Sehyun Choi

ML Research Scientist @ TwelveLabs.

prof_pic.jpg

(Photo: HKUST)

Yongin-si, South Korea

I have a huge interest in Machine Learning (ML) and Artificial Intelligence (AI), and especially Natural Language Processing (NLP) and Multi-Modal AI, working with Language Models (LMs). Other orthogonal fields of interest include Trustworthy AI, such as eXplainable AI (XAI) and robustness. I am also interested in more engineering focused works, such as large-scale LLM pretraining engineering.


Research Questions:

Language Model +

  1. Inference-time optimization
  2. Knowledge Grounding
  3. compositional generalization
  4. World Model

In more details:

First topic is something I call “inference-time optimization”. Language Models have come a long way, and they are extremely strong nowadays. However, their behaviors are strongly dependent on the training dataset’s property. It is impossible to design dataset that is “perfect”, so they will always exhibit undesired attributes. I believe this should be addressed with inference-time adaptation method which includes controllable generation or constrained modeling for given constraints / desiderata (objectives).

Second, I am also interested in “Knowledge Grounding”. Hallucination, or confabulation (credit to Dr. Hinton), has become the hot potato in the era of generative AI. I believe reference-grounded generation is an important direction of solving this problem in the safe AI direction.

Third topic is reasoning ability, or more precisely, “Compositional generalization”. Dr. Chomsky states that human language has “infinite uses of finite means”, which suggests that humans compose finite set of functions to create infinite different possibilities. There are evidences that current SOTA AI (arguably LLMs) still may lack such ability.

The final topic is “World Model”, which is related to understanding commonsense about how the world works. I have a few publications about commonsense reasoning and building commonsense knowledge bases (CSKBs).


📄 Resume

(Last Update: Nov 16, 2023)


My another Passion: Photography

During my free time, I love doing street photography. You can check out my photos on my Instagram @sync.hpoto (More Anything) & @sync.hcut (More Street). I’m also big into analog film photography; which is exclusively showcased in my personal gallery.

news

Oct 29, 2024 Career Update I will be joining TwelveLabs as a full-time ML Research Scientist, working on Pegasus, a foundational Video-LM.
Apr 3, 2024 My joint work with NucleusAI on Efficient Pretraining of RetNet Architecture* has been released as a preprint!
Oct 7, 2023 🚨 NEW PAPER! 🚨 My first author paper about “Knowledge Constrained Decoding” has been accepted to EMNLP 2023 main conference!
Aug 26, 2022 :tada: Won 2nd place in Naver Clova AI Rush 2022, Unknown Document Detection Task! :tada:
May 11, 2022 Successfully finished my FYT project, “Explaining NLI with Feature Interaction Attribution”, in which I developed MAsk-based Feature Interaction Attribution (MAFIA), with an interactive demo.

selected publications

2024

  1. Cross-Architecture Transfer Learning for Linear-Cost Inference Transformers
    Sehyun Choi
    2024

2023

  1. EMNLP 2023
    KCTS: Knowledge-Constrained Tree Search Decoding with Token-Level Hallucination Detection
    Sehyun Choi, Tianqing Fang, Zhaowei Wang, and 1 more author
    EMNLP 2023, Oct 2023

2021

  1. EMNLP 2021
    Benchmarking Commonsense Knowledge Base Population with an Effective Evaluation Dataset
    Tianqing Fang, Weiqi Wang, Sehyun Choi, and 4 more authors
    EMNLP 2021, Sep 2021

experiences

Oct, 2024 ~ ML Research Scientist @ TwelveLabs, Building Pegasus, Video-LLM of TwelveLabs (Seoul, Full-time)
Sep, 2023 ~ Jul, 2024 Research Intern @ Nucleus AI, Foundational Language Model Pretraining (California, U.S., Remote)
Dec, 2021 ~ Feb, 2022 Research Intern @ Naver Corporation, Papago (Seongnam, South Korea)
Jul, 2021 ~ Aug, 2021 Research Intern @ SAI-Lab, KAIST, led by Prof. Jaesik Choi (Seongnam, South Korea)
Sep, 2020 ~ Dec, 2022 Research Intern (UROP) @ HKUST-KnowComp, led by Prof. Yangqiu Song. (Hong Kong)
Dec, 2019 ~ Jul, 2020 ML Engineer Intern @ Skelter Labs (Seoul, South Korea)

education

Aug 25, 2022 Master of Philosophy @ HKUST (2022~2024)
  • Highlights: CGA: 4.12/4.3
Jul 14, 2022 Bachelor’s Degree @ HKUST (2017~2022) [ Certificate | Transcript ]
  • Hightlights: GGA: 4.01/4.3; Academic Achievement Medal (Summa Cum Laude equivalent); First Class Honours; Dean’s List
Dec 23, 2016 High School Diploma @ Handong International School (2011~2016)