Researchers from the University of Washington and the Allen Institute for AI have trained the neural network to interpret and predict dog behavior. Scientists came to the conclusion that animals can become a new source of data for training artificial intelligence systems, including those used in robots.
To train the neural network, the head of a dog named Kelp was attached to a GoPro camera, and sensors for recording movements were mounted on the legs and body. Scientists used a system similar to that used in Hollywood to capture the movements of completely computer characters. They got 380 short clips, which could be seen where Kelp moved.
After that, scientists used the method of deep learning – the way in which AI allocates templates from the finished data. The algorithm compared the movements of the dog to the visual information that GoPro recorded. As a result, the neural network has learned to predict what the animal will do in different situations. For example, the algorithm knows that if the dog saw the ball, it would run after him.
The head of the research team, Dr. Kiana Ehsani (Kiana Ehsani), explained toThe Verge that the neural network predictions were fairly accurate, but only in some short time. She noted that it is very difficult to foresee random events in the long term.
However, the scientists did not stop at this and decided to find out what else the algorithm could learn about the surrounding world independently – without preliminary instructions. As explained in the study, dogs “demonstrate a clear visual intelligence, recognizing food, obstacles, other people and animals”, but the authors were wondering whether the neural network could acquire this quality.
Researchers came to the conclusion that the neural net was partially received by the “dog’s mind”: they conducted two tests, asking the algorithm to recognize different environments (house, street, staircase, balcony) and surfaces on which it is possible to walk. In both cases, the neural network was able to successfully complete the tasks using only the basic data from Kelp’s dog.
Our intuition was that dogs actually can understand where they are allowed to walk and where not. This is a very difficult task for the computer, because it requires prior knowledge.
To teach the robot to recognize surfaces, it would take a lot of information and rules, but the dog already knows them all, the researchers noted. Therefore, the neural network, having received information about Kelp’s behavior, mastered the rules without additional training.
The way to train neural networks with the help of in-depth training is not in itself new. However, as noted by Ehsani, the experiment of scientists is probably the first time that someone successfully taught the algorithm with the help of a dog. This means that animals can be a useful source of data for training robots, dogs know enough information that would be useful to robots: for example, they can distinguish children from adults, avoid cars and orient themselves on stairs.