The Technologies with The That Google Will Simulate Be You Writing Post

Manage your mail is often heavy and tedious task: many of the messages that we receive need quick and urgent responses, but when their number is very high that process takes time. Or he wore, because Google has announced a new feature that will make that we can from Google Inbox generate automatic responses default to different messages.

Smart Reply that gets underway these days and to be limited for now to the English support is one demonstration of the progress the company is making in automatic learning (or machine learning) and that is combined with the use of natural language that Google servers can generate to simulate something surprising: that the responders seem really our.

The deep neural networks come into action

A detailed article in which Greg Corrado, responsible for this development in particular, explained how it had managed to implement this feature in Google Inbox found yesterday on the Google Research blog. The basis of everything is in machine learning and in particular in deep learning algorithms.

Such systems have already been used in other Google services extensively. It happens for example in the scope of the search by voice of Google, that leverages the deep neural networks (DNNs) as the technology on which these models are based. Improvements on the Gaussian mixture models (GMM) were evident, and allowed improve the speed and accuracy of speech recognition It is notable.

Another of those DNNs application has been made to obtain something as seemingly random as thumbnails of the videos on YouTube. Behind that process is really detailed studies on how to choose the image that will make users Finally click on that video because this miniature draws attention.

Works Smart Reply

The basis of this characteristic is the use of the so-called learning sequence-to-sequence, a system that allows the automatic conversational synthesis, which among other things served as in Google create a nice experiment: a chatbot that debated about the meaning of life with a human, and did remarkably well.

The application of this system of learning for the generation of responses to all kinds of post was a challenge, but they managed to solve the problem with the use of recurrent neural networks. One of them “encodes” the incoming mail, and the other generates possible responses.

Email encoder It works studying the words of an email and generating a vector that allows you to “understand” the machine what is being said and with what tone. Not only that: is able to identify two sentences formulated differently to mean the same thing. Thus, this part of the system knows that “are you free tomorrow?” and “How about if we meet tomorrow?” have the same goal.

The second system, the “decoder”, takes this vector and synthesizes a correct answer Word for Word. To avoid responses of tens and hundreds of words in Google used a variant of a neural network of the type “Long short-term Memory” (LSTM) that allows the system focus on the part of the email that is more useful then predict the response, downplaying to least relevant phrases before and after.

The privacy flag

Throughout this system had a clear potentially controversial component: the privacy. The proper functioning of this system of learning machine or machine learning is based precisely on the training with millions of e-mails.

However, these emails are never read by a human, and as stated Corrado “that means that the researchers had to work learning on a set of data that no could read, it was something as well as treat of” solve a puzzle with the covered eyes”.

Even so the prototype ran after the correction of several nice initial failures. For example, in the early stages of developing a response suggested consistently by the system was “I love you” (“I love you”). The system, said Corrado, “it was just what they had trained le do, generate probable answers, and answers like”thank you”,”I feel good”, or”I love you”are super common, so the system would support them as a safe bet if I wasn’t sure in other cases”.

Solving it normalized the likelihood of a candidate response contrasting them with past answers, something that made the system was less “loving”, but as says Corrado, much more useful. If you want, you can test the results – in English – using Inbox for Android and iOS. We will see if your acquaintances detected that this response is automatic.