SAN FRANCISCO, Calif. Tomorrow's semiconductors will mimic the human brain to solve problems in novel ways. That's the view of Jeff Hawkins whose startup Numenta (Menlo Park, Calif.) is shipping tools to create smart algorithms that learn the same way as the human neocortex.
"The idea is to take the way the neocortex works and apply it to new problems," said Hawkins in a keynote address at the International Solid State Circuits Conference here Monday (Feb. 4). "I sincerely believe this is similar to building the first computers," said Hawkins.
The new algorithms could someday open the door to novel memory designs that can directly act on data structures, delivering answers to problems without the intervention of a traditional logic processor. Initial applications for the tools are more likely to be classical artificial intelligence problems such as machine vision.
Although the term "artificial intelligence" was first coined at a conference at Dartmouth College in 1957, the AI field has made little progress in the past 50 years, Hawkins said. "I think the root cause is the lack of a way to represent knowledge," Hawkins said.
Recent advances in neural science suggest the neocortex uses a distributed hierarchy of memory to learn. The hierarchy stores common sequences at each node and can make spatial predictions based on them.
Last year, Numenta shipped a set of software tools to create in software what it calls hierarchical temporal memories. Although the tools are still at an early stage of development, more than 100 companies have accessed them. Eight partners are using the tools on a variety of projects that range from motion capture for computer games to lane-change predictors for cars and voice ID systems for defense.
"We are making a lot of progress understanding how brains work and building machines based on them. I think this is going to have an enormous impact on semiconductors," Hawkins said.
Nevertheless, the tools are still in what Hawkins characterized as a pre-beta stage. Designing chips based on the new principals could be two to four years away, he added.
"Right now we are trying to do all this in software," Hawkins said. "People are coming to us wanting to build this in hardware, but I am telling them it is too soon," he said.
"We are very confident we know how this memory system works, but we still have a long way to go," Hawkins said.