Much of the work done by artificial intelligence involves an educational process known as machine learning.
The AI gets better at tasks like recognizing something or mapping a route the longer it takes.
Now the same technique is being used to create new AI systems without human intervention.
For many years, Google engineers have been working on an unusually smart machine learning system known as AutoML (or automatic machine learning) that is already capable of creating AI.
Now researchers have made changes to Darwin's concept of evolution and have shown that it is possible to create AI programs that continue to improve faster if humans could code them by hand.
The new system is called AutoML-Zero, and it could lead to the rapid development of more intelligent systems – for example, neural networks designed to more accurately mimic the human brain.
“Today it is possible to automatically detect complete machine learning algorithms simply by using basic math operations as building blocks,” the researchers write in their paper. “We are demonstrating this by introducing a new concept that significantly reduces human influence across the shared search space.”
The original AutoML system was designed to make it easier for applications to use machine learning, and it already includes many automated features, but AutoML-Zero requires little or no human-written code.
Using a simple three-step process – tuning, predicting, and training – it can be thought of as machine learning from scratch.
The system begins with a selection of 100 algorithms made by randomly combining simple mathematical operations. A complex process of trial and error then determines the best ones, which are saved – with some modifications – for the next round of trials. In other words, the neural network is constantly evolving.
When new code is generated, it is tested against AI tasks – for example, detecting the difference between a truck image and a dog image – and the most efficient algorithms are then saved for future iteration. Like the survival of the fittest.
And this is fast too: the researchers believe that up to 10,000 possible algorithms can be loaded per second per processor (the more computer processors available for a task, the faster it can run).
In the end, this should lead to the fact that artificial intelligence systems will become more widely used and accessible to programmers with no experience in AI development.
Work continues to improve AutoML-Zero, in the hope that it will eventually be able to develop algorithms that simple programmers would never have thought of.
“While most people were taking small steps, [the researchers] took a giant leap into the unknown,” Edd Gent told Science, a scientist at the University of Texas at Austin. “This is one of those papers that could start a lot of future research.”
The work has not yet been published in a peer-reviewed journal, but can be viewed at arXiv.org.
Sources: Photo: uscybersecurity.net
