anatomically constrained model, but the other two architectures both showed significant impairment with ATL pathology, averaged across all levels of damage.

Figure 7. 

The effect of lesioning hub units on accuracy for the mechanical puzzle task in patients and sighted models. The overall accuracy (M ± SEM) on the mechanical puzzle task in tool selection (A) and tool use (B) aspects from Bozeat et al. (2002), Hodges et al. (2000), and three models. Patient data from Bozeat et al. were replotted from original data, and those from Hodges et al. were replotted from Figure 1B (Hodges et al., 2000).

Figure 7. 

The effect of lesioning hub units on accuracy for the mechanical puzzle task in patients and sighted models. The overall accuracy (M ± SEM) on the mechanical puzzle task in tool selection (A) and tool use (B) aspects from Bozeat et al. (2002), Hodges et al. (2000), and three models. Patient data from Bozeat et al. were replotted from original data, and those from Hodges et al. were replotted from Figure 1B (Hodges et al., 2000).

Thus, although all three architectures exhibited increasing dyspraxia with greater ATL pathology, only the anatomically constrained model captured the preserved knowledge of affordanced praxis for familiar objects and the preserved ability to select and use novel objects.

These results raise two questions. First, why does ATL pathology disrupt vision-to-praxis mappings for familiar items even in the anatomically constrained model, where praxic representations do not communicate with the damaged hub? The pattern arises because the network is interactive. Activation from visual input flows toward both the hub and the praxic representations and then echoes back to the visual hidden units. Distorted feedback from the damaged hub causes the visual patterns to differ from those arising in the healthy model. The differences propagate to the praxic units, causing some to deactivate and others to incorrectly activate. Greater pathology in ATL produces greater distortion of visual representations, hence greater disruption to praxic patterns generated through the intact direct pathway.

Second, why are afforded properties (and novel items that exploit afforded properties) only spared in the anatomically constrained model? Affordances arise from the systematic co-occurrence of visual and praxic features, correlations that do not always align with conceptual structure. Objects with straight cylindrical handles—hammers, dustpans, spatulas, swords—afford a similar grasp but need not be conceptually related. Moreover their semantic neighbors—nails, brooms, whisks, shields—may lack either or both properties. Thus, affordances may cross-cut semantic category structure in important ways.

Weights in a direct path between visual object and praxis representations can support robust coactivation of correlated visual–praxic pairs. When this mapping must go through the ATL hub, these correlations are less robustly encoded—representations in the hub express conceptual structure and so cannot exploit visual–praxic correlations that cross-cut this structure. Thus, when hub representations degrade, knowledge of both affordanced and unaffordanced praxic properties is rapidly lost. In the anatomically constrained model, disrupted feedback from the ATL can distort visual representations, but because the direct path from vision to praxis robustly encodes affordances, this distortion mainly affects knowledge of unaffordanced properties. The ability to retrieve afforded properties of both familiar and novel items remains unaffected, even with severe ATL pathology.

GENERAL DISCUSSION

The simulations show that the category-sensitive activation patterns seen in both sighted and blind populations can arise from learning in an interactive neural network when (a) there exists a direct but graded pathway between pFG and LP such that medial aspects of pFG connect more effectively than lateral aspects (Mahon et al., 2007, 2013) and (b) there is no direct pathway from LP to the ATL hub, as suggested by classical anatomy and recent diffusion tractography studies (Binney et al., 2012; Caspers et al., 2011; Wakana et al., 2004). A neural network model adopting this connectivity explained both the similarities and differences between sighted and blind populations, whereas two models adopting different connectivity among the same pools of units did not capture the full pattern of imaging results. The results suggest that category-sensitive activation reflect joint effects of learning and base connectivity in a cortical network that is not innately organized by conceptual domain.

Perhaps surprisingly, the same model also captured the dyspraxic pattern caused by anterior temporal pathology in SD. To explain dyspraxia in SD, prior work has suggested that praxic representations communicate directly with the ATL hub (Lambon Ralph et al., 2007; Patterson et al., 2007; Bozeat et al., 2002; Hodges et al., 2000). Although models adopting this architecture did show serious praxic impairments with simulated ATL damage, they failed to capture preserved knowledge of affordanced features of familiar items or the preserved ability to select and use novel tools in simple mechanical puzzles (Bozeat et al., 2002; Hodges et al., 2000). These patterns suggest that knowledge of systematic visuo-praxic relations remains robust when concepts degrade in SD. The anatomically constrained model explains why: Correlations between visual and praxic properties are encoded by weights in the direct path from vision to praxis, which remain unaffected by pathology in SD.

These results are important both in providing the first model of dyspraxia in SD and also in validating the conclusions drawn from the imaging simulations. A critic might reasonably view the anatomically constrained model as having been “tailored” to produce category-sensitive activation in pFG layer, but the same cannot be said of the dyspraxia results: The anatomically constrained model directly contradicts prior proposals about the architecture needed to explain dyspraxia in SD. The fact that this model best captures this very different set of data thus validates the hypothesis about network connectivity developed to explain the imaging data.

Finally, the simulations suggest a novel hypothesis about the forces that constrain initial connectivity in the semantic network. Conceptual and affordance structure, although both important for aspects of behavior, are independent. Conceptual structure promotes knowledge generalization, as when the carpenter realizes she can use either a hammer or a nailgun to build a wall, despite the different praxis these objects engage. Affordance knowledge is critical for interactions with objects generally, and with novel items particularly: The first time one encounters a nailgun, one must infer that the item is to be gripped and pointed differently from the hammer, despite sharing the hammer's function. Conceptual structure, arguably, does not cohere well with affordance structure. A network in which there exist separate pathways for extracting knowledge of affordances versus concepts may best promote both kinds of generalization.

