- Tram Ho
Artificial intelligence (AI) is evolving, and this sentence is literally correct. Researchers create software that embraces the concept of Darwin's evolutionary theory, such as "survival of the fittest" – that is, only the most powerful individuals will emerge in the selection process. fierce nature; This AI program can be improved over many generations without the need for human input. With its unique "genetic" ability, it can replicate the efforts of decades of AI research in just a few hours.
The creators of the groundbreaking artificial intelligence program believe that one day, this system will give us a completely new approach to AI.
“ While the majority of researchers take small steps, they [refer to the researchers in the new project] make a big step into the unknown. This is one of the scientific reports that can create many new aspects for research , ”said Risto Miikkulainen, professor of computer science, who did not contribute to the new research.
Writing an artificial intelligence algorithm is not an overnight affair. Just look at neural networks, machine-learning networks often found in translation research or self-driving cars: these artificial neural networks are a low-level copy of the human brain structure, learning data by changing connections between artificial neurons. Researchers can take months to reconnect neurons, allowing the artificial intelligence system to identify objects.
In recent years, the process of teaching machine learning has been less challenging as researchers have been able to automate a number of steps. But these programs are still based on the coupling of existing circuits, that is, the output still depends on the imagination of the computer engineer and the industry-specific habits that persist.
Therefore, Dr. Le Viet Quoc, a computer scientist working at Google and many other colleagues, has developed a program called AutoML-Zero with the ability to develop AI programs without using intervention. Human cards, making use of mathematical concepts that high school students can understand.
" Our highest goal is to develop a high-end machine learning concept that researchers today still don't understand, " said researcher Le Viet Quoc.
Dr. Le Viet Quoc.
The program detects the algorithm itself by estimating how it will "evolve" during the research process. First, it generates 100 potential algorithms by randomly combining operations. Next, AutoML-Zero tries to apply the newly created algorithm for simple tasks, such as image recognition.
After each round of such testing, the program compares the results obtained with the available algorithms carefully created by researchers. AutoML-Zero selects the quality results closest to the sample algorithm to proceed with the mutation: randomly place, edit, or delete some code to create a different version of the best algorithms. . Mutant algorithms are included in the list of highly efficient lines of code, the program eliminates the old algorithms and the cycle continues.
Every second, the system generates tens of thousands of algorithms to find the most worthy candidate to hold. Researchers also use "tips" to speed up the process of finding algorithms, such as moving algorithms back and forth between groups to avoid evolution going to dead ends and automatically eliminating repetitive algorithms.
In a research report (not yet censored) published in arXiv , the team talked about how in the evolution of the algorithm, the program created even classical machine learning techniques, including Building neural networks. Dr. Le Viet Quoc said that although this solution is simple compared to other advanced algorithms, he is optimistic that when increasing the system size, the program will produce more complex AI systems. much more.
Another computer scientist, Joaquin Vanschoren, working at Eindhoven University of Technology, thinks AutoML-Zero needs more time to perfect himself before he has the same potential as powerful artificial intelligence systems. first in the world. One of the possible improvements is not to let the system start from scratch, but to add the AI training techniques we have discovered.
That is the future plan of Dr. Le Viet Quoc. He also added that instead of looking at the whole picture of the algorithm, focusing on the details also contains many promises. In early April, Dr. Quoc and his colleagues published a new study , suggesting a design modification of an element that appears to be common in neural networks.
Dr. Le Viet Quoc believes that by increasing the number of operations in the data library, increasing computational resources for AutoML-Zero will allow the program to find new possibilities of artificial intelligence.
" We are extremely excited about this direction, in order to discover something very basic, but it will still take a very long time for us to figure it out ," said Dr. Le Viet Quoc.
According to ScienceMag
Source : Genk