TITLE: Pleasantly Surprising Interconnections AUTHOR: Eugene Wallingford DATE: June 11, 2006 2:05 PM DESC: ----- BODY: The most recent issue of the Ballast Quarterly Review, on which I've commented before, came out a month or so. I had set it aside for the right time to read and only came back around to it yesterday. Once again, I am pleasantly surprised by the interconnectedness of the world. In this issue, editor Roy Behrens reviews John Willats's book Making Sense Of Children's Drawings. (The review is available on-line at Leonardo On-Line.) Some researchers have claimed that children draw what they know and that adults draw what they see, and that what we adults think we see interferes with our ability to create authentic art. Willats presents evidence that young children draw what they see, too, but that at that stage of neural development they see in an object-centered manner, not a viewer-centered manner. It is this subjectivity of perspective that accounts for the freedom children have in creating, not their bypassing of vision. The surprising connection for came in the form of David Marr. A vision researcher at MIT, Marr had proposed the notion that we "see by processing phenomena in two very distinct ways", which he termed viewer-centered object-centered. Our visual system gathers data in a viewer-centered way and then computes from that data more objective descriptions from which we can reason. Where's the connection to computer science and my experience? Marr also wrote one of the seminal papers in my development as an artificial intelligence researcher, his "Artificial Intelligence: A Personal View". You can find this paper as Chapter 4 in John Haugeland's well-known collection Mind Design and on-line as a (PDF) at Elsevier. In this paper, Marr suggested that the human brain may permit "no general theories except ones so unspecific as to have only descriptive and not predictive powers". This is, of course, not a pleasant prospect for a scientist who wishes to understand the mind, as it limits the advance of science as a method. To the extent that the human mind is our best existence proof of intelligence, such a limitation would also impinge on the field of artificial intelligence. I was greatly influenced by Marr's response to this possibility. He argued strongly that we should not settle for incomplete theories at the implementation level of intelligence, such as neural network theory, and should instead strive to develop theories that operate at the computational and algorithmic levels. A theory at the computational level captures the insight into the nature of the information processing problem being addressed, and a theory at the algorithmic level captures insight into the different forms that solutions to this information processing problem can take. Marr's argument served as an inspiration for the work of the knowledge-based systems lab in which I did my graduate work, founded on the earlier work on the generic task model of Chandrasekaran. Though I don't do research in that area any more, Marr's ideas still guide how I think about problems, solutions, and implementations. What a refreshing reminder of Marr to encounter in light reading over the weekend. Behrens was likely motivated to review Willats's book for the potential effect that his theories might have on the "day-to-day practice of teaching art". As you might guess, I am now left to wonder what the implications might be for teaching children and adults to write programs. Direct visual perception has less to do with the programs an adult writes, given the cultural context and levels of abstraction that our minds impose on problems, but children may be able to connect more closely with the programs they write if we place them in environments that get out of the way of their object-centered view of the world. -----