In an experiment, AI researchers from Google Brain allowed software to develop the design for a machine learning system to recognize human speech. This software produced better results than software designs for machine learning that had previously been created by people, the researchers write in a scientific paper submitted to a conference. This paper has not yet been subject to peer review – so testing of these results by other researchers still remains to be done.
Other research groups have also reported similar progress in recent months in the field of machine learning using artificial neural networks – the most promising technology in artificial intelligence in recent years. These include researchers at the non-profit research institute cofounded by Tesla and SpaceX founder Elon Musk, OpenAI, prestigious technology university MIT, the University of California Berkeley and the internal AI research department competing with Google Brain, Deepmind, Technology Review reports.
The idea of a self-learning artificial intelligence that can create its own AI has been around for a long time. But until now researchers had been unable to generate programs that could compete with those written by people. Many AI researchers believe that it will someday be possible to create an artificial intelligence that will solve general problems just like the human brain. Some believe that this will result in a constantly self-improving super intelligence, whose rise is designated as the technological singularity .
Google's Deepmind team beat a human Go player for the first time with AI in early 2016, attracting great attention from the public in the huge progress made in the field of artificial intelligence through machine learning using artificial neural networks. This approach is also known as deep learning .
If self-developed AI systems become usable beyond the scope of mere scientific experiments, this could significantly speed up the development of machine learning software. Machine learning experts are currently urgently sought by many companies. With the constantly increasing number of sensors in our world, ever more data is being generated. But AI software is needed to derive rational business decisions from all this information.
One aspect of the work facing AI software programmers could now be automated, as Google Brain Group director Jeff Dean told Technology Review. He calls the research branch his team is working on "automated machine learning."
At the AI Frontiers conference in Santa Clara, California (USA), Dean presented additional progress achieved by his research group, the report says. Experiments by the Google researchers indicate that the massive data volumes currently needed to teach AI something could be significantly reduced in the future.
The Google researchers were able to use a series of specific problems, such as solving a labyrinth, to teach AI broader problem solving abilities for other types of tasks as well. Smaller data sets are then needed to train neural networks for other problems.
The availability of specialized, extremely powerful chips has driven the deep learning approach towards this AI research breakthrough in recent years. For example, the Google Brain team uses 800 connected high performance graphic cards for their experiments.