Microsoft's research in machine and deep learning on display at NIPS 2015

Staff Writer

Microsoft’s recent forays into improving machine learning and deep learning behind projects such as Skype Translator, Cortana, and Clutter have come to greater fruition as time passes. The company’s work and advancements in deep learning and machine learning have been held in high regard at the Neural Information Processing Systems, or NIPS for short, conference in 2015. According to Microsoft, however, the era of machine and deep learning ate just beginning.
Skype Translator works to translate spoken words in real-time during Skype calls, and we have seen that the likes of this technology has previously been the work of science fiction universes like Star Wars and Star Trek. Since her initial introduction in Windows Phone 8.1, Cortana has grown into a personal digital assistant that can handle complex questions, as well as notice aspects of your query and offer to keep you up-to-date on said aspects. Cortana can also handle package and flight tracking and even allow SMS sending from a desktop interface to a phone powered by Windows 10 Mobile. Clutter, contrary to the connotation of disorder and messiness, is Microsoft’s implementation of an inbox-cleaner that hands important and unimportant emails in Outlook 2016 while making the distinction of importance for itself, rather than requiring a user to set filters or parameters.
All of these are just implementations of deep and machine learning, and Microsoft is continuing to improve the learning processes and methods themselves. Li Deng, a partner research manager in Redmond, said that “People see machine learning as more and more important – and deep learning is increasingly central to business”, and he appears to be right. Cynthia Dwork, a cryptographer and scientist at Microsoft Research, has discussed and written how just one improvement in a computer’s algorithm can lead to numerous errors being avoided, and thus both saving time and allowing new data to be processed. The algorithm she describes is one that is at work currently within Microsoft, and it works to prevent data from being able to register its irrelevant relationships and allowing its actual relationship to be discovered instead.
Microsoft’s research lab in Cambridge, U.K., also has made advancements in machine learning, specifically with computer vision. If a computer is allowed to see a single perspective or angle of a three-dimensional object, the computer, thanks to advancements made in 3D graphical rendering engines, can rotate the object to show another perspective or angling of it, even if it’s never encountered the object before.
A large hurdle for machine learning is its exposure to the outside world. Unfortunately, those working on it can’t accommodate for all the random and necessary real-world occurrences that non-experts might introduce. Instead, researchers are now working on machine teaching, where the computer provides resources and guidelines for a non-expert user. This all allows for non-interaction with algorithms that sit as a much larger part of the improvements in machine and deep learning. Deng is also quoted by Technet as saying that “In speech recognition, visual object recognition, and a few other areas of AI if you’re not in deep learning, you’re outside the mainstream”. He related a number of roughly 100 people at Microsoft working on deep learning, with another portion working in related and overlapping fields not accounted for in the estimation. Both what he had to say about the integral nature of the focus on deep learning and Microsoft’s increased attention to both deep and machine learning prove that, if nothing else, computers can be taught to understand as a human can, but only with a human’s help.