Google Brain researchers are looking for ways to soon put forth a software that will be creating machine learning software as well.
There's a relevant reason why researchers are getting their heads on this. A lot of money is still needed to hire experts who can work with it. What's more is that building an AI still requires a significant amount of time and effort to develop AIs using machine-learning. Relieving some of that stressful work to other machine learning systems could greatly cut the human input that needs to be dedicated to the entire process.
According to Google Brain's research group leader, Jeff Dean, "automated machine learning" is one of the most promising research avenues the team was exploring.
"Currently the way you solve problems is you have expertise and data and computation," Dean said during the AI Frontiers conference in Santa Clara Convention Center, California last January 11 to 12. The conference which was hosted by Impact Deep LLC, is a training and education organization dedicated to AI and deep learning. He follows that up with a revolutionary question: "Can we eliminate the need for a lot of machine-learning expertise?"
MIT researcher, Otkrist Gupta, says that easing the burden from data scientists would result into an increase in productivity, better models, and allow a lot more exploration of ideas.
Professor Yoshua Bengio from the University of Montreal previously explored the idea sometime in the 90s. Apparently, the concept of having a software that learns to learn has already been circulated. It's just that experiments done could not really stand against what humans were capable of doing. Currently, AI is advancing with what is called deep learning. Deep learning software learns to recognize patterns in digital representations of sounds, images, and similar data.
Bengio says, however, extreme computational power is still needed. So, it's still impractical to think of relieving work from machine-learning experts.