Microsoft: with me than Google brain monkey-AI, AI-IT information Microsoft: compared to my Google brain monkey

Coincidentally, all the major technology companies have recently with "artificial intelligence" and "deep learning" field. "Father of the Google brain" and now Baidu's Chief Scientist, Andrew Ng, a month ago in PingWest SYNC was mentioned in a, the General Assembly held in San Francisco, "deep learning is that many modern-day technology behind the" engine ", as a core technology supporting business application, such as Web searches, such as machine translation, product recommendations and medical image, bringing considerable economic value. ”

Now, Microsoft has also joined the "arena".

Researchers from Microsoft and their deep learning system--named Adam (ADAM)-achieved breakthrough results, and more mature than previous deep learning system. Researchers at Microsoft, for example, in image recognition, for example, the system can be identified not only the specified items, also under the category classification, a more precise identification. And the previous "Google brain" for comparison, if you say "Google brain" can do is, after watching the Youtube video of the week, identify cats, so Adam can see cats and cat breeds, and using only the number of machines before the one-thirtieth.

Pictured above from Wired. "Adam project" team members: from left to right as Karthik Kalyanaraman, Trishul Chilimbi, Johnson Apacible, Yutaka Suzue.

Microsoft Research Director Dr Peter Lee, "said Adam plan" sought not to break through analog neural network scale, but rather build on the improvements. These improvements is: Optimizing and fine-tuning the machine the way data is processed and the data communication between machines. When computing systems become more complex, the transfer information between each component at the same time – this matter will more and more difficult as the systems become large, and Adam system allows asynchronous algorithm existed. Asynchronous algorithm refers to a larger system is split into different sections, and in sharing and output results before allowing them to operate independently, and share results in the partial results cover each other when there is no problem.

The school of computer science at the University of Wisconsin researcher Feng Niu, and Benjamin Recht, who studies pointed out that stochastic gradient descent algorithms are used in traditional (SGD) to several parallel computing task, usually need to ensure synchronization of output, otherwise it will cause a data conflict. However, if it is used in small computing systems, first of all, data conflicts is very low. Secondly, since the data write tends to be incremental results, output and stack each section in a different order, so even conflict but the results are often the same. In this research report called "Hogwild!" Technology makes it possible to asynchronous computation in depth learning system.

In addition, the depth of Microsoft's learning system, different from the Google brain using GPU for building components, using instead the cost more expensive CPU, these machines are supported by Microsoft's Azure cloud service.