Microsoft

Contribution to AI

A residual connection is now so ordinary that it sits inside almost every deep network, transformers included, and it arrived in a 2015 paper from Microsoft Research Asia. He, Zhang, Ren and Sun's residual networks took first place in ImageNet classification, detection and localisation and in COCO detection and segmentation in the same season , and settled a practical question — whether networks hundreds of layers deep could be trained at all — that had been blocking everyone. The Beijing lab that produced it was itself a contribution: founded in 1998 under Kai-Fu Lee , it trained a large share of the people who later ran AI groups across the Chinese technology industry. Other Microsoft work changed what researchers could buy. Kinect put a depth camera at a games-console price into vision and robotics labs in 2010, and the pose-estimation and reconstruction work built on it moved out of the laboratory quickly. In 2016 a Microsoft system reached a 5.9 percent word error rate on the Switchboard conversational transcription task, matching human transcribers . The later contribution is structural rather than scientific. The billion dollars committed to OpenAI in 2019, half of it in Azure credits , and the ten-thousand-GPU machine built for it the following year , tied frontier model training to a cloud business and to Office, Windows and GitHub. That coupling is why a research result now reaches hundreds of millions of desks within months.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).