An Ai Capable Of Rewriting Its Own Code Is A Reality

The modern AI philosophy puts learning on the first place. Machine and deep learning are the two most popular catchphrases in today’s field of AI research. By repeating, in the form of feeding AIs huge amounts of data and them letting them learn on examples and making them capable of doing different things, we can today make AIs capable of recognizing different objects, predicting user’s preferences, coming with solutions for different problems, and even recognizing faces, one of the hardest things an AI can accomplish.

But for making an AI to “learn” something we must feed it with immense amounts of data, and the question of whether it is possible of making an AI to learn something with fewer data arises. Well, one startup from Boston, Gamalon, managed to develop a new technology that teaches AI to new things by feeding it with less data than usual deep-learning techniques do.

The company uses the technique called Bayesian program synthesis. It is based on workings of Thomas Bayes, an 18th-century mathematician. Bayesian probability is used all over the world, and we use similar techniques to make our own predictions. The probability is based on experience, and the tech Gamalon made is based on probabilistic programming, a type of programming that uses probabilities instead of specific variables.

By using probabilistic programming, Gamalon can teach AI new things by exposing it to fewer examples than it would take by using classic deep learning techniques. An AI system can get better by being exposed to further examples, and it can even improve itself by tweaking its own code in order to adjust probabilities after being exposed to more examples. For instance, after learning that the sky is blue an AI can recognize the sky in most pictures. But, if learning that the sky is blue, but that there are clouds covering it in most cases, an AI can rework its code so that it includes a high chance of clouds appearing on an image depicting the sky.

“Probabilistic programming will make machine learning much easier for researchers and practitioners,” stated Brendan Lake, an NYU researcher who worked on probabilistic programming. “It has the potential to take care of the difficult [programming] parts automatically.”

In a live demo, Gamalon CEO and cofounder Ben Vigoda showed how a drawing app using the company’s new programming method can predict what a person is trying to sketch. This is similar to Google’s app showed last year, but the app made by Google uses previously analyzed sketches for making predictions, while the app made by Gamalon uses probabilistic techniques. For instance, if you draw an object consisting out of two circles and rectangular shapes on top of them, the app will recognize that you are trying to draw a car because it recognized certain features of the drawing and then predicted the chances of those features to appear one next to another.

The new programming technique can make big data obsolete. In the future, smartphones could come up with user’s preferences just by analyzing the data stored on the smartphone instead sending it to huge data farms, and the future smart shops could determine what a certain customer tends to buy after analyzing data from just a couple of recent visits to the shop. And that’s just the start since further refining the Bayesian program synthesis could make  machines smarter than they ever were.

Leave a Reply

Your email address will not be published. Required fields are marked *