Skip to content

Jinho Choi

Ph.D. Candidate, KAIST

Visiting Researcher, Helmholtz Munich

Munich, Germany

Open to research collaborations and internships.

I am a Ph.D. candidate at KAIST and a visiting researcher at Helmholtz Munich. My research focuses on mechanistic interpretability and data-centric AI to advance safe and trustworthy AI.

My Ph.D. research uses sparse dictionary learning to uncover human-interpretable concepts in foundation models. My work includes interpreting vision models (PatchSAE), uncovering dataset biases (ConceptScope), and analyzing vision-language models (VisualScratchpad). More recently, I developed ConSDL to improve the consistency of learned concepts across training runs.

I also spent five years in industry as an AI scientist, developing biomedical imaging models at Tomocube and AI interview systems and conversational agents at Genesislab.

During my M.S. at Korea University, I worked on visual analytics and human-computer interaction (HCI).

Recent

  • Oct 2026 Released ConSDL, a sparse coding approach for consistent concept discovery in foundation models.
  • Apr 2026 VisualScratchpad presented at the ICLR 2026 Workshop on Trustworthy AI.
  • Feb 2026 Joined the Dynamical Inference Lab at Helmholtz Munich as a visiting researcher.
  • Sep 2025 ConceptScope accepted to NeurIPS 2025.
  • Jan 2025 PatchSAE accepted to ICLR 2025.

Research

Mechanistic interpretability

Discovering interpretable and reproducible concepts in foundation models through sparse dictionary learning.

Trustworthy AI

Understanding and mitigating model biases through interpretable representations and fairness-aware learning.

Data-centric AI

Understanding and improving datasets through concept-based analysis and human-in-the-loop annotation.

Publications

Selected publications

  1. ConSDL

    Under review

    Beyond SAEs: Consistent Concept Discovery with Sparse Coding

    Jinho Choi, Hyesu Lim, Jaegul Choo, Steffen Schneider

    Uses sparse coding to learn more consistent and reproducible concepts across independent training runs.

  2. ICLR 2026 Workshop on Trustworthy AI

    VisualScratchpad: Grounding Visual Concepts in Large Vision Language Models

    Hyesu Lim, Jinho Choi, Taekyung Kim, Byeongho Heo, Jaegul Choo, Dongyoon Han

    Enables concept-level interpretation of large vision-language models during inference.

  3. NeurIPS 2025

    ConceptScope: Characterizing Dataset Bias via Disentangled Visual Concepts

    Jinho Choi, Hyesu Lim, Steffen Schneider, Jaegul Choo

    Analyzes image datasets through interpretable visual concepts to uncover previously overlooked biases.

  4. ICLR 2025

    Sparse Autoencoders Reveal Selective Remapping of Visual Concepts During Adaptation

    Hyesu Lim, Jinho Choi, Jaegul Choo, Steffen Schneider

    Reveals that adapting vision models primarily remaps existing visual concepts rather than learning new ones.

  5. WACV 2024

    Slice and Conquer: A Planar-to-3D Framework for Efficient Interactive Segmentation of Volumetric Images

    Wonwoo Cho, Dongmin Choi, Hyesu Lim, Jinho Choi, Saemee Choi, Hyun-seok Min, Sungbin Lim, Jaegul Choo

    Enables efficient 3D image annotation by propagating sparse 2D labels and selectively requesting corrections.

  6. IEEE Access 2023

    Fairness-aware Multimodal Learning in Automatic Video Interview Assessment

    Changwoo Kim, Jinho Choi, Jongyeon Yoon, Daehun Yoo, Woojin Lee

    Reduces demographic bias in AI video-interview assessment while preserving predictive accuracy.

  7. bioRxiv 2021

    Label-free Three-dimensional Analyses of Live Cells with Deep-learning-based Segmentation Exploiting Refractive Index Distributions

    Jinho Choi, Hye-Jin Kim, et al.

