Home | People | Papers | Projects | Methods | Links | Graduate Programs | Subjects
Neural network model. We constructed a simple 2-layer neural network that uses a Hebbian learning rule to modify the weights of the connections between the input and output layers. Specifically, if two units are both firing (correlated) then their connection is strengthened, if only one unit of a pair is firing (anticorrelated) then their connection is weakened. The input layer represents the visual forms of input characters (letters and digits) using a localist representation (each unit represents a different visual form). We also constructed a network that used a distributed input representation, but because it produced the same results and because the localist representation makes the operation of the model more transparent, I will focus on the localist network for the rest of this discussion.
Initially, the network's output layer does not represent anything (because the connections from the input layer are initially random), but with training it should self-organize to represent letters and digits in segregated areas. Neighboring units in the output layer were connected via excitatory connections and units further away were connected via inhibitory connections, in keeping with previous models of cortical self-organization (von der Malsburg, 1973, 1979). (Other architectures would also be consistent with our explanation, e.g., normalization of output activations as opposed to long-range inhibitory connections. What is critical is that the architecture provide a cooperative mechanism to produce clusters of activity and a competitive mechanism to inhibit multiple clusters. For a review of a variety of such models see G. Goodhill (1992). (For related models see Cottrell & Fort, 1986; Durbin & Mitchison, 1990; Linsker, 1986abc; Miller, Keller, & Stryker, 1989; Ritter, 1990; von der Malsburg, 1973, 1979).
Neuroimaging | Theory | Cognitive Modeling | Behavioral Studies | Discussion | Papers