Hancock, N.H. - If successful, a Darpa initiative to develop technology for a "perceptive assistant that learns," or PAL, could kick off a new phase for artificial intelligence, enabling devices that would peruse large databases and assemble their own knowledge bases to assist people in decision-making.
As the 22 labs that have received initial funding from the Defense Advanced Research Projects Agency work out the thorny artificial intelligence (AI) issues to realize the agency's vision, a critical piece of the puzzle may already be in place, in the form of a patent granted last month to author and inventor John E. LaMuth for an "ethical" AI system. LaMuth said he has approached Darpa's Information Processing Technology Office about his expert system, but its proprietary nature has been a stumbling block.
The inventor believes his system addresses a crucial facet of any human-oriented automated personal assistant: an understanding of human motivation and ethics. "This AI patent allows for information processing in an emotive, motivational specialization. As such, it represents a quantum leap in the AI field, allowing for abstract reasoning and long-term planning in matters of a motivational nature," said LaMuth, who believes the personal assistant envisioned in the Darpa initiative (see www.eet.com/at/news/OEG20030717S0040) would be an ideal first application for his expert system.
LaMuth said he has been talking to Ron Brachman, director of Darpa's Information Processing Technology Office, about the invention and asserted that Brachman has no problem with the expert system itself but is concerned because the technology is proprietary. "I fear that this shortsighted attitude could prove detrimental to America's current preeminence in the field," LaMuth said. "I feel strongly that the newly issued patent eventually will prove its merit."
The inventor's system tackles some of the less-defined areas of mental ability. Generally, expert systems take some well-defined area of expertise and implement rote rule-execution algorithms. Because emotion, ethics and motivation are relatively esoteric concepts that defy hard definitions, capturing them in a digital system that represents discrete rules and procedures is a challenge.
Rather than bytes as the basic unit of data, LaMuth uses the sentence. "This AI entity is readily able to learn through experience [by] employing verbal conversation or a controller interface," he said. "Technically it would not be [reflectively] aware of its own existence, as this is a strictly subjective determination. It would, however, be able to simulate this feature through language, thereby convincing others of the fact."
The system is based on affective language analysis, a branch of linguistics in which language is characterized in terms of goals, preferences and emotions. LaMuth has automated this aspect of linguistics using conventional ethical categories drawn from Western religion, philosophy and ancient Greek thought.
After working out a basic set of ethical categories, LaMuth created a hierarchy of definitions based on the human cognitive ability to construct emotional and motivational models of someone else's state of mind. For example, a customer talking to a salesperson at a car dealership will be aware of his or her own motivation and expectations but will also construct a model of the salesperson's motivation and ethical values to evaluate information presented about a car. The same two people might meet in different circumstances-say, at a party -and the ethical/motivational models they construct then would be different. But the process would be the same in both cases.
Part of the ethical model building involves successive levels of "indirection." In the car salesperson example, the customer might also construct a cognitive model of how the salesperson is thinking about the customer's own motivation and expectations. The human mind can only take this process a few steps, but logically it can be extended indefinitely. LaMuth's expert system uses a 10-level hierarchy, resulting in 32 pages of "schematic definitions" in the patent application.
David Patterson, known for his pioneering research that led to RAID, clusters and more, is part of a team at UC Berkeley that recently made its RISC-V processor architecture an open source hardware offering. We talk with Patterson and one of his colleagues behind the effort about the opportunities they see, what new kinds of designs they hope to enable and what it means for today’s commercial processor giants such as Intel, ARM and Imagination Technologies.