-
Zhihao Ouyang
+1 (571) 523-5851
EDUCATIONAL BACKGROUND
Carnegie Mellon University
Pittsburgh
07/2020
Ø Master of Computer Science (MSCS)
Tsinghua University
Beijing
07/2021
Ø Master of Computer Science , THU-CMU double degree program
Beihang University (BUAA)
Beijing
07/2018
Ø Bachelor of Engineering in Software Engineering & Minor in English; Rank 1 out of 141
Ø GPA: Major 3.88/4.0 & Minor 3.94/4.0; Graduate with Shenyuan Honor (Highest honor for students)
PUBLICATIONS
[1]. Zhihao O, Yuyuan Z, Zihao H, Yinghua G, Tao D, Shu-Tao X. Towards Randomness in Learning Rates. AAAI Conference on Artificial Intelligence (AAAI) 2020 (under review).
[2]. Zhihao O. Haoyu L, Yuyuan Z, Zihao H, Hang S, Jun Z, Shu-Tao X. Channel Specified Convolutional Neural Network. International Conference on Learning Representations (ICLR) 2020 (under review).
[3]. Zhihao O. Yan F, Zihao H, Tianbo H, Tao D, Shu-Tao X. AttentionDrop for Convolutional Neural Networks. International Conference on Multimedia and Expo (ICME) 2019.
[4]. Zhihao O. Yihang Y, Kun Y, Jian W, Xiaolin H. XiaoJoy: Using Locally Connected CNN to Cooperate with Human Composers (Artwork). Conference on Neural Information Processing System (NeurIPS) Workshop 2018.
[5]. Lu Y, Zhihao O, Xu S. Website Design and Development Case. Tsinghua University Press 2018.
EXPERIENCE
Interpretability and distangle of Convolutional Neural Network
03/2019– Present
Research Assistant, Advisor: Prof.Jun Zhu
Department of Computer Science & technology, Tinghua University
Ø Used channel-class gates and loss function to interprate the behavior of CNN kernels.
Ø Disentangle the CNN kernals using limit function and staged training methods.
Regulization and Robustness of Deep Learning Model
09/2018– Present
Department of Computer Science & Technology, Tsinghua University
Research Assistant, Advisor: Prof.Shutao Xia
Ø Proposed Adaptive Dropout for better regularization than original Dropout for CNN models.
Ø Utilized random learning rate to bring variance to the backpropagation gradient for better optimization and enhance the model’s robustness.
Music Generation Research and Engineering
04/2018– 12/2018
Tsinghua University & Deep Music company
Research Assistant, Advisor: Prof. Xiaolin Hu
Ø Used Bi-LSTM, Seq2Seq and other LSTM models for music generation.
Ø Designed CNN based on resNet and Locally Connected CNN structure to replace the existing RNN generation model, achieving higher precision and speed in the task of sequence generation.
Ø Models were successfully used in Deep Music company and appeared on TV show “The Voice of China”.
Face Recognition Research
04/2018- 2/2019
Megvii(Face++) Company
Research Intern, Mentor: Dr.Dong Li
Ø Designed face feature extraction network, using series structure such as Inception Block, Group Conv, IBNNet to improve network robustness and speed.
Ø Implemented face feature regularization methods from NormFace series of articles, and improved the optimization objective of the classification from the perspective of the angle.
Microsoft Congnitive Service Face API
05/2017– 08/2017
Microsoft Research Asia (MSRA)
Intern, Mentor: Gang Chen (Principle Manager)
Ø Maintained Microsoft Face API’s customer-related code, developed face recognition related technology.
Ø AI composer MelodyMaster were selected as the demoware in the MSRA showroom and won Award of Most Creative in Microsoft Hackathon (2/12824 ).
Projects and Entrepreneurship
Industrial defect detection system
1/2019– Present
Wise Sense Company
Co-Founder & CTO
Ø Designed a defect detection system for the iPhone microphone production line involves the use of latest computer vision algorithms such as classification, segmentation, and detection.
Ø Based on DenseNet and InceptionResNet, implemented a fast classification model. The accuracy of 99.5% and the speed of 50 milliseconds per component were achieved. Core algorithm for a 0.2 billion $ detection system.
Music Auto-Accompaniment Software “SoundMix” and “ChordMixer”
08/2017-04/2018
SoundMix and ChordMixer Project
Team Leader
Ø Developed a smart audio recognition software on iOS by using Fast Fourier Transform Algorithm (FFT) to realize the recognition for the pitch of instruments.
Ø Finished road show in Tencent headquarter; won the second place (2/4000) and the Outstanding Creative Award (5/4000) in Tencent Innovation and Entrepreneurship Competition,.
Ø Purchased by the Beijing Mousse Company.
Skills
Ø Programming Language: Python, C/C++, C#, HTML, JavaScript, Swift etc.
Ø Frameworks: Tensorflow, Keras, Pytorch, Flask, JQuery, Bootstrap
Ø Additional Skills: Product Presentation, Violin, Piano, Guitar, Composition, Design, Painting.