Microsoft Research

Contribution to AI

Deep learning's first commercial payoff came in speech rather than vision. In 2011 a collaboration between the Redmond and Beijing labs grafted deep neural networks onto the context-dependent tied-state HMMs that recognisers already used, and cut the word-error rate on the Switchboard telephone-conversation benchmark from 27.4 per cent to 18.5 — a third off a figure that a decade of Gaussian-mixture refinement had barely moved. That happened before AlexNet, and it is the reason large companies with shipping products started staffing neural-network groups. The Beijing lab then supplied the more portable idea. Residual connections, introduced in 2015, let a network learn a correction to its input rather than a fresh mapping, which made 152 layers trainable when depth beyond about thirty had been making things worse; ResNet took every ImageNet and COCO track it entered that year, and the same shortcut now sits inside every transformer block. Other work set the terms of measurement: the COCO dataset of 2014 annotated everyday scenes with per-instance segmentations and became the standard against which detection was scored, and LightGBM became a default for tabular problems that neural networks never took over. Kinect's pose tracker, classifying each depth pixel into a body part using forests trained on synthetic renderings, put a learned vision system in millions of living rooms. Microsoft Research Asia, opened in 1998, mattered as an institution: much of the leadership of China's AI industry trained there first.

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