Google AutoML is an artificial intelligence system that aims to design and train other artificial intelligence systems to improve them beyond human capabilities. However, although it reads very simple to do, in reality it is a very complex process.
Years ago Google already claimed that their AI systems were better than humans at doing this task, but there was still a lot of work to be done to make this completely true.
The big news then, is that now Google’s engineers claim to have taken an essential step: AutoML has the potential to operate without the bias of human beings involved and following a process similar to Charles Darwin’s evolutionary theory.
A new study published by engineers in Google’s Brain division explains the breakthrough they have made with AutoML. The report suggests that Google’s AI could automatically discover new algorithms that are completely unknown to humans because humans would basically not be involved in the research process.
Currently, AutoML is able to improve existing artificial intelligence systems and train them, automatically looking for patterns to choose the ones that have the best performance. Best of all, it is a tool that anyone can use, but because of this, humans are still involved in the process of refinement and data entry and the ways to influence it throughout the process.
So far, from all this intervention, AutoML chooses the best IAs according to the patterns that the engineers provide.
Now, if you remove humans from this whole process, AI has a “free way” to unleash its full potential and find better solutions than those it can find with human bias, and therefore the limits set by human engineers disappear, and you can go much further. As explained in the published document:
“Man-made components bias search results in favor of man-made algorithms, possibly reducing the innovation potential of AutoML”
And this is exactly what Google researchers claim to be able to change. To do this, an AutoML update uses a new method to find and improve AI without any pre-defined human parameters.
This is how Google AutoML works
As indicated in ScienceMag, AutoML-Zero (the name of this new version) follows a selection process similar to that proposed by Charles Darwin in his Theory of Evolution. Google AutoML-Zero chooses 100 candidate algorithms at random and tests them with relatively simple tasks. The algorithms with the best performance go to a next phase where their code is slightly modified at random and then tested again. At the same time, the variants that have been found to be effective are added to the rest of the algorithms. And so phase after phase until we have a “population” of more varied algorithms and with better capabilities to solve the proposed tasks.
It is a process similar to that followed by living beings, where only the best adapted survive, receiving those best adaptations of mutations that come from previous generations. The difference here is that the mutations that are shown to be useful for survival cannot be transmitted directly to existing organisms, but only to the descendants of future generations.
AutoML-Zero is a proof of concept and still has a lot of work and development ahead of it. We will have to see in the next years how this technology evolves, to see if it can indeed create artificial intelligence by evolution without the slightest human intervention, which makes it an interesting and certainly curious step if we consider how it uses the mutation process to evolve the AIs it generates autonomously.
Without a doubt, Google AutoML is one more step in the race of science to allow machines to control absolutely everything.