ABSTRACT
To account for the observations made in patient ELM, Dixon et al. (1997; see also Dixon et al., 1998, Dixon & Arguin, 1999) have proposed an alternative model where shape representations are based on collections of discrete features. This model was largely inspired by the ALCOVE model, which was initially proposed as an account of various visual categorisation data (Kruschke, 1992). The system of Dixon et al. (1997) encodes visual shapes through a series of input nodes, each coding a feature value defining the item on a particular shape dimension. Activation from these input nodes is then transferred to a hidden layer that represents exemplars as points in a multidimensional psychological space. This psychological space acts as a longterm memory that has the dual responsibility of coding stored properties about objects on both visual and semantic dimensions. The hidden exemplar layer connects to output units responsible for the production of responses identifying a particular target shape applied on the input units. Two key features largely determine the operation of the model. One is the assumption of a limited pool of attentional resources in the connections between input and hidden units (Nosofsky, 1986). Thus, if a particular condition requires the processing of multiple shape dimensions for correct discriminations among objects (as in conjunction shape sets), the overall attention pool is divided across these dimensions. Less attention is therefore available for each relevant stimulus dimension than if correct performance can be supported by the processing of a single stimulus dimension. In that case, all of the attentional resources can be directed to that dimension and none is allocated to the irrelevant dimensions. The other major feature of the model is that activation within the hidden layer is not an all-or-none matter, but rather is a graded function of the similarity between the exemplars stored in this long-term memory and the stimulus presented on the input layer. Specifically, it was assumed that activation in the hidden layer falls off exponentially as the
similarity between the stimulus values coded at input and in the hidden layer decrease. The rate of this fall-off of activation is a function of a specificity parameter that controls the selectivity of units in the hidden layer. The simulations conducted by Dixon et al. (1997) using this model replicated the effect of shared-shape features shown by ELM in the shape location task as well as the interactive effects of shared-shape features and semantic relatedness in the shape-name task (the latter result was also found in IL [Arguin et al., 1996a] as discussed earlier). These observations were produced by decreasing the selectivity of units in the hidden layer, without affecting the connections between input and hidden exemplar units or the attention weights that modulate their function. Crucial to the production of the above results is the fact that the hidden exemplar layer codes both visual and semantic properties of known objects. Impairment of this level of processing by reducing the selectivity of units therefore renders the network overly sensitive to both visual similarity and semantic proximity. As these two factors affect the same level of processing, they will also interact with one another (Sternberg, 1969, 1998), thus replicating the results of ELM in the shape-name task. However the very feature of Dixon et al.’s (1997) model that appears crucial in simulating the findings from ELM is also problematic. Indeed, the long-term memory store that is assumed to be impaired in ELM codes both visual and semantic knowledge of objects. This predicts an impairment affecting stored visual as well as semantic object properties. This assumption is contradicted by the dissociation exhibited by ELM (as well as by other CSVA patients) between impaired access to stored structural descriptions but intact semantic knowledge.
