Try NeuralDivergence (PacificVis'19, Poster #1008)



Haekyu Park (박해규)

Ph.D. Student in Computer Science at Georgia Tech
Advisor: Dr. Polo Chau
haekyu@gatech.edu ・ Curriculum Vitae
Research Interests
  • Machine Learning
  • Interpretability of Machine Learning
  • Information / Data Visualization

News


Last updated: Feb 19, 2019

Education


Georgia Institute of Technology

Ph.D., Computer Science
Advisor: Dr. Polo Chau
Aug 2018 - Present

Courses: Information Visualization, Machine Learning, Computer Vision




Seoul National University

B.S., Computer Science and Engineering
Graduated with honors (Cum Laude)
Mar 2012 - Aug 2017

Courses: Computer Vision, Introduction to Data Mining, Probabilistic Graphical Models, Software Engineering, Creative Integrated Design, System Programming, Operating Systems, Data Communications, Algorithms, Programming Language, Principles of Programming, IT Leadership Seminar, Computer Engineering Seminar, Electrical and Electronic Circuits, Computer Programming, Data Structures, Programming Practice, Discrete Mathematics, Logic Design, Logic Design Lab., Digital Computer Concept and Practice

Publication

2019

Poster [1]
NeuralDivergence: Exploring and Understanding Neural Networks by Comparing Activation Distributions
Haekyu Park Fred Hohman, and Polo Chau
IEEE Pacific Visualization Symposium (PacificVis), 2019, Bangkok, Thailand
[PDF] [Website] [Details]

As deep neural networks are increasingly used in solving high-stake problems, there is a pressing need to understand their internal decision mechanisms. Visualization has helped address this problem by assisting with interpreting complex deep neural networks. However, current tools often support only single data instances, or visualize layers in isolation. We present NeuralDivergence, an interactive visualization system that uses activation distributions as a high-level summary of what a model has learned. NeuralDivergence enables users to interactively summarize and compare activation distributions across layers, classes, and instances (e.g., pairs of adversarial attacked and benign images), helping them gain better understanding of neural network models.

2018

Conference [2]
SIDE: Representation Learning in Signed Directed Networks
Junghwan Kim, Haekyu Park, Ji-Eun Lee, and U Kang
The Web Conference (Previously known as WWW, World Wide Web Conference) 2018, Lyon, France
[PDF] [Website] [Details]

We propose SIDE, a network embedding algorithm for signed directed networks. Network embedding learns a mapping of each node to a vector. SIDE carefully formulates and optimizes likelihood over both direct and indirect signed connections. We provide socio-psychological interpretation for each component of likelihood function and prove linear scalability of our algorithm.

2017

Conference [1]
A Comparative Study of Matrix Factorization and Random Walk with Restart in Recommender Systems
Haekyu Park, Jinhong Jung, and U Kang
IEEE International Conference on Big Data (BigData) 2017, Boston, MA, USA
[PDF] [Website] [Slides] [Details]
  • Main Question:
    Between matrix factorization (MF) or Random Walk with Restart (RWR), which method works better for recommender systems?
  • Specific tasks:
    We compare MF and RWR for the following recommendation scenarios.
    • Which method performs better when using explicit feedback data?
      • → MF is better.
    • Which method performs better when using implicit feedback data?
      • → RWR is better.
    • Do global bias terms improve performance?
      • → Yes.
    • Which method performs better when exploiting global bias terms?
      • → MF is better.
    • Does side information enhance performance?
      • → Yes with explicit ratings, and No with implicit ratings.
    • Which method performs better when employing side information?
      • → MF is better with explicit rating data, and RWR is better with implicit rating data.
    • Which method solves the cold start problem better when employing side data?
      • → MF is better with explicit rating data, and RWR is better with implicit rating data.
  • Discussion:
    What are the reasons for good or bad performance of MF and RWR in various settings of recommendations? Details are in Section 4.G in our paper. Please read that part!

Project

2017

Recommender System for Videos on Oksusu Application

  • With SK Telecom, Seoul, Republic of Korea
  • Keywords: Tensorflow, Feedforward Neural Network, Sequence Embedding, Word Embedding, Approx. k-NN
  • [Details]
    We recommend videos to users of Oksusu application. Our model uses deep learning to achieve quality recommendation. Our recommendation engine handles massive data on users' behaviors and heterogeneous information on videos.

Aug - Dec, 2017
A Fast and Cost Efficient Data Compression Algorithm with Shared Virtual Memory in Heterogeneous System Architecture

  • For my undergraduate graduation essay
  • Keywords: OpenCL, GPGPU, SVM, HSA
  • [Details]
    • I used GPGPU and Shared Virtual Memory (SVM) in Heterogeneous System Architecture (HSA).
    • I acquired parallel programming thinking and skills, especially with C++ and OpenCL.
    • We propose fast data deduplication methods that use general purpose computing on graphics processing units (GPGPU) and heterogeneous system architecture (HSA) to efficiently execute the great amount of computations. GPGPU and HSA provide a powerful basis for parallel computing in an easy programmable and efficient way.

Nov 2016 - Aug 2017

2016

Personalized Recommendation for Credit Card Rewards

  • With Hyundai Card, Seoul, Republic of Korea
  • Keywords: Coupled Matrix Factorization, Time Series Data
  • [Details]
    We provide personalized recommendations for credit card rewards to customers. The main algorithm is TCMF (Time Coupled Matrix Factorization). We use various side information of users and items and highly improve the recommendation performance.
  • [News article (in Korean)]

June - Dec, 2016
Social Recommender System with Graph and Rating Information

  • For final project of Probabilistic Graphical Model course
  • Keywords: Matrix Factorization, Network Embedding, Social Network, PGM
  • [Details]
    We designed and implemented a novel recommendation model leveraging social network by network embedding.

Sep - Dec, 2016

Awards & Honors

  • National Scholarship for Science and Engineering
    Merit-based, 2015

Research Experience

  • Graduate Research Assistant
    • Aug 2018 - Present
    • Georgia Institute of Technology
    • Advisor: Dr. Polo Chau

  • Undergraduate Research Assistant
    • June 2016 - Aug 2017
    • Seoul National University

Technical Skills

  • Programming Languages
    • Python, JavaScript, HTML, R, Matlab, Java, C, C++, Ocaml, Scheme

  • Machine Learning
    • TensorFlow, Keras, Numpy, OpenCV, SciPy, scikit-learn

  • Data Visualization
    • D3.js, Matplotlib, ggplot

  • Parallel Computing
    • OpenCL

Patents

  • Apparatus and Method for Representation Learning in Signed Directed Networks
    U Kang, Junghwan Kim, and Haekyu Park, Korean Patent 10-2017-0130914, 2017.

Getting in Touch

I would like to talk to you! Please let me know you more, through
  • Email: haekyu@gatech.edu
  • Address: Klaus Advanced Computing Building 1305, Georgia Institute of Technology, 266 Ferst Dr NW, Atlanta, GA, 30332