1984–, Computer scientist
Kaiming He is a Chinese-American computer scientist whose 2015 paper with Xiangyu Zhang, Shaoqing Ren and Jian Sun Deep Residual Learning for Image Recognition introduced the residual connection, a "shortcut" identity path around each block of layers that lets gradients flow directly through the network. ResNet won the 2015 ImageNet challenge with a 152-layer network at 3.6% top-5 error (below the human estimate of ~5%) and made networks of arbitrary depth practically trainable.
Residual connections are now ubiquitous: every modern transformer block uses them, every modern image model uses them, every modern speech model uses them. The architectural innovation is arguably the single most influential contribution of the AlexNet-to-Transformer era.
Kaiming He was at Microsoft Research Asia and then Facebook AI Research, where his subsequent work covered Mask R-CNN (2017, the standard instance-segmentation model), Momentum Contrast (MoCo, 2019, a leading self-supervised learning method) and many other contributions. In 2024 he moved to MIT as a faculty member.
Video
Related people: Alex Krizhevsky
Works cited in this book:
- Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification (2015) (with Xiangyu Zhang, Shaoqing Ren, Jian Sun)
- Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks (2015) (with Shaoqing Ren, Ross Girshick, Jian Sun)
- Deep Residual Learning for Image Recognition (2016) (with Xiangyu Zhang, Shaoqing Ren, Jian Sun)
- Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour (2017) (with Priya Goyal, Piotr Dollár, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia)
- Focal Loss for Dense Object Detection (2017) (with Tsung-Yi Lin, Priya Goyal, Ross Girshick, Piotr Dollár)
Discussed in:
- Chapter 11: CNNs, CNNs in Vision