Google is trying harder than anyone else to make machine brains work

Backchannel has a fairly in-depth write up about Google’s scramble to hire or train programmers with experience in “machine learning” (aka artificial intelligence and neural networks). It’s betting hard on the technology’s ability to mine and interpret data in a way that’s not just novel, but also useful to it’s flagship Search app and more.

Deepmind, a British technology start-up that was bought by Google in 2014, and whose motto is “Solve intelligence,” was an early move for them. The company caught Google’s attention after they published an article detailing how one of their machine babies learned to play seven different Atari games well enough to pose a threat to human video game dominance. DeepMind also created AlphaGo, a machine learning program that whooped a “9 dan rank” human Go player in 2015. Additionally, the company is partly responsible for the DeepDream phenomenon of last summer. It was a neural network system that was able to take user-fed images and figure out terrifying ways to fuse them with puppies and slugs.

Google Brain / DeepMind HQ

Per the Backchannel article, Google has been “obsessed” with machine learning this year. Among their efforts so far: Pilfering any and all students of artificial intelligence guru Pedro Domingos of the University of Washington; allowing artificial intelligence engineers out of the mysterious Google X R&D basement and into daylight; posting flyers in the cafeteria bulletin boards of technical colleges across the country and the world: “Do you want to be a machine learning ninja????!” Google also made a powerful neural network building tool, TensorFlow, available to the public in November 2015, ostensibly to help build a generation of amateur programmers already familiar with machine learning, to be hired en-masse once they have degrees.

Machine learning is a new way of programming. From my limited understanding, it seems to be about creating semi-intelligent data golems to do your bidding, often to superhuman effect. Not only can a neural network learn and in a way understand what a shiba inu looks like (and find pictures of it for all your doge memes), it can use that understanding to then go through a pile of, say, a million pictures of dogs and doges of different breeds and then delineate between each different breed. “If it learns one breed,” reads the article, “it can use the same technique to identify the other 9999 using the same technique.”

Birth of a new God???

Machine learning can also, to a limited extent, understand language, which Google has put to use with its SmartReply feature. Per the article: “Traditional AI methods of language understanding depended on embedding rules of language into a system, but in [SmartReply], as with all modern machine learning, the system was fed enough data to learn on its own, just as a child would.”

Personally I like typing my replies and I think I always will, and, from what I read, SmartReply initially had the tendency to suggest “I love you” as a reponse whenever it sensed tension in the conversation. But other machine learning could be important, and, as this article will explain, Google is desperate to be at the forefront.