Foxconn may be looking to replace human workers with robots but they’ll at least need some flesh & blood nursemaiding; if the Hasegawa Research Group at the Tokyo Institute of Technology has its way, future ‘bots won’t even require that. The team has developed a so-called self-replicating neural network that recreates the educated guesses that the human mind is capable of. Called SOINN (Self-Organizing Incremental Neural Network), it means the robot can learn from previous experience, understand its own limitations and figure out possible solutions.
For instance, if asked to prepare a glass of cold water for someone, even if it has never been instructed on how to do that before, the robot can pick up a glass, fill it with water and then – recognizing that it only has two hands and can’t then pick up ice as well – put down the glass so as to drop the ice in. While it may sound basic, that sort of problem-solving has generally been beyond currently implemented robotics.
“Thinking about artificial intelligence in the real world, actual environments are inevitably more complex, and they change quickly. So it’s necessary to have a learning mechanism that adapts to the situation. Also, because new situations emerge, it’s also necessary to have the ability to keep learning new information on the spot” Hasegawa Group
Just as we can consult experts and other sources, like the internet, the SOINN system can educate itself using web-based information or skills already learned by other robots, in a similar way to how the open-source robot Qbo can crowdsource object-identification so as to recognize things individual units are only seeing for the first time.
“For example, suppose this robot doesn’t know how to make tea, and it’s sent to an elderly person who lives alone. And suppose that person asks it to make a cup of green tea. The robot doesn’t know how, so it asks robots around the world how to make tea. Suppose, for example, that a robot in the UK tells it how to make British-style tea. We think this robot will become able to transfer that knowledge to its immediate situation, and make green tea using a Japanese teapot” Hasegawa Group
According to the team responsible for SOINN, it’s also computationally-light so shouldn’t require a super-powerful PC in order to do all the crunching. No word on when the system might show up in commercial ‘bots, however.