In late October, we wrote about the Neural Turing Machine, a Google computer so smart it can program itself. In the time since, it’s become clear that this is only the beginning and we should expect a lot more from DeepMind Technologies, the little-known startup acquired by Google who developed the human-like computer and sports the mission “Solve intelligence.”
In discussing DeepMind Technologies’s delve into the future of computers with MIT, founder Demis Hassabis detailed the company’s research and mentioned that he wants to create “AI scientists.”
He explained that although they’re currently working on some smaller AI activities like searching for ways to apply DeepMind techniques to existing Google products such as Search and YouTube recommendations, his plans for the future are bigger than a better search engine. He dreams of creating artificially intelligent “scientists” that could develop and test their own hypotheses in the lab. He mentioned that there’s also a future for DeepMind’s software in robotics.
“One reason we don’t have more robots doing more helpful things is that they’re usually preprogrammed,” he told MIT. “They’re very bad at dealing with the unexpected or learning new things.”
DeepMind developed software that learns by taking actions and receiving feedback on their effects by combining “deep learning” with “reinforcement learning”—which researches have been tinkering with but failing to use as well as DeepMind for decades. DeepMind Software has learned to play the classic Atari games Pong, Breakout and Enduro better than a human expert without ever being programmed with information on how to play.
The Neural Turing Machine we wrote about in October also completed simple learning tasks. In tests where the computer was asked to learn to copy blocks of binary data and learn to remember and sort lists of data, it was found that the computer learned faster and produced longer blocks of data with fewer errors than more basic neural networks.
If DeepMind combines this type of learning technology with robotics, “AI Scientists” could be a very real future.