Why Microsoft’s big bet on deep learning is not a sure thing
When Microsoft looks into the near and far future, it sees artificial intelligence, particularly the hottest branch of it, called deep learning. The company has spent billions on AI, including a spate of acquisitions such as the deep-learning start-up Bonsai in late June and Semantic Machines in May.
Microsoft CEO Satya Nadella has said that AI is the “defining technology of our times”. He also said at an investor’s conference this spring: “It’s going to be AI at the edge, AI in the cloud, AI as part of SaaS applications, AI as part of in fact even infrastructure.”But there’s a chance that the big bet on deep learning could go bad if Microsoft fails to invest in other branches of AI that could be more useful. The issue isn’t that deep learning isn’t important – it certainly is. But there’s a chance it’s also been oversold and is already up against the limits of its capabilities.
That’s the argument that an increasing number of computer scientists and researchers are making. Deep learning works by taking a massive amount of data and essentially learns on its own, without developers writing specific programs for the task at hand. In one of its simplest uses, a machine could be taught to recognise the difference between cats and dogs by having hundreds of thousands of pictures fed into it, with each photo labelled as either a cat or a dog. The premise of deep learning is that this input is all the machine needs to learn how to correctly distinguish between them. How it does that is essentially a black box.
Deep learning is being applied to self-driving cars, medical research, and back-office automation, among other tasks. But as The New York Times notes in its article Is there a smarter path to Artificial Intelligence? Some experts hope so, “Some scientists are asking whether deep learning is really so deep after all.” The article goes on to say that “a growing number of AI experts are warning that the infatuation with deep learning may well breed myopia and overinvestment now – and disillusionment later.”
The problem, these researchers say, is that deep learning is by its very nature a limited technique, and only works on a specific set of problems, ones in which there are vast amounts of easily accessible data and where the tasks that need to be done are clear and well-defined, “like labelling images or translating speech to text,” in the words of the Times.
Because of that, Michael I. Jordan, a professor at the University of California, Berkeley, told the newspaper, “There is no real intelligence there. And I think that trusting these brute force algorithms too much is a faith misplaced.” And New York University professor Gary Marcus, warned in an article that “the patterns extracted by deep learning are more superficial than they initially appear”.
For evidence of that superficiality, look no further than Microsoft Office. Microsoft has built AI directly into Office, which sounds impressive. But take a close look at what that AI supposedly does and you might not be that impressed. It helps power Microsoft’s proofing tools in Word, for example – and you would be hard pressed to notice any difference in those proofing tools before and after AI was baked in. In Outlook, AI recommends when you should leave for an appointment. In PowerPoint, an AI-powered Designer feature recommends slide designs, which don’t seem like much more than templates. On the positive side, AI is useful for advanced Excel users, who can use it for extracting insights from complex sets of data.
AI will improve in the coming years and offer more useful tools. But because of the hype, Microsoft and other companies may be spending their resources on deep learning rather than on less touted AI technologies that may be more useful in the long run. That could hurt Microsoft and slow the benefits we all may get from the new technologies.
IDG News Service