Implications for Domain-specific Approaches to Knowledge Representation

As noted in the Introduction, our proposal adopts many elements central to the DDS view developed by Mahon and Caramazza (2009, 2011). It also resonates with domain-specific proposals more generally in invoking initial constraints to explain category sensitivity in brain activation. Thus, our model might be viewed as an implementation or extension of the DDS hypothesis and as generally consistent with domain-specific approaches. Our model also differs in important respects, however, from both the DDS and other domain-specific views articulated in different writings. Specifically, on our view:

  • Different network components are interactive and not functionally independent. To explain the apparent double dissociation of knowledge for different semantic categories, domain-specific approaches often propose that separate knowledge systems in the mind and brain are functionally independent: The healthy functioning of one system does not depend on the intact functioning of other systems (e.g., Caramazza & Shelton, 1998; Hillis & Caramazza, 1991). Such an argument relies on the modularity assumption adopted in classical cognitive neuropsychology (e.g., Caramazza, 1984). Our model proposes interactive rather than modular processing, with computations in different network components mutually interdependent. In the current context, interactivity explains why anterior temporal pathology causes praxic impairments for familiar items. More generally, interactive processing is a core claim of the domain-general approach to semantic representation and processing (Rogers & McClelland, 2014; Rogers et al., 2004). Double dissociations can arise in interactive networks but do not transparently reveal the functional organization of the system (Plaut, 1995; Farah & McClelland, 1991). On our reading, the question of functional independence is unclear in the DDS. On one hand, Mahon and Caramazza (2008) argue that domain-specific organization does not preclude interactive processing, but elsewhere they argue that category-specific knowledge impairments arise from selective damage to category-specific knowledge systems (Mahon & Caramazza, 2009, 2011), a view that seems to hew closely to the modular assumptions of classical cognitive neuropsychology.

  • Processing for all domains depends upon a central cross-modal hub. Some domain-specific approaches emphasize that different cortical regions are dedicated to different conceptual domains and/or different modalities, with no single region important for all kinds of concepts across all modalities (Caramazza & Shelton, 1998; Damasio, Grabowski, Tranel, Hichwa, & Damasio, 1996). This makes it difficult to understand neuropsychological syndromes like SD, in which circumscribed pathology in the ATL erodes conceptual knowledge for all domains across all modalities. Articulations of the DDS hypothesis have remained silent on this issue and appear to make no mention of this well-studied syndrome (see, e.g., Mahon & Caramazza, 2009). Although category sensitivity emerges in our model in the visuo-praxic pathway, the network still adopts a cross-modal hub important for conceptual representation across all domains and so offers an account of such findings as investigated extensively in prior work (Bozeat et al., 2002; Hodges et al., 2000). If the distributed domain-specific hypothesis permits some parts of the network to be domain-general while others are domain-specific, that would be more consistent with the current view, but we are unaware of any work suggesting this possibility.

  • Learning and experience are important. Some domain-specific accounts of knowledge acquisition downplay the role of learning and experience in advocating for innate knowledge modules, especially in studies of conceptual development (Carey & Spelke, 1994). Cognitive neuropsychology and cognitive neuroscience have generally been more open to the likelihood that neural organization is highly plastic, but the precise role of learning in domain-specific approaches to knowledge in the brain remains unclear. In different articulations of the DDS hypothesis, the authors clearly allow a considerable role for learning and experience (e.g., Mahon & Caramazza, 2011) but also suggest their view is compatible with developmental approaches that eschew such influences (e.g., Mahon & Caramazza, 2009). In our model, initial constraints on connectivity and the effects of learning and experience jointly shape emergent category sensitivity, explaining both the similarities and differences between sighted and congenitally blind groups. We also note that initially specified connectivity does not preclude the possibility that white matter connectivity itself can be partly shaped by experience (Tang et al., 2010; Scholz, Klein, Behrens, & Johansen-Berg, 2009), further highlighting the potentially important role of experience.

  • Category effects are overlapping, graded and fuzzy. Because category effects in our model arise from learning and experience, they are not all-or-nothing but overlapping and graded. Within the model, artifacts that are not frequent targets of praxis will not strongly drive the visuo-praxic system. For individuals who interact physically with animals (e.g., farmers), these animals may drive the visuo-praxic system and so elicit similar patterns to manipulable manmade objects. Moreover, units in medial FG that activate more strongly for tools may nevertheless still contribute somewhat to the representation of animals and vice versa, similar to Plaut's (2002) account of graded functional specialization in optic aphasia. This perspective seems very different from a domain-specific view under which different neural systems are natively tailored to only process a particular conceptual domain and do not contribute to other domains. Whether our view is consistent with the DDS hypothesis in this respect depends upon issues noted in the introduction—in particular, on what is intended by the proposal that initial connectivity is “concentrated around” different conceptual domains. If the claim is that a given tract only propagates information about items in a given domain, this seems clearly in opposition to our proposal. Likewise, if the claim is that initial connectivity expresses innate knowledge of property co-occurrences—for instance, knowledge that things with faces likely can move on their own and have emotional states, which might be reflected by innate connectivity between face, motion, and limbic regions—this also differs from our proposal, under which such co-occurrences are learned. If, however, the claim is simply that different conceptual domains come to rely more on some tracts than others by virtue of the surface representations they engage, this seems more consistent with our proposal, although less consistent with earlier domain-specific views.

