Google Brings Artificial Intelligence to Your Inbox

November 3, 2015 Updated: November 8, 2015

Google’s Inbox app, a streamlined version of Gmail, already comes with plenty of features meant to expedite the emailing process, filtering out low-priority items and giving you the option to hit “snooze” on a message you want to deal with later.

Smart Reply is a new feature that aims to take Inbox to the next level. Using deep learning algorithms, Smart Reply reads your emails and then automatically generates texts you can pick as your replies.

Smart Reply was built on the foundations of two existing programs—spam detection and email categorization—which were woven together to create an algorithm that could synthesize conversations.

A preview of Smart Reply, Google's new semi-automatic email function (Google Research Blog)
A preview of Smart Reply, Google’s new semi-automatic email function (Google Research Blog)

The replies don’t follow preprogrammed rules, instead they rely on machine learning techniques that have become a trademark for Google. Voice search, YouTube thumbnails, and Google Translate, all deploy neural networks.

“In practice, any engineer’s ability to invent ‘rules’ would be quickly outstripped by the tremendous diversity with which real people communicate,” Greg Corrado, a senior research scientist, wrote on Google’s research blog. “A machine-learned system, by contrast, implicitly captures diverse situations, writing styles, and tones.”

Neural networks differ from conventional programs in that the machine learns to formulate its own rules by drawing from patterns found in large amounts of data. First invented in the 1950s, neural networks have experienced a renaissance as improved computational power has allowed for wider applications, including in diagnostic medicineimage recognition, and self-learning complex tasks, like playing expert-level chess.

Smart Reply uses “end-to-end sequencing learning,” where machines learn how to have a conversation by training on sample dialogue. The same technology has been used by Google to make a chatbot that “learned” to talk by emulating the patterns found in movie scripts.

End-to-end sequencing learning can result in machines that can churn out appropriate dialogue, but have also generated lists of optional responses that lack common sense. The programmers had to make some major tweaks to the prototype after the program would suggest alternate replies that all shared the same semantic meaning, such as “How about tomorrow?” “Wanna get together tomorrow?” “I suggest we meet tomorrow.”

Another bizarre feature of our early prototype was its propensity to respond with “I love you” to seemingly anything. 
— Greg Corrado, Google

In addition to what researchers call the”response diversity problem,” the Smart Reply prototype was also inclined to give answers that were overly friendly, and the probability weights for different types of answers had to be adjusted accordingly.

“Another bizarre feature of our early prototype was its propensity to respond with ‘I love you’ to seemingly anything. As adorable as this sounds, it wasn’t really what we were hoping for,” Corrado wrote.

“Some analysis revealed that the system was doing exactly what we’d trained it to do, generate likely responses—and it turns out that responses like ‘Thanks,’ ‘Sounds good,’ and ‘I love you’ are super common—so the system would lean on them as a safe bet if it was unsure.”

Google will release Smart Reply for Inbox on iOS and Android later this week.