A Retrieval Model Using Feature Selection

Abstract

Most case-based memory models assume retrieval is an automatic process where an input is matched to the contents of memory and an episode is activated based on similarity; however, there are many instances where overall similarity alone is not sufficient to identify a single prior instance that corresponds to an input instance. Instead, goal context appears to affect the process of retrieving cases from memory: the same information, in the presence of different processing goals such as explanation or planning, will result in retrieval of different prior cases from memory. The evidence for this claim consists of a series of experiments demonstrating that activation of cases does not occur automatically, but is dependent on strategic processing goals. (Seifert, McKoon, Abelson, and Ratcliff, 1985). A hybrid model of retrieval is proposed that incorporates the content-addressable character of distributed memory models with goals as a controlling mechanism to focus attention on features that will produce relevant remindings. A weighting scheme is designed to help the network override features that prevent a match by focusing on features that are related to the goals of the model. This general notion of feature selection appears useful in incorporating expectations about the usefulness of features with the general learning approach in distributed models of learning where the input features are given equal weight as the model learns various associations within the task domain. In this paper, I propose that such systems may be able to speed the learning process by using a scheme to weight features before learning, based on prior expectations of the relevance of particular features within the task domain.