  • Gross connectivity may be innately specified, but knowledge is not. Domain-specific approaches in cognitive development propose that infants are born with knowledge that allows them to construe items as belonging to distinct core conceptual domains, such as physical objects, animals, agents, or social beings (for discussion, see Mahon & Caramazza, 2009; Carey & Spelke, 1994). Under our proposal, the gross connectivity of the cortical network is initially specified, but this does not reflect innate knowledge about distinct domains and their properties. Indeed, before learning our model treats all items as similar to one another. The proposed constraints on initial connectivity instead promote, with experience, efficient acquisition of different kinds of internal representations, specifically conceptual and affordance-based representations. To us, it is unclear whether the initial constraints on connectivity proposed by the DDS hypothesis constitute innate knowledge as proposed in nativist approaches to development, or whether these constraints operate only in conjunction with learning and experience as in our proposal.

  • Initial constraints distinguish conceptual from affordance structure. Finally, initial connectivity in our model promotes acquisition of two different kinds of knowledge structures—conceptual structures and affordance structures. Contra the domain-specific view, these do not align with different semantic categories. Conceptual structures describe the degree to which different items are similar in kind and thus promote conceptual generalization, for all semantic domains. Affordance structures describe systematic relations between an item's visual appearance and the kinds of actions that it can support. Because tools are frequently associated with praxis, they will strongly drive this pathway, but the affordance system should also support action directed toward items in other semantic domains. For instance, although faces are indeed important properties of animals and must contribute to their semantic representation, they can also be targets of praxis, as anyone who brushed crumbs from a child's cheek or kissed their partner can attest. Clearly actions are also directed toward foods and toward other body parts—categories that, under some domain-specific views, are supported by distinct neural systems (Caramazza & Shelton, 1998). Thus, although our proposal accords with the DDS in suggesting that initial connectivity supports acquisition of different kinds of knowledge, the particular structures invoked are very different: not distinct semantic domains, but distinct similarity structures.

In summary, we believe our proposal occupies a useful middle ground between strict domain-specific views and blank-slate domain-general views, similar to both the DDS hypothesis and to recent proposals in other cognitive domains (Plaut & Behrmann, 2011; Ueno, Saito, Rogers, & Lambon Ralph, 2011; Zevin & Seidenberg, 2006). We further believe, however, that it differs from the DDS hypothesis in several respects and so is best viewed as an alternative proposal rather than as a model implementation of that view.

Implications for the ATL “Hub” Theory

Despite invoking separate pathways for visual–praxic and visuo-conceptual interactions, our model remains consistent with the ATL hub theory. The theory's central claim is that a cross-modal hub is needed to learn the conceptual structure that governs knowledge generalization for all conceptual domains. We have argued that behavioral interactions with objects require knowledge of a different kind of structure—affordance structure—that can often cross-cut conceptual structure. Where different behaviors depend upon cross-cutting statistical structure, it makes sense that different pathways should be involved in encoding those structures. This proposal is broadly similar to arguments developed in the context of reading (Harm &Seidenberg, 2004; Plaut, McClelland, Seidenberg, & Patterson, 1996) and spoken language processing (Ueno et al., 2011), where the systematic structure existing among perceptual/motor representations of words cross-cuts the conceptual structure expressed by word meaning. To exploit both forms of structure, contemporary models likewise propose separate pathways for encoding the different mappings.

Implications for Other Theories of Category Specificity

Finally, we note that the current proposals about category-sensitive functional activations in the brain are similar in spirit to the well-known sensory/functional hypothesis about category-specific semantic impairment observed in some forms of neuropathology. This view, first articulated in Warrington and Shallice's classic work (1984) and subsequently developed in computer simulations by Farah and McClelland (1991), proposed that knowledge of different semantic domains might rely differentially on knowledge of perceptual versus functional properties of items in the environment. This view was appealing partly because it explained the interesting neuropsychological patterns with reference to a division of labor in the brain that was already well established—specifically, the anatomical division between sensory and motor systems. There was no need to postulate additional cortical and cognitive modules dedicated to particular semantic kinds to explain the selective disruption of particular categories of knowledge. The sensory functional view came under fire for a variety of reasons in the 1990s (see Caramazza & Shelton, 1998), but there is now considerable evidence from both patient and functional imaging work for an elaborated variant of this perspective, under which there are not merely two but many different kinds of sensory and motor properties represented in different regions throughout cortex (see Chen & Rogers, 2014; but see Mahon & Caramazza, 2009, for recent criticisms). The current work illustrates how differential reliance on different kinds of sensory–motor properties, coupled with additional constraints arising from patterns of connectivity among the regions that code these properties, can explain a range of phenomena in both brain imaging and semantic disorders. We view the work as highly consistent with the original motivations for the sensory-functional hypothesis.

Limitations and Future Directions

Our simulations illustrate how base network connectivity can produce category-sensitive activation patterns in both sighted and blind populations within a network that is not innately organized by conceptual domain. In so doing, it also provides a novel account of dyspraxia in SD. We view these contributions as promising, but the model is also limited in several respects.

First, the literature contains several additional claims of category- and modality-sensitive activation patterns in sighted and blind groups (Bedny et al., 2012; Mahon et al., 2010). Some results are highly provocative, such as the finding that the “visual word-form area” of congenitally blind individuals becomes active during tactile reading (Reich, Szwed, Cohen, & Amedi, 2011). In focusing on mappings among visual, verbal, and praxic representations of objects, our model omits many other kinds of information and so cannot directly address these new findings. An important question for future work will be whether a better elaborated model that incorporates haptic and orthographic representations can extend the same principles to these and other related findings (Yildirim & Jacobs, 2013).

Second, our simulations motivate further study of connectivity between ventral temporal and parietal regions. We have argued that medial pFG connects to praxic representations in the parietal cortex, but object-directed action also involves knowledge of the functions to which objects are put. Like conceptual structure, functional structure varies independently of praxis (Boronat et al., 2005; Schwartz et al., 1998; Buxbaum, Schwartz, & Carew, 1997)—whereas the typewriter and piano engage similar movements toward different ends, the typewriter and pen engage different movements toward similar ends (Kellenbach, Brett, & Patterson, 2003). Functional properties critically constrain conceptual representations of artifacts—so much so that artifacts are often viewed as changing in kind when they change in function (Ahn, 1998; Keil, 1992). Thus, whereas praxic representations may not strongly constrain the representations that arise in the ATL hub, function representations likely do.

