I am currently a Staff Applied Research Scientist and tech lead at Cruise, working on LiDAR detection and segmentation models.

Previously, I was a Staff Research Scientist and employee #7 at Ambient.ai, building computer vision models for video understanding. Prior to that, I worked as a Research Engineer at Google Brain, publishing research in large-scale distributed training and neural architecture search.

I completed my M.S. in Machine Learning at Carnegie Mellon University in 2017 and my B.S. in Computer Science also at CMU in 2016.

My research interests are:

• Solving multi-task problems in computer vision using compositional data primitives.
• Continuous learning paradigms for robotics, particularly in autonomous driving.
• Using "outer-loop" optimization techniques like evolution to search for more accurate and efficient neural architectures.
• Combining the fields of deep learning and high-performance computing to train larger and faster neural networks.
• Applying machine learning techniques to solve the climate crisis.

Outside of work and academia, I'm an amateur photographer and a fan of anything outdoors (e.g. hiking, biking, camping). Fun fact: I really enjoy playing Dance Dance Revolution.

• NAS-Bench-101: Towards Reproducible Neural Architecture Search
  
  Chris Ying, Aaron Klein, Esteban Real, Eric Christiansen, Kevin Murphy, Frank Hutter

  ---

  NAS-Bench-101 is the first large-scale neural architecture search benchmark dataset, with over 423K unique architectures and 5M trained models. This dataset enables researchers to test their NAS algorithms in milliseconds by querying a pre-computed tabular dataset. ICML 2019
  [arXiv]   [poster]
• Image Classification at Supercomputer Scale
  
  Chris Ying, Sameer Kumar, Dehao Chen, Tao Wang, Youlong Cheng

  ---

  Utilized distributed batch normalization, input pipeline optimizations, and 2D torus all-reduce to train ResNet-50 at petaFLOPS scale on a TPUv3 pod. We achieve 76.3% accuracy in 2.2 minutes, with throughput of over 1.05M images per second. NeurIPS 2018 Systems for ML
  [arXiv]
• Depth-Adaptive Computational Policies for Efficient Visual Tracking
  
  Chris Ying, Katerina Fragkiadaki

  ---

  Proposed multi-depth fully convoltuional Siamese networks with parametric gating to perform efficient visual tracking. We train with a custom soft-budgeting joint loss but run inference with conditional computation to reduce runtime cost. EMMCVPR 2017
  [arXiv]   [poster]
• Don't Decay the Learning Rate, Increase the Batch Size
  
  Samuel L. Smith, Pieter-Jan Kindermans, Chris Ying, Quoc V. Le

  ---

  Demonstrated both theoretically and practically that learning rate, momentum, and batch size are closely related and proposes an alternate strategy for training which enables larger batch training without accuracy loss. ICLR 2018
  [arXiv]
• Large-Batch Training for LSTM and Beyond
  
  Yang You, Jonathan Hseu, Chris Ying, James Demmel, Kurt Keutzer, Cho-Jui Hsieh

  ---

  Developed a novel warmup scheme for training neural networks at extremely large batch sizes which enables speeding up training on large distributed systems without accuracy loss.
  [arXiv]
• Enumerating Unique Computational Graphs via an Iterative Graph Invariant
  
  Chris Ying

  ---

  Proposed a novel iterative graph invariant for identifying unique computational graphs (colored directed acyclic graphs). This algorithm can be used for identifying computationally unique neural networks.
  [arXiv]
• Improving Event Coreference using Knowledge Bases
  
  Chris Ying, Bishan Yang

  ---

  Utilize knowledge bases like YAGO and DBpedia to improve event coreference.
  [research poster]
• ParaBDD: Parallel Binary Decision Diagrams for Efficient Model Checking
  
  Chris Ying, Matt Bryant

  ---

  Parallel BDD library using Intel CilkPlus and lockless hash tables.
  [Writeup]
• Hypercube: Breaking the 4th Wall in Minecraft
  
  Chris Ying, Andrew Benson

  ---

  Minecraft mod which adds a custom block that controls with a real-world light.
  [YouTube demo]
• PacAttack
  
  Chris Ying, Taehoon Lee

  ---

  PaperJS-based web game that utilizes flocking behavior to create interesting emergent mechanics.
  [Play game]

Coming soon™!

Email: contact (AT) chrisying (DOT) net

Github: @chrisying

LinkedIn: ying-chris

Résumé: Click here (last updated 2021 Q4)