img/profile.jpg

T.M. Harry Hsu

Tzu-Ming Harry Hsu, Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology

Research Interests

Deep Learning

Computer Vision Machine Learning Signal Processing

Highlights

  • 1.5+ years of experience in deep learning model deployment for Bite! APP
  • 2 majors acquired in undergraduate education
  • Ranked #1 in International Physics Olympiad 2011 in both theory and experiment

Education

Massachusetts Institute of Technology
Ph.D. Student in Computer Science

Sep. 2017 - Now

National Taiwan University (NTU)
B.S.E. IN ELECTRICAL ENGINEERING AND B.S. IN PHYSICS

Sep. 2011 - Jun. 2016
Overall GPA
3.99 / 4.00
Last-60-unit GPA
3.99 / 4.00
Overall Class Rank
1 / 190

Publications

Conference Paper

  1. Transfer Neural Trees for Heterogeneous Domain Adaptation. Wei-Yu Chen, Tzu-Ming Harry Hsu, Yao-Hung Hubert Tsai, and Yu-Chiang Frank Wang, in ECCV 2016.
  2. Unsupervised Domain Adaptation With Imbalanced Cross-Domain Data. Tzu-Ming Harry Hsu, Wei-Yu Chen, Cheng-An Hou, Yao-Hung Hubert Tsai, Yi-Ren Yeh, and Yu-Chiang Frank Wang, in ICCV 2015.
  3. Connecting the dots without clues: Unsupervised domain adaptation for cross-domain visual classification. Wei-Yu Chen, Tzu-Ming Harry Hsu, Cheng-An Hou, Yi-Ren Yeh and Yu-Chiang Frank Wang, in ICIP 2015.
  4. Robust Motion Artifact Reduction of Photoplethysmographic Signal with Trajectory Space Circular Model. Tzu-Ming Harry Hsu, Wei-Yu Chen, Kuan-Lin Chen, Mong-Chi Ko, You-Cheng Liu, An-Yeu Andy Wu, in ICASSP Signal Processing Cup 2015.

Research Experiences

Signal Kinetics Lab, Media Lab, MIT
Research Assistant

Sep. 2017 - Now

Multimedia and Machine Learning Lab, Academia Sinica
INTERN STUDENT UNDER THE INSTRUCTION OF DR. YU-CHIANG FRANK WANG

Apr. 2014 - Jun. 2016
  • Deep Learning for Heterogeneous Domain Adaptation

    Transfer knowledge across different feature domains and build classifiers above the transferred knowledges.

    Transfer Neural Trees is proposed to transfer classifiers to a different dimensional space with deep neural network.

  • Unsupervised Domain Adaptation with Imbalanced Cross-domain Data

    Information of labeled source-domain data is transferred to the unlabeled target-domain, which may be a small set with imbalanced label counts.

    Closest Common Space Learning is proposed to combine sub-domain level classifiers to identify better source data applicability.

  • Unsupervised Domain Adaptation with Balanced Cross-domain Data

    A set of labeled source-domain data is used to construct classifier for the unlabeled target-domain data.

    An algorithm is proposed to address source-target mismatch and project them to a common space.

  • External Review

    Review papers as external reviewer for IEEE ICCV, IEEE ECCV, IEEE AAAI, and IEEE IJCAI.

Access IC Lab
INTERN STUDENT UNDER THE INSTRUCTION OF DR. AN-YEU ANDY WU

Sep. 2014 - Jun. 2015
  • Noise Removal of Photoplethysmographic (PPG) Signals

    Remove noises in PPG signals induced by motions by decorrelating the PPG with accelerometer signal.

    An algorithm is proposed to project the signal into a complex plane, in which a temporal filter will be performed, followed by ensemble voting for the optimal beat counts.

Laboratory for Applied Logic and Computation in System Design (ALCom Lab)
INTERN STUDENT UNDER THE INSTRUCTION OF DR. JIE-HONG ROLAND JIANG

Jul. 2013 - Jun. 2014
  • Compressed Sensing

    Compress the data perceived by a sensor array using less data storage than what it used to consume.

  • Mathematical Neural Models

    Establish a time-continuous model of human neurons to simulate the biological effects at stimulus and message passing.

Honors & Awards

Group

Altera Innovate Asia FPGA Design Competition
Silver Medal Award

2015

Ranked 2nd among 20 teams, team Taipei Amoeba designed a custom PCB named EZBud with algorithms integrated inside, which communicates with the FPGA. This piece of hardware modulates music according to measured user sporting statistics.

ICASSP Signal Processing Cup
Tenth Place

2015

Ranked 10th globally in sports heartbeat detection with an error of 4.89 beats per minute (BPM), team Taipei Amoeba had proposed an algorithm called Trajectory Space Circular Model.

Individual

Presidential Award (5 times)
Issued by the Department of Electrical Engineering

2011 - 2014

Awarded per semester to the top 5% students.

International Physics Olympiad (IPhO)
World’s First Place and Gold Medal Award

2011

Ranked 1st in both theory section and experiment section among 401 national representatives from senior high schools of over 80 countries.

International Junior Science Olympiad (IJSO)
Gold Medal Award

2008

Ranked top 10% among 300 national representatives from junior high schools of over 60 countries.

Work & Teaching Experiences

Digital Drift Corporation
COOPERATIVE RESEARCHER

Mar. 2016 - Now
Deep Neural Networks for Recognition and Matching

Build deep models for cuisine images using TensorFlow on multi-GPU machines, providing a backend with an API.

Olympiad Tutoring Community
PRIVATE TUTOR

Sep. 2011 - Jun. 2015

Offer tutoring for high school physics, competition physics, GRE subject test (physics), and SAT II subject test (physics).

Two students became national representatives for Taiwan in International Physics Olympiad (IPhO).