Knowledge about praxis and function can doubly dissociate with brain damage and evoke different activation patterns within parietal cortex (Buxbaum & Saffran, 2002; Schwartz et al., 1998). Buxbaum (Binkofski & Buxbaum, 2013; Boronat et al., 2005; Buxbaum & Saffran, 2002) thus proposes a division of labor within LP cortex, with a “dorso-dorsal” stream mainly supporting praxic knowledge and an “ventro-dorsal” stream mainly supporting functional knowledge. Both superior and inferior aspects of the LP lobe appear to engage in tool concept processing (Anzellotti, Mahon, Schwarzbach, & Caramazza, 2011; Almeida et al., 2010; Canessa et al., 2008; Noppeney, Price, Penny, & Friston, 2006) and show category-specific activation in naming (Chouinard & Goodale, 2010). If inferior regions like supramarginal gyrus mainly encode tool-related function while praxis is mainly supported by more superior regions, an important test for our hypothesis concerns the nature of the connectivity between medial pFG and superior versus inferior parietal regions. Our framework suggests that medial pFG should be more effectively connected to praxic (superior parietal) than functional (inferior) representations.

On our reading the literature here is somewhat mixed. Mahon has emphasized functional connectivity between medial pFG and supramarginal gyrus in two studies (Garcea & Mahon, 2014; Mahon et al., 2007) but also reports significant nonzero connectivity from medial pFG to more superior parietal cortex in a third (Mahon et al., 2013; Figures 2 and 4B). To our knowledge there is no evidence from any method for direct white matter projections between medial pFG and supra-marginal gyrus. Animal studies (Zhong & Rockland, 2003; Seltzer & Pandya, 1994) have found a long projection from the inferior parietal lobe into various temporal regions, although with little evidence of a lateral/medial gradient. In humans, both dissection studies (Martino, Brogna, Robles, Vergani, & Duffau, 2010) and diffusion tensor imaging (Wakana et al., 2004) have shown white matter pathways that branch up from medial infero-posterior temporal cortex to terminate in superior parietal regions. We believe this supports the face validity of our proposal, although this is clearly an avenue for further research.

Third, theoretical approaches to knowledge representation in the brain have been strongly influenced by patterns of spared and impaired semantic memory across different neuropsychological groups. The current model builds on an existing account of semantic impairments in SD and herpes simplex viral encephalitis, but some semantic deficits—most notably cases in which knowledge of manmade objects is more impaired than knowledge of animals—have not been explored in this framework. A model that incorporates both praxic and functional representations and whose architecture is constrained by what is known about the relevant connectivity of cortex may allow us to better understand the full spectrum of semantic knowledge impairments observed in neuropsychology. One interesting direction for future research could be investigating how semantic impairment of different neuropsychological groups may or may not be modulated by their premorbid experience (i.e., congenitally blind patients). Although these individual cases would be rather difficult to encounter, a recent study has demonstrated that this approach can be extremely fruitful for answering important theoretical questions (Jefferies, Rogers, & Lambon Ralph, 2011).

Finally, our training environment consists of artificial representations that capture important aspects of category structure within and across modalities. Although widely used in simulation research (Lambon Ralph et al., 2007; Rogers et al., 2004; Plaut, 1997), such patterns are highly abstracted from the real sensory–motor structure of the world. The development of models that can operate on representations derived from real images (Plaut & Behrmann, 2011; Gale, Done, & Frank, 2001) may provide further insight into the relation between learning, connectivity, and representational structure.

Conclusion

We have developed an approach to knowledge representation in the brain under which activation patterns generated by words or pictures are influenced both by learning and by the native connectivity of the network architecture. The approach explains similarities and differences in activation patterns observed in sighted and congenitally blind participants and further accounts for the dyspraxia arising from anterior temporal pathology in SD. The work provides a new perspective on the interaction of nature and nurture in conceptual knowledge acquisition, suggesting that evolution may have shaped the base connectivity of the cortical knowledge network so as to promote efficient learning of both conceptual and affordance structure.

Acknowledgments

This work was conducted with generous support from the University of Wisconsin–Madison in the form of a University Fellowship to the first author and a Romnes Award to the second author. We thank Bradford Mahon and Matthew Lambon Ralph for kindly sharing their empirical data with us.

Reprint requests should be sent to Lang Chen, Stanford Cognitive and Systems Neuroscience Lab, Stanford University, 1070 Arastradero Road Suite 220, Palo Alto, CA 94304, or via e-mail: lchen32@wisc.edu.

Notes

1. 

This scoring protocol is very commonly used in neural network modeling; however, other common protocols produced a qualitatively similar pattern of behavior.

2. 

We also considered “blind” models given more extensive training with nonvisual patterns, but their behavior was qualitatively very similar to the “blind” models reported here so we do not discuss these further.

3. 

An alternative approach in which we simply divided the visual hidden units into a “medial” and “lateral” group produced qualitatively similar results.

