AI programs are becoming Darwinian. They are evolving on their own.

Software programs have now become sophisticated enough to build AI programs that can improve and build upon themselves without human intervention. In testing, the program could leapfrog decades of AI research. Perhaps a seminal achievement, the Google program called AutoML-Zero could possibly discover more about AI than humans themselves. (Science)

Manually building an algorithm

Manually building an AI algo can take many months because of the complex processes involved. For example, building and assembling a machine learning neural network involves an architecture of many smaller subcircuits. The process can take a long time to optimize. Even so, the finished product can still suffer from the designers’ bias and limitations.

Enter: Google’s AutoML-Zero

Computer scientist Quoc Le led a team of researchers at Google (NASDAQ: GOOGL) to develop the rather expressively-named software program called AutoML-Zero.

Perhaps it signifies the automatic creation of machine language programs with zero human effort. The program indeed develops AI programs with effectively zero human input, using only basic mathematical concepts a high school student would know.

But that’s not enough for Quoc Le. “Our ultimate goal is to actually develop novel machine learning concepts that even researchers could not find,” he says.

That rings a bell. Dailyalts wrote about an AI bot, also by Google, that will itself design the next-generation AI chips. “We believe that it is AI itself that will provide the means to shorten the chip design cycle, creating a symbiotic relationship between hardware and AI, with each fueling advances in the other.”

How AutoML-Zero works

AutoML-Zero starts with building a collection of  algorithms using mathematical operators. These are checked against hand-designed algorithms. It makes adjustments and then repeats the cycle. The software creates thousands of these collections at once, which lets it sift through tens of thousands of algorithms a second until it finds the optimal solution.

Though the software is in a preliminary stage it could still find a number of classic machine learning techniques, including neural networks.

However, according to Le the work is a proof of principle and that it can be ultimately scaled up to create more complex AIs.

That could be achieved by adding more mathematical operations to the software’s library and by beefing up the available computing resources.

Perhaps the software could even stumble upon totally new AI concepts.

Related Story:  Artificial Intelligence: An AI Bot Will Design the Next-Gen AI Chips