We can already see examples of how artificial intelligence technologies are capable of manifesting certain features that seem at first glance and are peculiar only to humans. We create humanoid robots, at least very similar to us, some are committed to creating algorithms that can perform something that people are usually able to do - write music, pictures, or engage in learning.
With the development of this area, companies and developers are beginning to look for an opportunity to change the very basis on which artificial intelligence algorithms are being created, and are taken for the study of real intelligence, as well as ways to effectively simulate it in engineering and creating new generation software. One of such companies is IBM, which has set itself an ambitious task to teach AI to behave (it would be more correct to say to work) more like a human brain, and not as a set of programmed algorithms.
Most existing machine learning systems are built around the need to use a huge set of different data. Whether it is a computer designed to find ways to win in a logic game of go, or a system built to detect signs of skin cancer based on digital images - this rule always works. But such a framework for work looks very limited and concise, and of course this is what essentially distinguishes such systems from how the human brain works.
IBM wants to change that. The DeepMind research team has created a synthetic neural network based on rational decision making when working on a particular task.
Rational machines
“By giving the artificial intelligence a multitude of objects and a specific task, we are forcing the network to detect existing correspondences,” comments DeepMind team specialist Timothy Lillikrap on the pages of Science Magazine.
In tests of the network, conducted in June, the system, in the presence of many factors, gave various tasks related to the digital image. For example, such: “Before the blue thing on the image is an object. It has the same shape as that tiny blue thing that is to the right of the gray metal ball? "
In this test, an artificial neural network was able to determine the desired object in 96 percent of cases, while conventional machine learning models were able to cope with the task in 42-77 percent of cases.
Recently, artificial neutron networks continue to improve in understanding human language. Researchers want, besides making intelligent decisions, such systems can demonstrate and maintain attention, as well as preserve memories.
According to Irina Rish, a researcher at IBM, the development of artificial intelligence could be significantly accelerated and expanded through the use of such tactics.
“Improving neural networks remains a matter of engineering, usually requiring a tremendous amount of time, to arrive at the right architecture that works best. In essence, this is a method of human trial and error. It would be great if these networks could create and improve themselves. ”
Some, of course, may be frightened by the idea of AI networks capable of creating and improving themselves, but if we find a competent way to monitor, control and manage this process, this will allow us to go beyond the current limitations. Despite the growing fear of the revolution of robots that will enslave us all, thousands of lives saved in medicine are predicted for the development of the AI field, opening up the opportunity for us to visit and even settle on Mars and much more.
The article is based on materials .
Comments
Post a Comment