Similarity in a distributed memory model (Q1902594)

An approach to similarity has been developed based on correlated random vectors rather than on the doctrine of identical elements. This approach has been used to work out the TODAM predictions for the list-strength effect (LSE) and list-length effect (LLE), two phenomena of considerable importance to global memory models. Although the incorporation of similarity in TODAM seems to result in an even smaller predicted LSE, with selected parameter values essentially no LLE at all is predicted when a scoring window is used. However, the model does give a good quantitative account of the LLE at the level of the individual subject. This analysis is meant to be illustrative, not exhaustive. The main point is that similarity, long known to be a central variable in memory and learning, can be implemented in TODAM (and any other model that uses a distributed representation) and applied to yield closed-form solutions for predicted \(d'\) values even in such complex experimental designs as those used to study the LSE and LLE. It is hoped that this analysis will prove useful in extending the domain of TODAM and other memory models to situations where similarity plays an important role.