Skip to content

Google has created software that creates artificial intelligence algorithms and makes them evolve following Darwin’s model

27 mayo, 2021

Artificial intelligence is evolving, but literally, with ‘survival of the fittest’ included. And it is that Google researchers have been borrowing concepts from Darwinian theory to create AI software capable of creating new generations of artificial intelligence successively more perfected than its predecessors.

Advancing AI takes time: first you have to create neural networks, loosely based on the structure of our brain, and its creators must spend months working on how to interconnect the ‘neurons’ that form them.

Although in recent years some parts of this task have been automated, its generation is still limited by the fact that be based on human-made designs; and that, therefore, reproduce their biases and prejudices. This means that “possibly [se esté] reducing the innovation potential “of AI, according to a paper published by Google Brain, the AI ​​division of the search engine company.


Don’t worry, it’s not Skynet

That’s where AutoML-Zero comes in., an open source program (you have its source code on Github) created by Quoc Le, a Google computer scientist, which is capable of randomly create thousands of new algorithms from scratch.

Moravec's paradox: why artificial intelligence makes the difficult easy (and vice versa)

“Our ultimate goal is to develop novel machine learning concepts that even researchers couldn’t find.” And it is that the steps in which humans stagnate they do not have to affect an algorithm that we have not shaped in the same waybut chance and selection.

Once created, you test them by training them to perform simple tasks (for example, image recognition) and evaluate their performance, moving the best to a new phase of the experiment. After that, the code of the algorithms is modified slightly and in a random way, again, to test them again.

Successful variants are incorporated into the rest of the algorithms so that, upon reaching the third phase, the program can offer its creators a varied ‘catalog’ of improved algorithms. (Although, well thought out, this, in addition to the mutation process inherently linked to Darwin’s natural selection, also has some of the symbiogenesis of Margulis).

Human beings will always have creativity, a machine cannot create myth or reality?

These processes (creation, selection, ‘symbiogenesis’) are carried out in a matter of seconds, which has allowed AutoML-Zero replicate decades of AI research in a matter of days by dint of ‘evolving’ millions of algorithms.

He has even “rediscovered” traditional AI techniques on his own. But now they want to go one step further. “This is an approach that really excites us,” explains Quoc Le: “discovering something really fundamental that would take a long time for humans to discover.”

Via | ScienceMag

Image | Wikipedia