Google machine learning software has learned self-reproduction

In May of this year, we wrote about the project AutoML – Artificial Intelligence (AI) technology from Google, designed specifically to create other AI. Now Google announced that its AutoML was able to outperform AI developers and is able to create software for machine learning, which is more effective and powerful than the best examples of similar systems developed by humans.

AutoML recently set a record in the efficiency and speed of image cataloging, under the specified conditions, showing a result of 82 percent efficiency. Although this task itself was relatively simple for the system, AutoML was also able to outperform automated systems and special augmented reality systems in a more complex task of locating multiple objects in an image. In this test AutoML coped in 43 percent of cases, while human-created systems showed an efficiency level of 39 percent.

The results are impressive, because even in such a giant company as Google, there are only a few people who have enough experience to lead the development of AI systems of this level. Automating this area requires a very wide set of skills, but once the result is achieved, it can completely change the industry, as noted in Google.

“Today, only a few thousand machine learning specialists worldwide can create such software. But we want to make so that hundreds of thousands of other developers, too, could take part in this, “- quotes the words of Google’s CEO Sundar Pichai Wired magazine.

A significant part of the meta-teaching is related to the imitation of the work of neural networks of the human brain, as well as the need to run huge amounts of various data through these networks. Of course, the most difficult task is precisely to simulate the structure of the brain and make it solve more complex problems.

Today, existing neural networks are still easier to upgrade or customize for certain tasks, rather than to develop new ones from scratch. However, research like the one we are talking about suggests that this is only a temporary phenomenon.

As it will be easier for the new AI to create more and more complex systems that are designed to perform tasks that a person is simply not able to perform now, it is very important that the person still remain as a key link without which these systems simply can not function. A truly high-grade AI can easily use prejudiced interpretations in various questions, for example, by stereotyping the parallel between ethical and gender characteristics. However, if engineers devote more time to solving this potential problem now, without leaving everything for later, then in the future it will have less chance of a real occurrence.

In general, Google is trying to perfect AutoML to such a state that developers can use it in real solutions to problems. If they succeed, the effect of using AutoML can affect far beyond the walls of the company itself.

“We want to democratize it,” Wired quoted Pichai as saying.