REFERENCES

Adlam
,
A. L. R.
,
Patterson
,
K.
,
Rogers
,
T. T.
,
Nestor
,
P. J.
,
Salmond
,
C. H.
,
Acosta-Cabronero
,
J.
, et al
(
2006
).
Semantic dementia and fluent primary progressive aphasia: Two sides of the same coin?
Brain
,
129
,
3066
3080
.
Ahn
,
W.
(
1998
).
Why are different features central for natural kinds and artifacts?: The role of causal status in determining feature centrality
.
Cognition
,
69
,
135
178
.
Almeida
,
J.
,
Fintzi
,
A. R.
, &
Mahon
,
B. Z.
(
2013
).
Tool manipulation knowledge is retrieved by way of the ventral visual object processing pathway
.
Cortex
,
49
,
2334
2344
.
Almeida
,
J.
,
Mahon
,
B. Z.
, &
Caramazza
,
A.
(
2010
).
The role of the dorsal visual processing stream in tool identification
.
Psychological Science
,
21
,
772
778
.
Anzellotti
,
S.
,
Mahon
,
B. Z.
,
Schwarzbach
,
J.
, &
Caramazza
,
A.
(
2011
).
Differential activity for animals and manipulable objects in the anterior temporal lobes
.
Journal of Cognitive Neuroscience
,
23
,
2059
2067
.
Bedny
,
M.
,
Caramazza
,
A.
,
Pascual-Leone
,
A.
, &
Saxe
,
R.
(
2012
).
Typical neural representations of action verbs develop without vision
.
Cerebral Cortex
,
22
,
286
293
.
Binkofski
,
F.
, &
Buxbaum
,
L. J.
(
2013
).
Two action systems in the human brain
.
Brain and Language
,
127
,
222
229
.
Binney
,
R. J.
,
Parker
,
G. J.
, &
Lambon Ralph
,
M. A. L.
(
2012
).
Convergent connectivity and graded specialization in the rostral human temporal lobe as revealed by diffusion-weighted imaging probabilistic tractography
.
Journal of Cognitive Neuroscience
,
24
,
1998
2014
.
Boronat
,
C. B.
,
Buxbaum
,
L. J.
,
Coslett
,
H. B.
,
Tang
,
K.
,
Saffran
,
E. M.
,
Kimberg
,
D. Y.
, et al
(
2005
).
Distinctions between manipulation and function knowledge of objects: Evidence from functional magnetic resonance imaging
.
Cognitive Brain Research
,
23
,
361
373
.
Bouhali
,
F.
,
de Schotten
,
M. T.
,
Pinel
,
P.
,
Poupon
,
C.
,
Mangin
,
J.-F.
,
Dehaene
,
S.
, et al
(
2014
).
Anatomical connections of the visual word form area
.
Journal of Neuroscience
,
34
,
15402
15414
.
Bozeat
,
S.
,
Lambon Ralph
,
M. A.
,
Patterson
,
K.
, &
Hodges
,
J. R.
(
2002
).
When objects lose their meaning: What happens to their use?
Cognitive, Affective, & Behavioral Neuroscience
,
2
,
236
251
.
Buxbaum
,
L. J.
,
Kyle
,
K.
,
Grossman
,
M.
, &
Coslett
,
B.
(
2007
).
Left inferior parietal representations for skilled hand-object interactions: Evidence from stroke and corticobasal degeneration
.
Cortex
,
43
,
411
423
.
Buxbaum
,
L. J.
, &
Saffran
,
E. M.
(
2002
).
Knowledge of object manipulation and object function: Dissociations in apraxic and nonapraxic subjects
.
Brain and Language
,
82
,
179
199
.
Buxbaum
,
L. J.
,
Schwartz
,
M. F.
, &
Carew
,
T. G.
(
1997
).
The role of semantic memory in object use
.
Cognitive Neuropsychology
,
14
,
219
254
.
Canessa
,
N.
,
Borgo
,
F.
,
Cappa
,
S. F.
,
Perani
,
D.
,
Falini
,
A.
,
Buccino
,
G.
, et al
(
2008
).
The different neural correlates of action and functional knowledge in semantic memory: An fMRI study
.
Cerebral Cortex
,
18
,
740
751
.
Caramazza
,
A.
(
1984
).
The logic of neuropsychological research and the problem of patient classification in aphasia
.
Brain and Language
,
21
,
9
20
.
Caramazza
,
A.
, &
Mahon
,
B. Z.
(
2003
).
The organization of conceptual knowledge: The evidence from category-specific semantic deficits
.
Trends in Cognitive Sciences
,
7
,
354
361
.
Caramazza
,
A.
, &
Shelton
,
J. R.
(
1998
).
Domain-specific knowledge systems in the brain: The animate-inanimate distinction
.
Journal of Cognitive Neuroscience
,
10
,
1
34
.
Carey
,
S.
, &
Spelke
,
E.
(
1994
).
Domain-specific knowledge and conceptual change
. In
L. A.
Hirschfeld
&
S. A.
Gelman
(Eds.),
Mapping the mind: Domain specificity in cognition and culture
(pp.
169
200
).
Cambridge
:
Cambridge University Press
.
Caspers
,
S.
,
Eickhoff
,
S. B.
,
Rick
,
T.
,
von Kapri
,
A.
,
Kuhlen
,
T.
,
Huang
,
R.
, et al
(
2011
).
Probabilistic fibre tract analysis of cytoarchitectonically defined human inferior parietal lobule areas reveals similarities to macaques
.
Neuroimage
,
58
,
362
380
.
Catani
,
M.
,
Jones
,
D. K.
, &
Ffytche
,
D. H.
(
2005
).
Perisylvian language networks of the human brain
.
Annals of Neurology
,
57
,
8
16
.
Chao
,
L. L.
,
Haxby
,
J. V.
, &
Martin
,
A.
(
1999
).
Attribute-based neural substrates in temporal cortex for perceiving and knowing about objects
.
Nature Neuroscience
,
2
,
