Laboratory Notes: Biologically Inspired Artificial Intelligence

Patrick H. Winston
Computer Science and Artificial Intelligence Laboratory

In the CSAIL Genesis Group, we aim to develop a computational theory of human intelligence, both for its own sake and because such a theory would take the engineering side of Artificial Intelligence to another level. Accordingly, we are especially interested in what happened 50,000 years ago when the human species began to become so much more capable than chimpanzees and other primates. In particular, we study our human ability to use words, phrases, and sentences to build descriptions, because that characteristic gives us humans an extraordinary capacity to tell complex stories, enabling us to construct both concrete and abstract analogies, making it possible to benefit from both our own experience and the experiences of others, all of which seems to set us humans apart from other species.

In the CSAIL Genesis Group we study our human ability to use words, phrases, and sentences to build descriptions, because that characteristic gives us humans an extraordinary capacity to tell complex stories, enabling us to construct both concrete and abstract analogies, making it possible to benefit from both our own experience and the experiences of others, all of which seems to set us humans apart from other species.

In the CSAIL Genesis Group we study our human ability to use words, phrases, and sentences to build descriptions, because that characteristic gives us humans an extraordinary capacity to tell complex stories, enabling us to construct both concrete and abstract analogies, making it possible to benefit from both our own experience and the experiences of others, all of which seems to set us humans apart from other species.

The words, phrases, and sentences of our language and symbol systems are important not just because of what they can do by themselves but also because of what they can marshal. Anyone told that “John kissed Mary,” and then asked “Did John touch Mary,” says “yes,” and reports that it must be true because they have imagined the event and read the contact off of the imagined event as if it were real.

Evidently our language faculty works with subsymbolic faculties in tightly coupled, imagination-stimulating loops, so steps toward a computational account of human intelligence cannot focus exclusively on the differentiating symbolic veneer. Rather, the steps must include work aimed at the contributions of a subsymbolic perceptual apparatus and the contributions of the tightly coupled loops that link the symbolic and subsymbolic.

To expose the issues and develop computational theories, we are building the Gauntlet system, a system of specialists, each of which watches sentences and images flow by, looking for particular aspects of what can happen in the physical world and in abstract worlds. For example, one of our dozen or so language-side specialists focuses on physical objects moving along trajectories (the bird flew to a tree); another focuses on change (the speed increased). Experiments using a 50,000 sentence sample from issues of the Wall Street Journal indicate that more than 25% of the sentences contain such a trajectory or change, usually in an abstract world (prices rose sharply; investor confidence increased).

Because the specialists are backed by memories that accumulate experience, the Gauntlet System has just become able to use its own imagination to answer questions such as “what happens when a bird flies to a tree,” “what happened when John kissed Mary,” and “what can happen to prices.” A cross-modal coupling supplies answers via various vision specialists that read the answers off of recalled images and videos contact appears between the bird and a branch; contact appeared between John’s lips and Mary’s; prices can go up and down over time).

Leave a Reply

Replies which add to the content of this article in the EECS Newsletter are welcome. The Department reserves the right to moderate all comments. If you would like to provide any updated information please send an email to newsletter@eecs.mit.edu.