Chenzhuo Zhu 朱宸卓

alt text 

Chenzhuo Zhu
Ph.D. Student
Department of Electrical Engineering
Stanford University

Email: czhu95 [at] stanford [dot] edu

Research Interests

My research focuses on deep learning and computer architecture. On one hand, I'd like to address vision tasks with specifically designed neural networks and further understand the learned knowledge behind their performance. On the other hand, I'm interested in improving the energy efficiency of large neural networks to expand its applications to mobile platforms.


  • Feb 5, 2017: “Trained Ternary Quantization” is accepted to appear at ICLR 2017 as a full paper.

  • Jan 16, 2017: “Shape-independent Hardness Estimation Using Deep Learning and a GelSight Tactile Sensor” is accepted to appear at ICRA 2017 as a full paper.

Research Experience


CVA Lab, Stanford
Oct. 2016 - Dec. 2016
Advisor: Prof. Bill Dally
Thesis: Neural Network Compression


Computer Science and Artificial Intelligence Lab, MIT
Jun. 2016 - Sep. 2016
Advisor: Prof. Edward Adelson
Thesis: Hardness Measurement Using a GelSight Sensor with CNN and LSTM


Sensetime Beijing
Feb. 2016 - Jun. 2016
Supervisor: Dr. Jianping Shi
Thesis: Instance-Aware Image Segmentation


3D Image Lab, Tsinghua University
Apr. 2015 - Dec. 2015
Advisor: Prof. Huimin Ma
Thesis: High-speed Pupil Tracking System | Boundary-Aware Box Refinement for Object Proposal Generation

Selected Projects


Trained Ternary Quantization
Deep neural networks are widely used in machine learning applications. However, the deployment of large neural networks models can be difficult for mobile devices with limited power budgets. To solve this problem, we propose Trained Ternary Quantization (TTQ), a method that can reduce the precision of weights in neural networks to ternary values. This method has very little accuracy degradation and can even improve the accuracy of some models. We highlight our trained quantization method that can learn both ternary values and ternary assignment. During inference, our models are nearly 16× smaller than full-precision models.


Shape-independent Hardness Estimation with CNN and LSTM
Tactile sensing is an important component of interactive learning, and hardness is among the most attributes of an object that humans learn about through touch. In this work, we address enable robots’ sense of hardness by introducing a novel method based on the GelSight touch sensor. We analyze it using a deep convolutional (and recurrent) neural network. We can directly observe through deformation of the gel how the geometry of the samples and the contact force changes over time. We also published a dataset of GelSight videos and annotations.
[website] [dataset]


Boundary-aware Box Refinement for Object Proposal Generation
We present an effective approach to improve the localization quality of object proposals. We leverage the boundary-preserving property of superpixels and design an efficient algorithm for object proposal refinement. Our approach first performs bounding box alignment to adapt proposals to potential object boundaries, and then diversifies the proposals via multi-thresholding superpixel merging. The algorithm only takes 0.15s and can be applied to any existing proposal methods to improve their localization quality.



  • Languages: Mandarin Chinese (Native), English

  • Programing Languages: C/C++, Cuda, Python, MATLAB, Lua, Verilog HDL, Java

  • Tools: Caffe, Torch, Tensorflow, OpenCV, Git, LaTeX