An Auto-Associative Neural Network for Sparse Representations: Analysis and Application to Models of Recognition and Cued Recall

The authors present the results of their analysis of an auto-associator for use with sparse representations. Their recognition model using it exhibits a list-length effect but no list-strength effect, a dissociation that current models have difficulty producing. Data on the effects of similarity and strengthening that indicate a dissociation between recognition and frequency judgments are also addressed. Receiver operating characteristic curves for the model have slopes between 0.5 and 1.0 and achieve this ratio in a novel way. The model can also predict latencies naturally. The authors' cued-recall model uses an architecture similar to that of the recognition model and where applicable the same parameters. It predicts appropriate amounts of retroactive interference, and analysis reveals an output competition process that relies on distributed representations and has not been proposed before