First, a model of memory for multiplication facts is described. A system is built to capture the response time and slips of adults recalling two digit multiplication facts. The phenomenon is thought of as spreading activation between problem nodes (such as "7" and "8") and product nodes ("56"). The model is a multilayer perceptron trained with backpropagation, and McClelland's cascade equations are used to simulate the spread of activation. The resulting reaction times and errors are comparable to those reported for adults. An analysis of the system, together with variations in the experiments, suggest that problem frequency and the "coarseness" of the input encoding have a strong effect on the phenomena. Preliminary results from damaging the network are compared to the arithmetic abilities of brain-damaged subjects.
The second model is of children's errors in multicolumn multiplication. Here the aim is not to produce a detailed fit to the empirical observations of errors, but to demonstrate how a connectionist system can model the behaviour, and what advantages this brings. Previous production system models are based on an impasse-repair process: when an child encounters a problem an impasse is said to have occurred, which is then repaired with general-purpose heuristics. The style of the connectionist model moves away from this. A simple recurrent network is trained with backpropagation through time to activate procedures which manipulate a multiplication problem. Training progresses through a curriculum of problems, and the system is tested on unseen problems. Errors can occur during testing, and these are compared to children's errors. The system is analysed in terms of hidden unit activation trajectories, and the errors are characterized as "capture errors". That is, during processing the system may be attracted into a region of state space that produces an incorrect response but corresponds to a similar arithmetic subprocedure. The result is a graded state machine---a system with some of the properties of finite state machines, but with the additional flexibility of connectionist networks. The analysis shows that connectionist representations can be structured in ways that are useful for modelling procedural skills such as arithmetic. It is suggested that one of the strengths of the model is its emphasis on development, rather than on "snap-shot" accounts. Notions such as "impasse" and "repair" are discussed from a connectionist perspective.
[table of contents]