Seqsee: a concept-centered architecture for sequence perception
Douglas R. Hofstadter, Abhijit Mahabal · 2010
One of the goals of this project is to design and implement a computer program that can extend integer sequences intelligently, and the project has resulted in the creation of the program named (pronounced sexy). Seqsee can extend a wide range of cognitively interesting sequences, including the following sequence (Seqsee is presented the sequence without the groupings indicated by the parentheses): ( (1) ) ( (1) (1 2) ) ( (1) (1 2) (1 2 3) ) If people are shown this sequence (without the parentheses), they quickly form a group consisting of the three initial 1s, but then realize that each plays a slightly different role in the sequence. Like people, Seqsee is initially distracted by the three consecutive 1s, but gradually figures out that the second 1 is ascending group, and that the initial 1 is ascending group made up of one ascending group. Architecturally, Seqsee is a descendant of Hofstadter & Mitchell's computer program Copycat, and adds several novel features that allow it to easily modify behavior in response to its recent perceptions, to form specific expectations such as an ascending group is likely to be located here, to more quickly understand sequences having previously seen similar sequences, to see entity as something else, and to do all this without the use of brute force. Seqsee uses several ideas in achieving its goals: William James' notions of the and the stream of thought; analogies between objects; categorization and labeling of objects and of situations, and the detection of categories without using brute-force tests for all sorts of categories; the notion of context which influences and is influenced by perception; the notion, similar to affordances, of the action fringe of object; and a category-based long-term memory. The dissertation describes the program and its principles, which are much more general than integer-sequence extrapolation, and compares its performance with human performance.