From MIT, the machine that provides human behavior

WRITE, dream, make calculations and play chess better than us, but when it comes to creativity and intuition we should continue to exceed even computer for a while '. Or at least, it is what we like to think. The latest invention of the researchers of the Computer Science and Artificial Intelligence Laboratory at MIT, however, could make us think again: it is a program that can guess how will a human being, with comparable (or better) than those of a person of flesh and bones.

The device, presented at the recent IEEE Data Science and Advanced Analytics Conference in Paris, is a Data Science Machine: an automatic system able to develop predictive models starting from large databases of raw data, as can be for example the information gathered in the server Google or Facebook.

As all the mechanisms of date science (ie scientific analysis of large data sets), the device is designed to extract information, identify patterns and relationships, and extrapolate predictions. Normally, to support a car in this activity is a human operator, who must choose the relevant elements, called variables, including those stored in the database (such as age, gender, or education of users of a website), intuit such correlations could hide in (habits, type of purchases, preferences, political or cultural) and then ask the computer to perform calculations.

MIT researchers have developed a technique, however, called Deep Feature Synthesis, which allows the machine to identify itself relevant variables within a database, and understand how to use them. To test the effectiveness of their device, MIT researchers have registered their car in three competitions to date science, in which the computer has competed against 906 teams of researchers in the flesh.

To get an idea, in one of the races, for example, participants were asked to predict which of the students of online learning program would abandon the course within 10 days from the beginning, using the information registered in the site server about interactions users with the content available. The MIT device has solved the problem, realizing that, to find out who he would retire from the course, just look at two variables: how long before the students went to work the exercises assigned by teachers, and the average time spent on site course. Thanks to this "intuition", the machine is able to overcome the 86% of the participants (human) in the competition. And again, the algorithm was able to predict whether a proposed crowdfounding would be considered as "exciting" by the public; and recognize what buyers of a online shopping site would become regulars. Not bad, in fact, to be the first attempts.

In total the device has defeated 615 teams entered the competition. It is not yet able to win consistently against human beings, but, say its inventors, offers similar results and in much less time: an expert on science working for months to its algorithms, while the machine MIT took only 12 hours to process those used in competitions.

For this, the researchers imagine that soon could find use in this field of study. "We think our Data Science Machine as a natural complement of human intelligence," says Max Kanter, one of the authors of the device. "There are a lot of data out there that could be analyzed, and are not used only for lack of time. But now we have a solution that would give us at least a clue, something from which to start with."