913
.
Chen
,
L.
, &
Rogers
,
T. T.
(
2014
).
Revisiting domain—General accounts of category specificity in mind and brain
.
Wiley Interdisciplinary Reviews: Cognitive Science
,
5
,
327
344
.
Chouinard
,
P. A.
, &
Goodale
,
M. A.
(
2010
).
Category-specific neural processing for naming pictures of animals and naming pictures of tools: An ALE meta-analysis
.
Neuropsychologia
,
48
,
409
.
Cree
,
G. S.
,
McNorgan
,
C.
, &
McRae
,
K.
(
2006
).
Distinctive features hold a privileged status in the computation of word meaning: Implications for theories of semantic memory
.
Journal of Experimental Psychology: Learning, Memory, and Cognition
,
32
,
643
658
.
Damasio
,
H.
,
Grabowski
,
T. J.
,
Tranel
,
D.
,
Hichwa
,
R. D.
, &
Damasio
,
A. R.
(
1996
).
A neural basis for lexical retrieval
.
Nature
,
380
,
499
505
.
Davis
,
M. H.
,
Di Betta
,
A. M.
,
Macdonald
,
M. J.
, &
Gaskell
,
M. G.
(
2009
).
Learning and consolidation of novel spoken words
.
Journal of Cognitive Neuroscience
,
21
,
803
820
.
Devlin
,
J. T.
,
Gonnerman
,
L. M.
,
Andersen
,
E. S.
, &
Seidenberg
,
M. S.
(
1998
).
Category-specific semantic deficits in focal and widespread brain damage: A computational account
.
Journal of Cognitive Neuroscience
,
10
,
77
94
.
Devlin
,
J. T.
,
Rushworth
,
M. F. S.
, &
Matthews
,
P. M.
(
2005
).
Category-related activation for written words in the posterior fusiform is task specific
.
Neuropsychologia
,
43
,
69
74
.
Farah
,
M. J.
, &
McClelland
,
J. L.
(
1991
).
A computational model of semantic memory impairment: Modality specificity and emergent category specificity
.
Journal of Experimental Psychology: General
,
120
,
339
357
.
Gale
,
T. M.
,
Done
,
D. J.
, &
Frank
,
R. J.
(
2001
).
Visual crowding and category specific deficits for pictorial stimuli: A neural network model
.
Cognitive Neuropsychology
,
18
,
509
550
.
Garcea
,
F. E.
, &
Mahon
,
B. Z.
(
2014
).
Parcellation of left parietal tool representations by functional connectivity
.
Neuropsychologia
,
60
,
131
143
.
Goodale
,
M. A.
, &
Milner
,
A. D.
(
1992
).
Separate visual pathways for perception and action
.
Trends in Neurosciences
,
15
,
20
25
.
Graves
,
W. W.
,
Desai
,
R.
,
Humphries
,
C.
,
Seidenberg
,
M. S.
, &
Binder
,
J. R.
(
2010
).
Neural systems for reading aloud: A multiparametric approach
.
Cerebral Cortex
,
20
,
1799
1815
.
Guo
,
C. C.
,
Gorno-Tempini
,
M. L.
,
Gesierich
,
B.
,
Henry
,
M.
,
Trujillo
,
A.
,
Shany-Ur
,
T.
, et al
(
2013
).
Anterior temporal lobe degeneration produces widespread network-driven dysfunction
.
Brain
,
136
,
2979
2991
.
Harm
,
M. W.
, &
Seidenberg
,
M. S.
(
2004
).
Computing the meanings of words in reading: Cooperative division of labor between visual and phonological processes
.
Psychological Review
,
111
,
662
720
.
He
,
C.
,
Peelen
,
M. V.
,
Han
,
Z.
,
Lin
,
N.
,
Caramazza
,
A.
, &
Bi
,
Y.
(
2013
).
Selectivity for large nonmanipulable objects in scene-selective visual cortex does not require visual experience
.
Neuroimage
,
79
,
1
9
.
Hillis
,
A. E.
, &
Caramazza
,
A.
(
1991
).
Mechanisms for accessing lexical representations for output: Evidence from a category-specific semantic deficit
.
Brain and Language
,
40
,
106
144
.
Hodges
,
J. R.
,
Bozeat
,
S.
,
Lambon Ralph
,
M. A.
,
Patterson
,
K.
, &
Spatt
,
J.
(
2000
).
The role of conceptual knowledge in object use Evidence from semantic dementia
.
Brain
,
123
,
1913
1925
.
Hodges
,
J. R.
,
Patterson
,
K.
,
Oxbury
,
S.
, &
Funnell
,
E.
(
1992
).
Semantic dementia progressive fluent aphasia with temporal lobe atrophy
.
Brain
,
115
,
1783
1806
.
Hwang
,
K.
,
Palmer
,
E. D.
,
Basho
,
S.
,
Zadra
,
J. R.
, &
Müller
,
R.-A.
(
2009
).
Category-specific activations during word generation reflect experiential sensorimotor modalities
.
Neuroimage
,
48
,
717
725
.
Jefferies
,
E.
,
Rogers
,
T. T.
, &
Lambon Ralph
,
M. A.
(
2011
).
Premorbid expertise produces category-specific impairment in a domain-general semantic disorder
.
Neuropsychologia
,
49
,
3213
3223
.
Kanwisher
,
N.
(
2010
).
Functional specificity in the human brain: A window into the functional architecture of the mind
.
Proceedings of the National Academy of Sciences, U.S.A.
,
107
,
11163
11170
.
Keil
,
F. C.
(
1992
).
Concepts, kinds, and cognitive development
.
Cambridge, MA
:
MIT Press
.
Kellenbach
,
M. L.
,
Brett
,
M.
, &
Patterson
,
K.
(
2003
).
Actions speak louder than functions: The importance of manipulability and action in tool representation
.
