Humans don’t need to look at thousands of pictures of Golden Retrievers to recognize one when they see it at the dog park. Computers don’t have that ability: Teaching A.I. how to categorize dogs, toys, and other things requires showing them many images depicting those objects. At least it did — now Google’s DeepMind has made a breakthrough that allows A.I. to figure out what an object is after viewing just a single picture of it.
This ability is called one-shot learning. It’s supposed to help A.I. see the world much like we do, and give them the skills needed to know what a dog is without requiring its creator to find millions of images of them. DeepMind still doesn’t learn exactly like we do — those capabilities are still outside of its grasp — but one-shot learning is a stepping stone on the path towards closing the gap between human and machine learning.
DeepMind researcher Oriol Vinyals told Technology Review that the breakthrough was made by adding a memory component to a deep learning system. It then viewed hundreds of categories of images (it still has to learn somehow) and was able to recognize images of dogs after seeing just one example.
Similar research from DeepMind was published in June. In that paper, DeepMind’s researchers explained that one-shot learning could also help A.I. learn the meaning of new words and recognize someone’s handwriting. These abilities would help it learn why people are searching for a specific term, for example, or perhaps verify that an autographed object is legitimate.
This brings A.I. much closer to behaving like humans. DeepMind has been working on that for a while: It’s also made robot voices more realistic and taught A.I. basic reasoning skills by combining neural networks with memory systems.
DeepMind’s A.I. is also famous for learning the notoriously complex game of Go, an achievement which Tesla CEO Elon Musk called a 10-year jump for A.I., in March. It still struggles with strategy games like Magic: The Gathering, but teaching it reasoning skills and how to recognize objects could change that.
It also brings A.I. closer to learning more like a baby, which DeepMind researchers note can “generalize the concept of ‘giraffe’ from a single picture in a book,” than like machines. Whether smarter A.I. is terrifying or exciting, though, depends on your point of view.