Microsoft Research

Microsoft Research (MSR) is the research division of Microsoft. It was founded in 1991 by Richard Rashid, Bill Gates and Nathan Myhrvold, and it employs more than 1,000 researchers.^[1] Peter Lee, a corporate vice president, directs the core research labs.^[1] Like other industrial labs, its job is to do open-ended research while also feeding ideas and people back into the company's products.

Where the labs are

The first lab opened in Redmond in 1991, next to the company's main campus. A second site followed in Cambridge, UK, in 1997, and Microsoft Research Asia opened in Beijing in November 1998. More sites came later: New England in 2008, New York City on 3 May 2012, Montreal after the 2017 acquisition of Maluuba, and Tokyo in November 2024.^[1] The division's own about page lists labs in Redmond, Cambridge, Asia-Pacific, India, Montreal, New England, New York City and a Catalyst Lab.^[6] The spread of locations matters for a research organisation: it lets the company hire in places where strong universities and a local AI industry already exist, rather than asking everyone to move to one headquarters.

What came out of it

Its best-known single result is an image-recognition method called ResNet. The paper "Deep Residual Learning for Image Recognition", by Kaiming He, Xiangyu Zhang, Shaoqing Ren and Jian Sun, won first place in the ILSVRC 2015 classification task with a top-5 error of 3.57%.^[2] The same paper also reports first places on the ImageNet detection and ImageNet localization tasks.^[2] It received the CVPR 2016 Best Paper Award.^[5]

The idea behind it is easy to state. Stacking more layers onto an image-recognition network ought to make it stronger, but in practice very deep networks become hard to train. Residual learning sidesteps this by having a block of layers learn a correction relative to its own input rather than the full transformation, so the signal can pass through many layers without degrading. That simple change made networks with far more layers practical and became a standard building block in computer vision.

Research from the labs has also gone into shipping products, including Kinect, Bing, HoloLens, Cortana and Microsoft Translator.^[1] This is the other half of an industrial lab's purpose: speech recognition, machine translation, vision and search ranking are all research problems before they are product features.

People who passed through

Many researchers spend a few years at MSR and then move on. Jacob Devlin, later the first author of BERT, was a research scientist at Microsoft Research from 2014 to 2017, where he helped move Microsoft Translator from phrase-based to neural machine translation.^[3] Tomáš Mikolov spent time at Microsoft Research in Redmond during his PhD.^[4] Visits and internships of this kind are one way a corporate lab stays connected to academic work; the people who pass through carry methods and habits in both directions.

Its position in the field

Microsoft's research effort has been unusually productive in AI. Between 2010 and 2018 the company held about 20% of the 154,000 AI patents filed worldwide, the largest share of any company.^[1] Patents are a rough measure at best, but the figure indicates how much of the company's research output in that period was in machine learning and related areas rather than only in systems and software engineering.

Read together, the record describes an institution built to do two things at once: publish work that other researchers build on, such as ResNet, and supply the techniques behind products that millions of people use.