Journal of Cognitive Neuroscience
,
15
,
30
46
.
Lambon Ralph
,
M. A.
,
Lowe
,
C.
, &
Rogers
,
T. T.
(
2007
).
Neural basis of category-specific semantic deficits for living things: Evidence from semantic dementia, HSVE and a neural network model
.
Brain
,
130
,
1127
1137
.
Lambon Ralph
,
M. A.
,
Patterson
,
K.
,
Garrard
,
P.
, &
Hodges
,
J. R.
(
2003
).
Semantic dementia with category specificity: A comparative case-series study
.
Cognitive Neuropsychology
,
20
,
307
326
.
Mahon
,
B. Z.
,
Anzellotti
,
S.
,
Schwarzbach
,
J.
,
Zampini
,
M.
, &
Caramazza
,
A.
(
2009
).
Category-specific organization in the human brain does not require visual experience
.
Neuron
,
63
,
397
405
.
Mahon
,
B. Z.
, &
Caramazza
,
A.
(
2008
).
A critical look at the embodied cognition hypothesis and a new proposal for grounding conceptual content
.
Journal of Physiology-Paris
,
102
,
59
70
.
Mahon
,
B. Z.
, &
Caramazza
,
A.
(
2009
).
Concepts and categories: A cognitive neuropsychological perspective
.
Annual Review of Psychology
,
60
,
27
51
.
Mahon
,
B. Z.
, &
Caramazza
,
A.
(
2011
).
What drives the organization of object knowledge in the brain?
Trends in Cognitive Sciences
,
15
,
97
103
.
Mahon
,
B. Z.
,
Kumar
,
N.
, &
Almeida
,
J.
(
2013
).
Spatial frequency tuning reveals interactions between the dorsal and ventral visual systems
.
Journal of Cognitive Neuroscience
,
25
,
862
871
.
Mahon
,
B. Z.
,
Milleville
,
S. C.
,
Negri
,
G. A. L.
,
Rumiati
,
R. I.
,
Caramazza
,
A.
, &
Martin
,
A.
(
2007
).
Action-related properties shape object representations in the ventral stream
.
Neuron
,
55
,
507
520
.
Mahon
,
B. Z.
,
Schwarzbach
,
J.
, &
Caramazza
,
A.
(
2010
).
The representation of tools in left parietal cortex is independent of visual experience
.
Psychological Science
,
21
,
764
771
.
Martin
,
A.
(
2006
).
Shades of Déjerine—Forging a causal link between the visual word form area and reading
.
Neuron
,
50
,
173
175
.
Martin
,
A.
, &
Chao
,
L. L.
(
2001
).
Semantic memory and the brain: Structure and processes
.
Current Opinion in Neurobiology
,
11
,
194
201
.
Martin
,
A.
,
Wiggs
,
C. L.
,
Ungerleider
,
L. G.
, &
Haxby
,
J. V.
(
1996
).
Neural correlates of category-specific knowledge
.
Nature
,
379
,
649
652
.
Martino
,
J.
,
Brogna
,
C.
,
Robles
,
S. G.
,
Vergani
,
F.
, &
Duffau
,
H.
(
2010
).
Anatomic dissection of the inferior fronto-occipital fasciculus revisited in the lights of brain stimulation data
.
Cortex
,
46
,
691
699
.
McRae
,
K.
,
de Sa
,
V. R.
, &
Seidenberg
,
M. S.
(
1997
).
On the nature and scope of featural representations of word meaning
.
Journal of Experimental Psychology: General
,
126
,
99
130
.
New
,
J.
,
Cosmides
,
L.
, &
Tooby
,
J.
(
2007
).
Category-specific attention for animals reflects ancestral priorities, not expertise
.
Proceedings of the National Academy of Sciences, U.S.A.
,
104
,
16598
16603
.
Noppeney
,
U.
,
Patterson
,
K.
,
Tyler
,
L. K.
,
Moss
,
H.
,
Stamatakis
,
E. A.
,
Bright
,
P.
, et al
(
2007
).
Temporal lobe lesions and semantic impairment: A comparison of herpes simplex virus encephalitis and semantic dementia
.
Brain
,
130
,
1138
1147
.
Noppeney
,
U.
,
Price
,
C. J.
,
Penny
,
W. D.
, &
Friston
,
K. J.
(
2006
).
Two distinct neural mechanisms for category-selective responses
.
Cerebral Cortex
,
16
,
437
445
.
Patterson
,
K.
,
Nestor
,
P. J.
, &
Rogers
,
T. T.
(
2007
).
Where do you know what you know? The representation of semantic knowledge in the human brain
.
Nature Reviews Neuroscience
,
8
,
976
987
.
Patterson
,
K.
,
Ralph
,
M. A. L.
,
Jefferies
,
E.
,
Woollams
,
A.
,
Jones
,
R.
,
Hodges
,
J.
, et al
(
2006
).
“Presemantic” cognition in semantic dementia: Six deficits in search of an explanation
.
Journal of Cognitive Neuroscience
,
18
,
169
183
.
Plaut
,
D. C.
(
1995
).
Double dissociation without modularity: Evidence from connectionist neuropsychology
.
Journal of Clinical and Experimental Neuropsychology
,
17
,
291
321
.
Plaut
,
D. C.
(
1997
).
Structure and function in the lexical system: Insights from distributed models of word reading and lexical decision
.
Language & Cognitive Processes
,
12
,
765
806
.
Plaut
,
D. C.
(
2002
).
Graded modality-specific specialisation in semantics: A computational account of optic aphasia
.
Cognitive Neuropsychology
,
19
,
603
639
.
Plaut
,
D. C.
, &
Behrmann
,
M.
(
2011
).
Complementary neural representations for faces and words: A computational exploration
.
Cognitive Neuropsychology
,
28
,
251
275
.
Plaut
,
