Advances in artificial intelligence have made it possible for software to take on some jobs for humans like proofreading, analysing and even truck driving. Moreover, some leading researchers have recently created a software that can replace them to do one of the most difficult tasks: designing machine learning software.
In an experiment, researchers from Google’s Google Brain group had the software designed a machine learning system and take a test used to evaluate language processing software. Surprisingly, the results surpassed previously published achievements from software designed by humans.
In recent months, other groups of researchers have also reported on the progress of software that learns how to learn. They come from the nonprofit research institute OpenAI (co-founded by Elon Musk), MIT, University of California, Berkeley and another Google’s artificial intelligence team called DeepMind.
If automated AI techniques are put into practice, they can accelerate the speed of machine learning software development across the economy. Currently, companies have to pay more for machine learning experts, a supply shortage when demand has increased gradually. But Jeff Dean, leading the research team Google Brain, stressed that some of the work of this target group could be replaced by software. Describing the concept of “automated machine learning”, Jeff showed that this is one of the most promising studies his team has been working on. “The way you solve a problem now is usually having expertise, data and computing” said Jeff Dean at AI Frontiers conference in Santa Clara, California. “Could it be possible that the need of having lots of machine learning expertise can be eliminated?”
In the face of the question, a series of experiments from Google’s DeepMind research team recommended that what researchers call “learning to learn” can also help reduce the problem of learning software having to consume a large amount of data for a specific task to do it well. Researchers have challenged their software to create a self-learning system to collect a multitude of different yet related problems like locating mazes. It produces results that demonstrate a generalisation and acceptance ability for new tasks with less additional training than before.
The idea of creating AI from the original AI has appeared before, but previous tests cannot produce results that outperform products designed by humans. Yoshua Bengio, a professor at Montreal University who experimented this idea in the 1990s, said it was “really interesting”. He also pointed out that thanks to the more powerful configuration of computers today and the support of machine learning techniques, the idea of building AI by AI could really work (for instance, the researchers from Google Brain described the use of 800 powerful graphics processors to help software create the most competitive image recognition systems with humans’ designs); however, it also requires a lot of computing power. Even if Google AI is learning how to create AI software, it may not be practical enough to reduce the burden or partly replace machine learning experts.
Nonetheless, Oktrist Gupta, a researcher from MIT Media Lab, believes that this will change in the future. He and his colleagues at MIT planned to open-source AI for their own tests. The experiment will include programming of deep learning systems that can be adapted to human-made ones and used in standardized tests to identify objects.
Gupta’s inspiration to deploy the project comes from hours of depression he spent to design and test machine learning models. Gupta thinks that companies and researchers are motivated to find a way to put automated machine learning into practice, because “Reducing the burden on data scientists is a good payoff. You can work more productively, have better models and freely explore your ideas at higher levels. ”
Source: technologyreview, value walk