Looks Like Google's AI Can Now Build Better Machine Learning Programs Than Humans
AutoML runs hundreds of simulations to improve and replicate its own code, over multiple cycles.
There are already experts in the field of artificial intelligence who believe it can be dangerous, particularly when creators don¡¯t take enough steps to restrict their capabilities. Now, they may have an even bigger concern.
Google¡¯s AutoML system, an AI the research team was teaching to replicate itself, has just produced a series of machine learning algorithms that are even more efficient than itself.
AutoML was originally started as a possible solution to the lack of developers in the cutting edge field. In order to keep up with the predicted demand for AI, the research team was attempting to build an AI program capable of self-replicating. It runs thousands of simulations to figure out which parts of the code it¡¯s creating can be improved, changes it, then repeats the process over and over. Indeed, the AI has shown it¡¯s capable of building machine learning programs in just hours, as compared to the weeks or months it takes humans.
More worrisome is the fact that AutoML seems to be even better than humans at building AI. In image recognition tests, AutoML¡¯s creations achieved a record-breaking 82 percent accuracy, much higher than anything developed by humans. It can mark multiple points in an image with 42 percent accuracy, whereas human-made software only reaches 39 percent.
Of course, if there¡¯s any worry of a Skynet-like AI awakening, we¡¯re far away from that. What we are certainly closer to is much better AI that we can produce at a much faster rate, so you can expect your smartphones assistants to get a lot smarter. And all of that in just the five months AutoML has been active.
