Most of the artificial intelligence research includes a cycle of training known as machine learning, where AI gets better at a task such as recognizing a cat or mapping a path the more it learns. The same approach is used today, without human interference, to build new AI programs.

Google engineers have worked for several years on an outrageously clever machine learning framework known as AutoML (or AutoML) that is now able to build AI that does better than everything we did.Now, researchers have modified it to integrate Darwinian evolution principles and showed it can create AI systems that keep building on themselves faster than they would if humans did the coding.

The latest framework is called AutoML-Zero, and while it might sound a little confusing, it may contribute to the rapid creation of smarter systems-for example, neural networked designed to replicate the human brain more accurately with multiple layers and weightings, which has been struggling with by human coders.Using a basic three-step method-setting up, forecasting, and learning-can be thought of from scratch as machine learning.

The program begins with a set of 100 algorithms created by combining basic mathematical operations at random. Then a complex trial-and - error procedure selects the best performers, who are retained for another round of trials-with some modifications. Or put it another way, the neural network mutates as it goes.

When new code is produced, it is tested on AI tasks-such as spotting the difference between a truck 's picture and a dog's picture-and then the best performing algorithms are kept for future iteration. Like the fittest life.