How Google improves itself as a “machine learning first” company

Traditional AI language comprehension methods depend on embedding the rules of the language into a system, but in this project, as with all modern machine learning, the system is provided with enough data. to self-study as a child. “I didn’t learn to speak from a linguist, I learned to speak while listening to others,” Corrado said. But what makes Smart Replies really possible is that success is easily identifiable – ideas aren’t created. a virtual Scarlett Johansson people who will engage in flirting chats, but respond reasonably to real-life emails. “What success looks like is a candidate feedback machine that people find useful enough to use as their real feedback,” he said. Therefore, the system can be trained by noting whether the user actually clicked on the suggested responses.

However, when the team started experimenting with Smart Reply, users noticed something odd: it often suggested inappropriate romantic responses. Corrado says: “One of the types of failures is this really crazy tendency to make it say, ‘I love you’ whenever it’s confused. “It’s not a software bug – it’s a bug in what we asked it to do.” The program somehow learned a subtle aspect of human behavior: “If you are cornered, saying, ‘I love you’ is a good defensive strategy.” Corrado was able to help the team reduce its enthusiasm.

Repl smarty, released last November, was a hit – Gmail Inbox users now often get a choice out of three potential replies to emails they can flip with. one touch. Often they look out of the ordinary on the label. Of the responses sent by mobile Inbox users, 1/10 was generated by machine learning. “I’m still very surprised it works,” said Corrado with a laugh.

Smart Reply is just one data point in a dense graph of the ML cases that have proven effective at Google. But perhaps the final turning point came when machine learning became an integral part of search, Google’s flagship product, and the font for almost all of its revenue. Search has always been based on artificial intelligence to some degree. But for many years, the company’s most sacred algorithms, the distribution algorithms that used to be called the “ten green links” in response to a search query, were deemed too important for the algorithms. learning maths of ML. “Because search is a very important part of the company, rankings are very evolving, and there’s a lot of skepticism that you can give,” Giannandrea said.

In part, this is cultural antagonism – a stubborn miniature model of the common challenge of getting master hackers to control the Zen-ish method of machine learning. Longtime seeker Amit Singhal, a self-proclaimed student of Gerald Salton, a legendary computer scientist with pioneering work in document retrieval inspired Singhal to help modify code learning. Brin’s and Page’s class into something that can scale in the modern web age. (This puts him in the school of “seekers”.) He teased the astounding results of those 20th century methods and doubted about integrating learners into complex systems. Impurity is the lifeblood of Google. David Pablo Cohn said: “My first two years at Google, I liked search quality, trying to use machine learning to improve rankings. “It turns out Amit’s intuition is the best in the world, and we did better by trying to encode anything in Amit’s brain. We couldn’t find anything as good as his approach. ”

In early 2014, Google’s machine learning gurus believed that would change. “We had a bunch of discussions with the rating team,” said Dean. “We said we should at least try this and see, if there’s any benefit.” The test his team thought of turned out to be central to search: how well a document in the rankings fits into the query (measured by how users clicked it). “We just said, try to compute this subscore from the neural network and see if that is a useful score.”

Leave a Comment