D. C.
,
McClelland
,
J. L.
,
Seidenberg
,
M. S.
, &
Patterson
,
K.
(
1996
).
Understanding normal and impaired word reading: Computational principles in quasi-regular domains
.
Psychological Review
,
103
,
56
115
.
Reich
,
L.
,
Szwed
,
M.
,
Cohen
,
L.
, &
Amedi
,
A.
(
2011
).
A ventral visual stream reading center independent of visual experience
.
Current Biology
,
21
,
363
368
.
Rizzolatti
,
G.
, &
Matelli
,
M.
(
2003
).
Two different streams form the dorsal visual system: Anatomy and functions
.
Experimental Brain Research
,
153
,
146
157
.
Rogers
,
T. T.
,
Hocking
,
J.
,
Mechelli
,
A.
,
Patterson
,
K.
, &
Price
,
C.
(
2005
).
Fusiform activation to animals is driven by the process, not the stimulus
.
Journal of Cognitive Neuroscience
,
17
,
434
445
.
Rogers
,
T. T.
,
Lambon Ralph
,
M. A.
,
Garrard
,
P.
,
Bozeat
,
S.
,
McClelland
,
J. L.
,
Hodges
,
J. R.
, et al
(
2004
).
Structure and deterioration of semantic memory: A neuropsychological and computational investigation
.
Psychological Review
,
111
,
205
235
.
Rogers
,
T. T.
, &
McClelland
,
J.
(
2004
).
Semantic cognition: A parallel distributed processing approach
.
Cambridge, MA
:
MIT Press
.
Rogers
,
T. T.
, &
McClelland
,
J. L.
(
2008
).
Précis of semantic cognition: A parallel distributed processing approach
.
Behavioral and Brain Sciences
,
31
,
689
714
.
Rogers
,
T. T.
, &
McClelland
,
J. L.
(
2014
).
Parallel distributed processing at 25: Further explorations in the microstructure of cognition
.
Cognitive Science
,
38
,
1024
1077
.
Rogers
,
T. T.
, &
Patterson
,
K.
(
2007
).
Object categorization: Reversals and explanations of the basic-level advantage
.
Journal of Experimental Psychology: General
,
136
,
451
469
.
Rogers
,
T. T.
, &
Plaut
,
D. C.
(
2002
).
Connectionist perspectives on category-specific deficits
. In
G. W.
Humphreys
&
E.
Forde
(Eds.),
Category-specificity in brain and mind
(pp.
251
290
).
East Essex, UK
:
Psychology Press
.
Rohde
,
D. L. T.
(
1999
).
LENS: The light, efficient network simulator
.
Pittsburgh, PA
:
Carnegie Mellon University, Department of Computer Science
.
Rumelhart
,
D. E.
,
Hinton
,
G. E.
, &
Williams
,
R. J.
(
1988
).
Learning representations by back-propagating errors
.
Cambridge, MA
:
MIT Press
.
Scholz
,
J.
,
Klein
,
M. C.
,
Behrens
,
T. E. J.
, &
Johansen-Berg
,
H.
(
2009
).
Training induces changes in white matter architecture
.
Nature Neuroscience
,
12
,
1370
1371
.
Schwartz
,
M. F.
,
Buxbaum
,
L. J.
,
Montgomery
,
M. W.
,
Fitzpatrick-DeSalme
,
E.
,
Hart
,
T.
,
Ferraro
,
M.
, et al
(
1998
).
Naturalistic action production following right hemisphere stroke
.
Neuropsychologia
,
37
,
51
66
.
Seltzer
,
B.
, &
Pandya
,
D. N.
(
1994
).
Parietal, temporal, and occipital projections to cortex of the superior temporal sulcus in the rhesus monkey: A retrograde tracer study
.
The Journal of Comparative Neurology
,
343
,
445
463
.
Tang
,
Y.-Y.
,
Lu
,
Q.
,
Geng
,
X.
,
Stein
,
E. A.
,
Yang
,
Y.
, &
Posner
,
M. I.
(
2010
).
Short-term meditation induces white matter changes in the anterior cingulate
.
Proceedings of the National Academy of Sciences
,
107
,
15649
15652
.
Tyler
,
L. K.
,
Moss
,
H. E.
,
Durrant-Peatfield
,
M. R.
, &
Levy
,
J. P.
(
2000
).
Conceptual structure and the structure of concepts: A distributed account of category-specific deficits
.
Brain and Language
,
75
,
195
231
.
Ueno
,
T.
,
Saito
,
S.
,
Rogers
,
T. T.
, &
Lambon Ralph
,
M. A.
(
2011
).
Lichtheim 2: Synthesizing aphasia and the neural basis of language in a neurocomputational model of the dual dorsal-ventral language pathways
.
Neuron
,
72
,
385
396
.
Wakana
,
S.
,
Jiang
,
H.
,
Nagae-Poetscher
,
L. M.
,
van Zijl
,
P. C. M.
, &
Mori
,
S.
(
2004
).
Fiber tract-based atlas of human white matter anatomy
.
Radiology
,
230
,
77
87
.
Warrington
,
E. K.
, &
Shallice
,
T.
(
1984
).
Category specific semantic impairments
.
Brain
,
107
,
829
854
.
Yildirim
,
I.
, &
Jacobs
,
R. A.
(
2013
).
Transfer of object category knowledge across visual and haptic modalities: Experimental and computational studies
.
Cognition
,
126
,
135
148
.
Zevin
,
J. D.
, &
Seidenberg
,
M. S.
(
2006
).
Simulating consistency effects and individual differences in nonword naming: A comparison of current models
.
Journal of Memory and Language
,
54
,
145
160
.
Zhong
,
Y.-M.
, &
Rockland
,
K. S.
(
2003
).
Inferior parietal lobule projections to anterior inferotemporal cortex (area TE) in macaque monkey
.
Cerebral Cortex
,
13
,
527
540
.