When Artificial Intelligence Learns to Build AI

Some researchers say they are now able to make artificial intelligence software design more AI software.

In one example, reports MIT Technology Review, researchers at Google Brain had software build a machine-learning system. The software did—and it outperformed human-made designs.

“On the CIFAR-10 dataset, our method, starting from scratch, can design a novel network architecture that rivals the best human-invented architecture in terms of test set accuracy,” the report, which was published in Cornell University Library in November, reads.

Researchers at Elon Musk’s OpenAI report similar success, as well as Berkeley, the University of California, and MIT. This could bring down one of the biggest barriers to developing AI, which is the cost—currently only a limited supply of human AI experts are available in economy to build complex machine learning systems.

“Easing the burden on the data scientist is a big payoff,” Otkrist Gupta, a researcher at the MIT Media Lab, told the Technology Review. “It could make you more productive, make you better models, and make you free to explore higher-level ideas.”

Microsoft Acquires Deep Learning Startup Maluuba to Advance Democratization of AI