GV

GeezerVison

The World Wide Web, restored to 800 pixels of glory.

Home|About|Gaming cabinet|Archives|How it works|Starter sites

On this site

  • alphaXiv
  • Researchers
  • NeurIPS 2026
  • Sign in
  • MCP Server
  • Autoresearch
  • Browser Extension
  • Blog
  • Send Feedback?

This view

  • Original page
  • New address

GeezerVison is showing www.alphaxiv.org — Kaiming He on alphaXiv

Kaiming He

Distinguished Scientist at Google DeepMind

Associate Professor, EECS at MIT

Google ScholarWebsiteDBLPGitHubOpenReviewWikipediaEmail

About

Kaiming He’s ResNet introduced residual connections to facilitate the training of deep neural networks with hundreds of layers. He is an Associate Professor of EECS at MIT and a Distinguished Scientist at Google DeepMind. His research in CV and ML focuses on visual perception and self-supervised representation learning. He created Faster R-CNN to integrate region proposals into end-to-end detection and Masked Autoencoders to scale vision learners through masked pixel reconstruction. Previously, he was a research scientist at Facebook AI Research and a researcher at Microsoft Research Asia. He holds a PhD from the Chinese University of Hong Kong and a BS from Tsinghua University.

News

  • Jun 2026ResNet · CVPR 2026 Longuet-Higgins Test of Time Award.
  • Mar 2025Transformers without Normalization · #7 most read in March · 81k+ reads on alphaXiv.
  • Oct 2017Mask R-CNN · Marr Prize at ICCV 2017.
  • Jun 2016ResNet · Best Paper at CVPR 2016.

ExperienceUnverified

Distinguished Scientist

Google DeepMind

2025 – Present

2025 – Present

Associate Professor, EECS

MIT

Feb 2024 – Present

Feb 2024 – Present

Research in computer vision and deep learning.

Research Scientist

Facebook AI Research (FAIR)

2016 – 2024

2016 – 2024

Researcher

Microsoft Research Asia (MSRA)

2011 – 2016

2011 – 2016

Information Engineering PhD Student

The Chinese University of Hong Kong

2007 – 2011

2007 – 2011

Research Intern

Microsoft Research Asia (MSRA)

2006 – 2007

2006 – 2007

Worked in the Visual Computing Group under Sun Jian.

Physics BS Student

Tsinghua University

2003 – 2007

2003 – 2007

Papers101

Deep Residual Learning for Image Recognition

Deep Residual Learning for Image Recognition

Kaiming HeKaiming HeXiangyu ZhangXiangyu ZhangShaoqing RenShaoqing RenJian SunJian Sun

Shortcut connections let image-recognition networks train effectively at extreme depths, improving classification and object detection without adding complexity to identity skips.

10 Dec 2015
21kviews350kcitations
Faster R-CNN: Towards Real-Time Object Detection with Region Proposal
  Networks

Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks

Shaoqing RenShaoqing RenKaiming HeKaiming HeRoss GirshickRoss GirshickJian SunJian Sun

Sharing image features lets a detector generate object proposals with little extra computation, making accurate object detection practical at near-real-time speeds.

04 Jun 2015
7kviews113kcitations
Focal Loss for Dense Object Detection

Focal Loss for Dense Object Detection

Tsung-Yi LinTsung-Yi LinPriya GoyalPriya GoyalRoss GirshickRoss GirshickKaiming HeKaiming He+1

Down-weighting easy background examples lets single-stage detectors match or surpass two-stage detectors’ accuracy without sacrificing comparable inference speed.

07 Feb 2018
4kviews57kcitations
Feature Pyramid Networks for Object Detection

Feature Pyramid Networks for Object Detection

Tsung-Yi LinTsung-Yi LinPiotr DollárPiotr DollárRoss GirshickRoss GirshickKaiming HeKaiming He+2

Combining features across a network’s existing scales enables accurate detection of both small and large objects without repeatedly processing resized images.

09 Dec 2016
2kviews45kcitations
Delving Deep into Rectifiers: Surpassing Human-Level Performance on
  ImageNet Classification

Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification

Kaiming HeKaiming HeXiangyu ZhangXiangyu ZhangShaoqing RenShaoqing RenJian SunJian Sun

Rectifier-aware initialization lets very deep image classifiers train from scratch, while learned activation slopes improve ImageNet accuracy with almost no extra computation.

