Kaiming He

Kaiming He

A Chinese computer vision researcher whose Residual Networks, published in 2015, let networks grow from tens of layers to hundreds. The trick was a skip connection that passes a layer's input forward unchanged alongside its output, so gradients have a short path back and each block only has to learn a residual correction. ResNets won ImageNet in 2015, and residual connections are now standard inside Transformers too. He also produced Faster R-CNN, Mask R-CNN and Masked Autoencoders, and teaches at MIT. (See also: Deep learning, Neural network, Backpropagation, Transformer)