Jeff Hawkins at Strange Loop
Jeff Hawkins recently gave a keynote address at Strange Loop 2012, a multi-disciplinary conference on diverse topics like emerging languages, alternative databases, concurrency and distributed systems.
This talk was similar to the keynote speech Jeff gave at the recent ISCA Conference, providing a good overview of the neuroscience and technology behind Grok (e.g.,sparse distributed representations and sequence memory). In this version, Jeff shows an example of how Grok adapts to unexpected change when predicting energy consumption in a building. At first, Grok learns a weekly pattern: energy usage climbs during the day, but drops at night and on weekends, when the building is closed. One week, however, Thursday and Friday are holidays, so the building is closed. Grok predicts that energy will rise on Thursday morning, only to see that its predictions are inaccurate, because energy consumption remains flat in the empty building. Instead of blindly forging ahead with the rest of the regular Thursday pattern, however, Grok adapts automatically and shifts its forecast to predict flat levels. Grok was not trained with data for holidays, but it recognizes that the pattern has changed and adapts accordingly.
Jeff also spends a little more time talking about the future of the technology, including the somewhat contrarian opinion that “The goal here is not to build …human-like things, it’s not to pass the Turing Test. I have no interest in that whatsoever. I want to build machines that are useful for humans.” The real potential of the technology lies in machines that can think millions of times faster than humans. These machines could have memory systems much larger than the human brain, with creative sensor arrays detecting patterns outside the realm of what humans can perceive intuitively. But they wouldn’t get tired and would have neither desires nor sentience. These will be tools that enhance knowledge discovery; they could be the machines that we can send out to explore the universe.
You can watch the video here.