# Kaiming He > ~1984– · Computer Scientist, AI Researcher > > **Recorded contribution:** ResNet (residual networks) — enabled training of 100+ layer networks; Mask R-CNN; MAE ## How to use this dossier Read for a causal chain, not a hero story: inherited problem → contribution → mechanism → downstream capability → limit. Then close the page and complete the reconstruction exercise from memory. ## 1. Historical orientation Chinese computer scientist Kaiming He led or co-authored influential computer-vision research including ResNet, Mask R-CNN, and masked autoencoders. The 2015 residual-network paper showed that very deep networks could be optimized more reliably by learning residual functions around identity shortcuts, enabling 100-plus-layer models and winning major ImageNet and COCO competitions. Residual learning changed the optimization target inside a deep network. If an ideal block should preserve its input, ordinary stacked layers must relearn identity; a shortcut supplies identity directly and asks the learned branch only for the difference needed to improve it. ## 2. The problem inherited Adding layers to neural networks should increase representational capacity, yet sufficiently deep 'plain' networks often trained worse because optimization degraded even when overfitting was not the cause. ## 3. The central contribution He and collaborators introduced deep residual learning, making identity shortcut connections a simple, scalable solution to degradation in very deep vision networks. ## 4. Reconstruct the mechanism 1. Let a block learn a residual transformation F(x) rather than a full mapping H(x). 2. Add the unchanged input through an identity shortcut to produce F(x)+x. 3. Backpropagate through both residual and identity paths so gradients have a direct route. 4. Stack many residual blocks and change dimensions only where required. ## 5. What changed downstream - ResNet became a standard backbone across vision and influenced architectures far beyond image classification. - Residual connections became a core component of Transformers and many other deep models. - Residual blocks enabled much deeper vision systems and became a reusable architectural pattern in detection, segmentation, generation, language modeling, and scientific machine learning. ## 6. Attribution, limits, and uncertainty - ResNet is a four-author Microsoft Research Asia result with predecessors in shortcut and highway networks. - Residual paths ease optimization but do not guarantee robustness, interpretability, efficient inference, unbiased data, or useful real-world objectives. - Shortcut connections improve gradient flow but do not explain every learned feature or eliminate data bias, adversarial vulnerability, compute cost, or the need to match architecture to scale. ## 7. Reconstruction lab Train a 20-layer plain network and residual counterpart on the same small image task. Compare training error, gradients, parameter count, and calibration rather than only final accuracy. Measure gradient norms by depth and include a control whose shortcut uses a projection, separating the effect of identity flow from merely adding parameters. ## 8. Evidence trail - [Deep Residual Learning for Image Recognition](https://arxiv.org/abs/1512.03385) — CVPR / arXiv - [Mask R-CNN](https://arxiv.org/abs/1703.06870) — ICCV / arXiv - [Kaiming He](https://en.wikipedia.org/wiki/Kaiming_He) — Wikipedia contributors --- *Research checked 2026-08-09. Dates, roles, and claims about living people are historical snapshots. Linked sources remain the authority; this dossier is original instructional synthesis.*