AI Going to the Dogs for Deep Learning Research

Canine Videos Teach AI about Dog Behavior

After trying for years to design an artificial intelligence patterned after humans, cats and now dogs, scientists have found an approach that could help with controlling robots.

Researchers at the University of Washington and Allen Institute for AI recently trained neural networks to interpret and predict the behavior of canines. Their results show that animals could provide a new source of training data for AI systems, especially those used to control robots.

Kelp, a Malamute that had a Go Pro strapped to his head, helped collect data that contributed to the predictions of dog behavior. A total of 380 short videos showed his dog’s point of view—walking, playing fetch, and going to the park.

A picture of Kelp the Malamute, showing the view from its head-mounted GoPro and the sensor data for its limbs.

Using deep learning, researchers analyzed Kelp’s behavior. Deep learning involves collecting vast amounts of data and sifting through it to identify patterns. In this case, that meant matching the data of Kelp’s limbs movement and the visual data from the GoPro. The resulting neural network was trained t0 predict what a dog would do in certain situations. If someone threw a ball, for example, it learned that the dog would chase it.

The paper’s lead author, Kiana Ehsani, told The Verge that the predictive capacity of the AI system was accurate, but only in short bursts. For instance, if the video shows a set of stairs, then the system can guess that the dog is going to climb them. Beyond that, there were too many variable options to predict the dog’s behavior. “Whether or not the dog will see a toy or an object it wants to chase, who knows,” said Ehsani, a PhD student at the University of Washington.

“Our intuition for this was that dogs are really good at finding where to walk—where they’re allowed to go and where they’re not,” says Ehsani. “This is a very hard task for a computer because it requires a lot of prior knowledge.”

The dog knows whether a surface is too steep to walk on or to avoid it if it’s spiky and uncomfortable. It would be time-consuming to program a robot with so many rules, so by watching Kelp’s behavior, the neural network learned the rules without having to be taught them. In other words, it learned from the dog.

Researchers plan to study other animals and use them to train AI, too, to teach it how to model behavior.

read more at theverge.com