Naiyan Wang

Timeline

1. 2013

   A deep learning tracker

   As a doctoral student at the Hong Kong University of Science and Technology, he coauthors “Learning a Deep Compact Image Representation for Visual Tracking” with Dit-Yan Yeung for NIPS 2013. The tracker trains a stacked denoising autoencoder offline on auxiliary natural images and transfers its encoder to online tracking.

2. August 2015

   A doctorate on visual tracking

   He defends his PhD thesis in the Department of Computer Science and Engineering at the Hong Kong University of Science and Technology in August 2015, with Dit-Yan Yeung as supervisor. The thesis, “A Unified Framework and Applications of Machine Learning Methods for Visual Tracking,” splits a tracker into five parts, tries dictionary learning, a stacked denoising autoencoder and a convolutional neural network for feature learning, and combines trackers with a factorial hidden Markov model.

3. 2015

   Core developer of MXNet

   In December 2015, he is one of ten authors of the MXNet paper, which describes a multi-language machine learning library for deep neural networks that runs on systems from mobile devices to distributed GPU clusters. He works on the framework as a core developer.

4. 2015–2024

   Chief scientist at TuSimple

   Right after his doctorate, he joins TuSimple as chief scientist in 2015 and builds its algorithm team in China from scratch. In September 2016, perception algorithms from his team win ten world firsts on the KITTI and Cityscapes datasets. From 2019, he serves as chief technology officer for TuSimple's China region.

5. 2016

   Adaptive Batch Normalization

   In March 2016, he coauthors a paper with Yanghao Li, Jianping Shi, Jiaying Liu and Xiaodi Hou that introduces Adaptive Batch Normalization (AdaBN). The method adjusts the statistics in all Batch Normalization layers of a network to adapt it to a new domain, and it adds no parameters.

6. 2021–2022

   Sparse detectors for LiDAR

   In December 2021, he coauthors the Single-stride Sparse Transformer (SST) for LiDAR-based 3D object detection, which keeps the original resolution from the start to the end of the network and reaches state-of-the-art results on the Waymo Open Dataset. In July 2022, he coauthors Fully Sparse 3D Object Detection (FSD), whose cost grows roughly linearly with the number of points and is independent of the perception range.

7. May 2024–September 2026

   L3 lead at Xiaomi Auto

   In May 2024, he joins Xiaomi Auto to lead autonomous driving technology development and L3 projects, reporting to Ye Hangjun, head of Xiaomi Auto's autonomous driving division. Before the move, he has published more than 40 papers in computer vision and machine learning conferences and journals.

8. Since September 2026

   An embodied AI venture

   In September 2026, he leaves Xiaomi Auto to start a company on the core perception, decision-making and planning abilities of robots, aiming to build embodied AI foundation models. His team of more than ten researchers and engineers includes several former TuSimple colleagues.