06 Feb 2015
774views31kcitations
Momentum Contrast for Unsupervised Visual Representation Learning

Momentum Contrast for Unsupervised Visual Representation Learning

Kaiming HeKaiming HeHaoqi FanHaoqi FanYuxin WuYuxin WuSaining XieSaining Xie+1

Unlabeled images can pretrain visual features that transfer effectively to object detection and segmentation, sometimes outperforming features trained with ImageNet labels.

13 Nov 2019
7kviews22kcitations
Spatial Pyramid Pooling in Deep Convolutional Networks for Visual
  Recognition

Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition

Kaiming HeKaiming HeXiangyu ZhangXiangyu ZhangShaoqing RenShaoqing RenJian SunJian Sun

Pooling convolutional features across spatial regions lets image classifiers and detectors handle varied image sizes while avoiding repeated feature computation for each candidate region.

18 Jun 2014
639views20kcitations
Masked Autoencoders Are Scalable Vision Learners

Masked Autoencoders Are Scalable Vision Learners

Kaiming HeKaiming HeXinlei ChenXinlei ChenSaining XieSaining XieYanghao LiYanghao Li+2

Training vision models to reconstruct heavily masked images produces transferable representations without labeled pretraining data, improving performance as model size grows.

02 Dec 2021
10kviews19kcitations
Aggregated Residual Transformations for Deep Neural Networks

Aggregated Residual Transformations for Deep Neural Networks

Saining XieSaining XieRoss GirshickRoss GirshickPiotr DollárPiotr DollárZhuowen TuZhuowen Tu+1

Increasing the number of parallel transformations improves image recognition accuracy, even when network depth, model size, and computation remain roughly constant.

16 Nov 2016
663views18kcitations
Identity Mappings in Deep Residual Networks

Identity Mappings in Deep Residual Networks

Kaiming HeKaiming HeXiangyu ZhangXiangyu ZhangShaoqing RenShaoqing RenJian SunJian Sun

Deep residual networks have emerged as a family of extremely deep architectures showing compelling accuracy and nice convergence behaviors. In this paper, we analyze the propagation formulations behind the residual building blocks, which suggest that the forward and backward signals can be directly propagated from one block to any other block, when using identity mappings as the skip connections and after-addition activation. A series of ablation experiments support the importance of these identity mappings. This motivates us to propose a new residual unit, which makes training easier and improves generalization. We report improved results using a 1001-layer ResNet on CIFAR-10 (4.62% error) and CIFAR-100, and a 200-layer ResNet on ImageNet. Code is available at: this https URL

16 Mar 2016
546views16kcitations

Writing1

MiniT2I: A Minimalist Baseline for Text-to-Image Generation

MiniT2I: A Minimalist Baseline for Text-to-Image Generation

XianbangXianbangHanhong ZhaoHanhong ZhaoYiyang LuYiyang LuKangyang Zhou+2
16 Jun 2026
24views6citations

Followers

See all
5,431

Metrics

899,779Citations
78h-index
95i10-index

Citations per year

111,106 citations in 2026
20162026

Coauthors

Ross Girshick
Ross Girshick28 shared papers
Jian Sun
Jian Sun16 shared papers
Piotr Dollár
Piotr Dollár16 shared papers
Saining Xie
Saining Xie10 shared papers
Xinlei Chen
Xinlei Chen10 shared papers
Tianhong Li
Tianhong Li9 shared papers
Shaoqing Ren
Shaoqing Ren7 shared papers
Xiangyu Zhang
Xiangyu Zhang7 shared papers
Haoqi Fan
Haoqi Fan7 shared papers
Jifeng Dai
Jifeng Dai6 shared papers
Christoph Feichtenhofer
Christoph Feichtenhofer6 shared papers
Yanghao Li
Yanghao Li5 shared papers
Yiyang Lu
Yiyang Lu5 shared papers
Hanhong Zhao
Hanhong Zhao5 shared papers
Alexander Kirillov
Alexander Kirillov4 shared papers
Yuxin Wu
Yuxin Wu4 shared papers
Zhuang Liu
Zhuang Liu3 shared papers
Georgia Gkioxari
Georgia Gkioxari3 shared papers
Wojciech Matusik
Wojciech Matusik3 shared papers
Zhengyang Geng
Zhengyang Geng3 shared papers
You are visitor 01402605 — GeezerVison does not store the pages you look up.
Best viewed with Netscape Navigator 4.7 or Internet Explorer 6 at 800×600. No cookies required.