Connectionist Approaches to Reading

Reading is a highly complex task, involving the rapid coordination of visual, phonological, semantic and linguistic processes. Computational models have played a key role in the scientific study of reading. These models allow us to explore the implications of specific hypotheses concerning the representations and processes underlying reading acquisition and performance. A particular form of computational modeling, known as connectionist or neural network modeling, offers the further advantage of being explicit about how such mechanisms might be implemented in the brain. In connectionist models, cognitive processes take the form of cooperative and competitive interactions among large numbers of simple neuron-like processing units. Typically, each unit has a real-valued activity level, roughly analogous to the firing rate of a neuron. Unit interactions are governed by weighted connections that encode the long-term knowledge of the system and are learned gradually through experience. Units are often organized into layers or groups; the activity of some groups of units encode the input to the system; the resulting activity of other groups of units encodes the system’s response to that input. For example, one group might encode the written form (orthography) of a word, another might encode its spoken form (phonology), and a third might encode its meaning (semantics; see Figure 1). The patterns of activity of the remaining groups of units—sometimes termed “hidden” units—constitute learned, internal representations that mediate between inputs and outputs. In this way, the connectionist approach attempts capture the essential computational properties of the vast ensembles of real neuronal elements found in the brain using simulations of smaller networks of more abstract units. By linking neural computation to behavior, the framework enables developmental, cognitive and neurobiological issues to be addressed within a single, integrated formalism. One very important advantage of connectionist models is that they deal explicitly with learning. Though many of these models have focussed predominantly on simulating aspects of adult, rather than childrens reading, many of the models do explicitly consider the process of learning (e.g., Plaut, McClelland, Seidenberg & Patterson, 1996; Seidenberg & McClelland, 1989). In essence, such models instantiate learning as a process as a slow incremental increase in knowledge, represented by increasingly strong and accurate connections between different units (e.g., the letters in printed words and the phonemes in spoken words to which they correspond). Another critical feature of many connectionist systems is that after learning they show the ability to generalize (e.g., to pronounce novel words which they have not been trained on). Finally, and related to this, such systems often show graceful degradation when damaged. Removing units or connections in such systems typically does not result in an all-or-none loss of knowledge; rather,