    Enables 3D segmentation of live cells and cellular structures from label-free refractive-index images.

  8. SPIE 2021

    3D Cell Instance Segmentation via Point Proposals using Cellular Components

    Jinho Choi, Junwoo Park, Hyun-seok Min, Hyungjoo Cho, Sungbin Lim, Jaegul Choo

    Separates individual cells in 3D microscopy images using point proposals derived from cellular components.

  9. Computer Graphics Forum (EuroVis) 2019

    Visualizing for the Non-Visual: Enabling the Visually Impaired to Use Visualization

    Jinho Choi, Sanghun Jung, Deokgun Park, Jaegul Choo, Niklas Elmqvist

    Makes chart information accessible to people with visual impairments through data extraction and interactive exploration.

  10. CHI 2018

    TopicOnTiles: Tile-based Spatio-Temporal Event Analytics via Exclusive Topic Modeling on Social Media

    Minsuk Choi, Sungbok Shin, Jinho Choi, et al.

    Supports exploration of local events in social media through spatio-temporal topic modeling and visual analytics.

  11. IEEE ICDM 2017

    STExNMF: Spatio-Temporally Exclusive Topic Discovery for Anomalous Event Detection

    Sungbok Shin, Minsuk Choi, Jinho Choi, et al.

    Discovers location- and time-specific topics to detect anomalous events in social media.

Experience

Feb 2026 — Present

Munich, Germany

Helmholtz Munich

Visiting Researcher · Dynamical Inference Lab

  • Developed ConSDL, a sparse coding approach that improves the consistency of learned concepts across independent training runs. Demonstrated improved dictionary and coefficient consistency over six sparse autoencoder baselines on vision and language models ConSDL.

Aug 2022 — Present

Daejeon, Korea

KAIST

Ph.D. Researcher · DAVIAN Lab

  • Developed ConceptScope to characterize image datasets through interpretable visual concepts, revealing previously overlooked dataset biases ConceptScope.
  • Applied sparse autoencoders to CLIP (PatchSAE) to study visual concept representations and their changes during task adaptation PatchSAE. Extended concept-level analysis to large vision-language models with VisualScratchpad VisualScratchpad.

Sep 2022 — Nov 2024

Seoul, Korea

Genesislab

AI Scientist

  • Developed a fairness diagnosis and mitigation pipeline for an AI video-interview platform used by 150+ companies, reducing gender bias by 35% without sacrificing accuracy Fair AVI.
  • Developed a talking-face generation model deployed as an AI interviewer avatar, with accurate lip synchronization and high visual quality.
  • Built LLM-based conversational agents for a creator-persona app with 500K+ downloads, using retrieval-augmented generation, memory, and automated persona construction.

Feb 2019 — Sep 2022

Seoul, Korea

Tomocube

AI Researcher

  • Developed 3D cell and organelle segmentation models for label-free holotomography, integrated into commercial cell analysis software 3D Cell Inst. Label-free 3D.
  • Built a human-in-the-loop 3D annotation system that propagates 2D labels into 3D and requests corrections on uncertain slices, and deployed it as the company's internal annotation system Slice & Conquer.

Mar 2017 — Aug 2019

Seoul, Korea

Korea University

M.S. Researcher

  • Developed a chart accessibility tool that extracts data from web-based visualizations and makes it accessible to visually impaired users through screen readers Non-Visual Vis.
  • Developed visual analytics systems for detecting anomalous local events in geo-tagged social media using topic modeling and interactive map visualizations STExNMF TopicOnTiles.

Education

  • 2022 — 2028

    Ph.D. in Artificial Intelligence

    KAIST · Advisor: Jaegul Choo · expected Feb 2028 · part-time 2022–2024

  • 2017 — 2019

    M.S. in Computer Science and Engineering

    Korea University · Advisor: Jaegul Choo

  • 2013 — 2017

    B.S. in Computer Science and Engineering

    Korea University

Service

  • Reviewer · NeurIPS, ICLR