Analogical Retrieval via Intermediate Features
Speaker: Mark A. Finlayson , CSAIL
Date: May 10 2004
Time: 4:10PM to 5:00PM
Location: Stata 397 (** UNUSUAL LOCATION **)
Host: Lilla Zollei, Medical Vision
Contact: Lilla Zollei, Louis-Philippe Morency, x32986, x34278, {lzollei, lmorency}@csail.mit.edu
Relevant URL: Analogy is widely considered a key process in human learning and reasoning. Within the currently accepted understanding of analogy, a first step is the recall of relevant precedents from memory. Surprisingly, psychological studies of recall have provided strong evidence that most people are hard pressed to retrieve precedents that are analogically related to probe information or useful for further analogical inference; rather, they retrieve information that is similar in surface, or superficial, features. Experiments have emphasized this predominance in retrieval of so-called ``literal-similarity'' over analogical or structural similarity, and models of recall to date have attempted, with a reasonable amount of success, to account for the effect computationally.
In contrast, evidence drawn from other literature is suggestive that certain sorts of people do consistently achieve analogical recall in particular domains: these people are commonly termed ``experts.'' Current models of recall do not account for this novice-expert difference in the predominance of analogical retrieval. In light of this gap, we present a computational model of how analogical retrieval in experts might be achieved, while still maintaining consistency with experiments on novices. Our model is based on a dot-product-like comparison between vectors of varying-complexity features, and was inspired by recent work by Ullman in the use of visual features for classification. Our result is the same as Ullman's, namely that an intermediate average feature complexity provides the most robust recall within category: we demonstrate this by running an implementation of our model on a dataset of descriptions of complex policial scenarios.
See other events that are part of CSAIL Student Seminar - Spring 2004
See other events happening in May 2004
