Li-Ren Hou
Love coding, Love sharing knowledge, straightforward and honest
Ability
Computer Vision / Deep Learning / Machine Learning
- Experience in optimizing neural network model latency and power efficiency
- Developing semi-automatic tool to search suitable model configs on MTK chip.
- Familiar with image classification, especially in face-related applications
- Participated international image classification competition - CVPR2020 Face Anti-Spoofing Attack Detection Challenge. Won the 8th place out of 186 contestants.
- Developed 2D/3D face anti-spoofing recognition systems for 2.5y and applied them in access control systems.
- Able to implement DL models from scratch according to papers and approximate the accuracy the papers claimed.
- Experience in building traditional machine learning models on time-series data for biomedical applications.
- [Master Thesis] Built a anomaly detection model to detect abnormal health condition through analyzing physical activities with accelerometer and load cell sensor.
- Master thesis was praised by Golden Smart Home Technology Crop. and transferred the technology to it
- Published a conference paper - "Detection and Assessment of Abnormal Circadian Rhythm by Analyzing Rest/Activity Cycle" to ICME 2014
Programming
- Have solid background in information engineering
- Both university and graduate school majored in Computer Science and Information Engineering
- Won five presidential awards in the university
- Served as a teaching assistant for UNIX system programming course
- Have about 7 years of working experience as a software engineer
- Developed neural network models and inference programs using Python/C++
- Developed Android applications on Zenfone and Zenbo robot using Java/C++
English
- TOEIC 815/990
Work Experience
MediaTek - Senior Software Engineer, 2021.07 - Present (0.6 years)
- AI model co-optimization, profiling and debugging with customers
ChungHwa Telecom Lab - Associate Researcher, 2018.03 - 2021.05 (3.2 years)
- Developed 2D/3D face anti-spoofing recognition models and inference module
ASUS - Sensor Software Engineer, 2014.09 - 2018.03 (3.5 years)
- Got a promotion to senior software engineer after 1.8 years
- Developed Reminder application on ASUS Zenbo robot
- Developed MP3 Tag Editor and FM radio application on Zenfone
- Ported Google Mobile Services (GMS) to different phones
Education
National Taiwan University - Computer Science & Information Engineering, 2011 - 2014
National Sun Yat-Sen University - Computer Science and Engineering, 2007 - 2011
National Tung-Shih Senior High School, 2004 - 2007
AI Competitions and Projects
[Tool Development] Model Config Tuner
- Profile model inferences on MTK chip with different model configs for searching the optimal config
- Main algorithms: Developing
- Key tools: NeuroPilot SDK
[Competition] CVPR 2020 Face Anti-Spoofing Challenge [OfficialWeb] [Code]
- Won the 8th place out of 186 contestants
- Main algorithms: MIMAMO Net, local patches, ensemble learning and cosine annealing
- Key tools: pytorch, opencv and imgaug
[Product Development] 3D Face Anti-Spoofing, 2021
- Reached average accuracy 99.9% in 10.3K test images
- Main algorithms: MobileNetV2, BBox augmentation, Quantization, Cosine Annealing and MTCNN
- Key tools: pytorch, libtorch, tensorflow and opencv
[Product Development] 2D Face Anti-Spoofing, 2018-2021
- Reached TPR=99.7% @ FAR=1.0% in about 400K test images
- Main algorithms: ensemble learning, resnet50, MTCNN
- Key tools: pytorch, libtorch, tensorflow and opencv
[Training Project] Intrusion Detection System, 2018 [Slides]
- Main algorithms: MLP and Autoencoder
- Key tools: keras, pandas, numpy
Reproduced Research
Central Difference Convolutional Network, Sep 2020 - Oct 2020
- Studied source code and reimplement preprocessing for applying to products
- Main algorithms: CDCN++, 3DDFA_v2
- Key tools: pytorch and opencv
- Reference Paper [Link]
Rebroadcast Attack Recognition CNN, Oct 2018 - Feb 2019
- Implemented the model from scratch and successfully approximate the accuracy the papers claimed
- Main algorithms: local patches
- Key tools: tensorflow and opencv
- Reference Paper [Link]
Thesis and Paper
[Master Thesis] Detection and Assessment of Abnormal Health Condition by Analyzing Circadian Rhythm and Physiologic Reserve, July 2014 [Link]
- Proposed a health assessment system, which detects and assess abnormal health condition by using circadian rhythm and physiologic reserves.
- Main algorithms: SVDD
- Key tools: Android SDK, Matlab, libsvm
- Sensors: Wearable accelerometer and load cell
Detection and Assessment of Abnormal Circadian Rhythm by Analyzing Rest/Activity Cycle, June 2014 [Link]
- A rhythm assessment system was proposed to
automatically detect and assesse abnormal circadian rhythm by
analyzing rest/activity cycle.
- Main algorithms: SVDD
- Key tools: Android SDK, Matlab, libsvm
- Sensors: Wearable accelerometer
- Conference: International Conference on Complex Medical Engineering (CME2014)
Tool Experience
Programming Language
[Advanced]
C/C++, Python, Bash Script
[Intermediate]
Java
[Basic]
JavaScript
AI-Related Tools
[Advanced]
PyTorch
[Intermediate]
OpenCV, Numpy
[Basic]
Tensorflow, Keras,
Scikit-Learn, Pandas
Others
[Advanced]
Git, Regex
[Intermediate]
Docker, HTML, Android SDK, SQLite
[Basic]
Flask, RESTful API
Technical Sharing
