DeepMind’s baby system, or “neural stack,” is a computer that’s designed to incorporate ideas from neuroscience into machine learning. It learns by building up a series of simple representations of the world around it.
The baby system’s performance on a simple task called the “visual reasoning” test was “unbelievable,” says Demis Hassabis, an artificial intelligence expert at DeepMind and the University of Cambridge. “I’ve never seen anything that learns so quickly.”
The test, developed by Google Brain researcher Tom Mitchell, involves showing a machine a series of images and asking it to identify which one is different from the others. It’s a task that’s easy for humans but difficult for computers.
The baby system got the answer right 99.8 percent of the time, while the best existing AI system got it right 98.5 percent of the time. The difference may not sound like much, but it’s significant, Hassabis says.
The baby system also did better than existing AI systems on a task called “alignment” – identifying whether two images are the same or different. And it was able to learn a simple language, called “WordNet,” with just a few hundred examples.
The baby system is a “proof of concept,” Hassabis says, and it’s not yet clear how well it would scale up to more complex tasks. But the results suggest that it’s possible to build AI systems that learn more like humans do.
The baby system is based on a theory of how the brain works, called “hierarchical temporal memory,” or HTM. The theory was developed by Jeff Hawkins, a computer scientist and co-founder of the startup Numenta.
HTM is a “bottom-up” approach to machine learning, in which the system starts with simple representations of the world and builds up to more complex ones. This is in contrast to the “top-down” approach of most existing AI systems, in which the system starts with a complex representation of the world and then tries to simplify it.
The baby system is “a very exciting and important step” in the development of HTM, Hawkins says. “It shows that you can build a working system that learns like a baby.”
But Hawkins caution
