Zhe Li

alt text 

Highlights:

  • Seasoned industry and academic expert with extensive experiences in developing large-scale, user-centric deep learning features on-device and strong publications.

  • Worked on large language model (LLM) training, inference, and infrastructure development, optimizing model deployment for scalability, latency, and efficiency.

  • Shipped the Face ID with mask feature, led the development of the face recognition system, and contributed to the implementation of spoof prevention and attention awareness feature.

  • Bridged the gap among learning algorithms (classification, detection, prediction, and others), feature specs (hand gesture, Persona, and others), hardware (GPU, ANE, sensors) capacities for Apple Vision Pro.

About me

I am a Member of Technical Staff at Patronus AI Inc., leading the development of reinforcement learning environments for Computer Use, Tool Use, and Finance Domain (e.g., Bloomberg Terminal and SAP) for frontier AI labs. Prior to joining Patronus AI, I was a Member of Technical Staff at Inflection AI Inc., where I worked on large scale model training, inference and infrastructure. Before Inflection AI, I was a staff machine learning and computer vision engineer at Apple, where my role was to bridge the gap between learning algorithms (classification, detection and others), features specs (hand gestures, Persona, Optic ID, others), hardware capacities (ANE, camera, ISP, SOC), and software implementation for Apple Vision Pro Product. Prior to this position, I was at the Face ID team, where I led the development of the first-generation face recognition for Face ID unlock with mask and the second-generation face recognition for Face ID in iPhone 13 and later version at Apple. I obtained my PhD from The University of Iowa in 2018, supervised by Prof. Tianbao Yang. My research focuses on optimizing deep neural networks from speeding up training, improving generalization ability, reducing model size and searching optimal network structure, and its application to computer vision and language tasks. I received my bachelor's degree and master's degree both in computer science from Xian Jiaotong University and South Dakota State University in 2010 and 2013, respectively. In case that you need to reach out me, my email is zhelixjtu/ [@] gmail [DOT] com.

Research Interest

My research interests fall into three levels:

  • understanding why machine can learn from learning theory perspective (statistical machine learning);

  • designing and analyzing optimization algorithms to facilitate machine learning efficiently;

  • utilizing machine learning models(Specifically, deep neural networks) to solve vision, language and other tasks.

Research and Work Experience

Recent Publications

  • Zhe Li, Tianbao Yang, Lijun Zhang, Rong Jin, "Fast and Accurate Refined Nystrom based Kervel SVM", AAAI 2016.
    [pdf][Bibtex][poster][slides]

Workshop

  • Zhe Li, Xiaoyu Wang, Xutao Lv, Tianbao Yang, "SEP-Nets: Small and Effective Pattern Networks".
    Beyond ISLVRC workshops 2017 Honululu, HI

  • Zhe Li, Tianbao Yang, Lijun Zhang, Rong Jin, "A Two-stage Approach for Learning a Sparse Model with Sharp Excess Risk Analysis"
    Learning Fast From Easy Data Workshop NIPS 2015, Montreal, Canada

Talks

  • “SEP-Nets: small and effective pattern networks” [pdf]

  • “Deep Learning With Caffe”[pdf], The University of Iowa, Apr 2017
    This lecture tried to teach students widely from The university of Iowa how to install caffe in the university cluster argon and how to train the very first deep neural network model for image classification using caffe. The installation in Argon clusteris quite tricky since students don't have root access to cluster. See Notes for detail information if you want to install caffe in your cluster environment.

  • “Improved Dropout for Shallow and Deep Learning”[pdf]

  • “A Two-stage Approach for Learning a Sparse Model with sharp Excess Risk Analysis”[pdf]

  • “Nystrom Based Kernel Classification for Big Data”"[pdf]

Demo

  • Object Detection: This is C++ implementation of RCNN based on caffe framework, in which sliding window, bounding box regression are utilized.

Notes

Here are some informal notes and blog posts I’ve written: