MIT Goes Back to the Drawing Board on Artificial IntelligenceS

A new research initiative at MIT intends to revisit some decades-old assumptions about artificial intelligence. It's part of a five-year plan to create an AI program that can comprehend a children's book.

The Mind Machine Project, a collaboration between some two dozen professors, researchers, and graduate students at MIT, hopes to fashion a new conceptual framework for AI development. Professor Neil Gershenfeld, the director of MIT's Center for Bits and Atoms, told reporters that the Mind Machine Project will "rewind to 30 years ago and revisit some ideas that had gotten frozen."

A running theme in the project's literature is the need for software that better resembles the human brain, which can store memories, integrate different cognitive functions, and make intuitive leaps in a way that no computer has been able to approximate. People tend to make decisions based on "data sets which are full of ambiguities and inconsistencies," in Gershenfeld's words. Rather than just create a program that can process a great deal of data very fast, the team wants to develop an artificial consciousness capable of nonlinear, human-style reasoning.

The project consists of four separate avenues of study: Mind, Body, Memory, and Brain/Intent. The "Mind" component might be the most interesting, because it seems to be concerned with teaching an artificial brain to interpret the kind of social cues that human beings learn through years of experience:

[Researchers will d]evelop a software model capable of understanding human social contexts — the signpost[s] that establish these contexts, and the behaviors and conventions associated with them.

If all goes well, the Mind Machine Project could have an AI program up and running by 2015. Don't expect anything incredibly sophisticated, at least at first — researchers aren't even all that concerned with passing a Turing test, wherein a person chats with the program without being told whether it's a machine or a human.

Instead, the five-year goal is to create a program that can read a children's book, explain the story, and ask questions about it or otherwise demonstrate a basic level of comprehension. This might seem modest compared with, say, Skynet, but it would be a major step toward what the project hopes eventually to accomplish: AI programs that can act as helper-entities for people with Alzheimer's disease, or other conditions that compromise the higher functions of the brain.

This kind of software, which the researchers describe as a "brain co-processor," would store and retrieve information for people with memory problems. The project team hasn't ruled out eventually making such programs available to anyone who wants them. This means that for those of us who aren't as organized as we'd like to be, or maybe had some trouble with the Amelia Bedelia books, help is potentially on the way.