MIT And IBM Engineers Find A Method To Train AI Faster

Machine learning has helped the human in various fields with our computer systems, from face identifying to reading medical scans. However, there has been a long way to go when it comes to processing events in the real world or interpreting videos because it could become cumbersome. Thankfully, we will soon have a solution as a team of researchers from the Watson AI lab, a partnership between MIT and IBM, have successfully found a way to resize video recognition models, speed up training as well as make it better when being used on smartphone devices.

The trick used by researchers is that they will adjust the view time of video recognition. Instead of using current models that encode images into bigger and computationally-intensive models, they made use of a temporal shift module to make the model pass a sense of time with no explicit representation. The tests resulted in a three-time-faster speed when training video-recognition AI, in comparison with existing methods. Furthermore, the new method could help mobile devices run video recognition models more easily. According to Song Han - an assistant professor at MIT, their goal is to help everyone be able to approach AI technology with, even with low-power devices. To complete this project, they will have to design AI models that are not only efficient but also less-energy consuming. If successful, AI can smoothly run on such devices in the near future.

Once the training needs less computing power, it is also expected to reduce the carbon footprint that AI produced. For example, YouTube and Facebook will be more effective in spotting violent and terrorist footage, or hospital will be able to use AI apps locally to keep private data away from prying eyes onto internal storage.