Why artificial intelligence is taught to rewrite your code?

Recently, one company , enabling the machine to effectively learn on examples in a small amount and hone your knowledge as new examples. It can be used everywhere, for example, to teach the smartphone to recognize user preferences or to help Autonomous propulsion systems to quickly identify obstacles.

The Old adage "repetition is the mother of learning" well applicable to machines. Many modern artificial intelligence systems operating in devices that rely on repetition in the learning process. Algorithms for deep learning allow AI devices to extract knowledge from data sets and then to apply what they have learned in specific situations. For example, if you feed the AI system is evidence that the sky is usually blue, and later she will begin to learn the sky among the images.

Using this method it is possible to carry out complex work, but it is, of course, leaves much to be desired. But it would be possible to obtain the same results if you omit system, deep learning AI using fewer examples? Gamalon Boston startup has developed a new technology to try to answer this question, and this week introduced two products that use the new approach.

Gamalon uses the technique of Bayesian programming, program synthesis. It is based on 18th century mathematics developed by mathematician Thomas Bayes. Bayesian probability is used to Refine predictions about the world using experience. This form of probabilistic programming is when code uses likely, but not a specific value requires fewer examples to draw a conclusion, for example, that the sky is blue with patches of white clouds. The program also clarifies your knowledge as further learning examples, and the code can be rewritten to correct the probability.

Probabilistic programming

While this new approach to programming still has its problems that need to be addressed, it has significant potential to automate the development of machine learning algorithms. "Probabilistic programming easier for machine learning researchers and practitioners," explains Brendan lake, researcher at new York University, who worked on probabilistic programming techniques in 2015. "He has the ability to take care of the complicated parts of programming."

Director General CEO and co-founder Ben Vigoda showed MIT Technology Review the sample application for drawing, which uses their new method. It is similar to what Google released last year, that predicts what the person is trying to draw. Read more . But unlike Google, which relies on the sketches has already been seen earlier, the application Gamalon relies on probabilistic programming in an attempt to identify the key features of the object. Thus, even if you draw a shape that is different from those in the application database before it can identify specific traits — for example, a square with a triangle at the top (house) — it will make correct predictions.

Two presented Gamalon product show that their methods can find commercial application in the near time. Product Gamalon Structure uses a Bayesian synthesis software to detect concepts from plain text and already knocking on the effectiveness of other programs. For example, after receiving the description of the TV from the manufacturer, it can identify the brand, product name, screen resolution, size and other features. Another application — Gamalon Match — distributes products and prices in store inventory. In both cases, the system quickly learns to recognize variations of acronyms or abbreviations.

Vigoda noted that there are other possible applications. For example, if you equip Bayesian machine learning model smartphones or laptops, they will not have to share personal data with large companies to determine the interests of users; the calculations will effectively carry out inside the device. Autonomous machines can learn to adapt to the environment much faster using this method of learning.

If you teach artificial intelligence to learn on their own, he doesn't have to be on a leash.