Google AutoML

2020: The artificial intelligence learned to evolve

In the middle of April, 2020 Google announced that it works on serious updating of the free-standing language of development of machine learning of AutoML. It is about AutoML Zero technology within which developers develop the system of evolution of artificial intelligence.

AutoML is a tool from Google which automates development process of algorithms of machine learning for different tasks. It is convenient, quite simple in use and completely an open source code on which Google continues to work still.

Google announced that it works on serious updating of the free-standing language of development of machine learning of AutoML

In the current iteration AutoML has several shortcomings as the programmer still needs to create and configure manually several algorithms which will serve as construction blocks for start. The engineer Risto Miikkulainen and the employee Kuok Le (Kuok Le) from Google submitted the AutoML-Zero program which could bypass this problem and develop programs of artificial intelligence practically without participation of the person, using only basic mathematical concepts.

The program creates algorithms, using the certain mechanism approximate to evolutionary process. Everything begins with creation of set from 100 possible algorithms by accidental consolidation of mathematical transactions. Then the program checks them on a simple task, such as image understanding problem which he should solve. Thereby he checks whether such system of artificial intelligence can work and compares performance of algorithms. Parts of different algorithms which work best of all will be connected further in one artificial intelligence which will be the best variation of algorithms.

Our ultimate goal is development of new concepts of machine learning which even researchers cannot find and which will be deprived of their systematic errors of perception — Kuok explained to Le.[1]

2017: The tasks solved by highly paid specialists people try to automate Google

As reported on October 13, 2017^[2], specialists in the field of artificial intelligence of Google company try to automate the tasks solved by highly paid specialists people.

Google developed self-trained in software for creation of the self-trainee of software

Within the AutoML project researchers of Google trained software for machine learning to create the same software for machine learning. In certain cases the software which was more powerful and more effective in comparison with software created by researchers turned out. According to Google, one of recently developed systems could categorize correctly images according to their contents with 82 percent probability. — putting on the image of marks of location of several objects which represents an important task during creation of augmented reality and independent robots — created by software of software coped with more difficult problem for 43%. Besides that the best of created by the person of a system — only for 39%.

Similar results are especially important because the examination necessary for creation of the similar modern systems with artificial intelligence, quite rare — even in Google.

Today these problems are solved by machine learning specialists whom in the world literally only several thousands — the CEO of Google Sundar Pichai reported, briefly having mentioned the AutoML project at the meeting devoted to release of new smartphones and other gadgets. — We want that the solution of these tasks became available to hundreds to thousands of developers.

On the website devoted to researches, Google specifies that all in the company researches employed about 1300 specialists, and only part of them are engaged in artificial intelligence. In total in Google thousands of engineers in the software development area work. According to the financial statement of parent company Google, Alphabet, in it 27,169 people are involved in research and development (data for 2016).

Robotics

Notes