Seong Joon Oh

I'm a Senior Scientist in Machine Learning at the University of Tübingen currently working on Trustworthy AI. I previously did research at Naver AI Lab and Max Planck Institute for Informatics, working with Mario Fritz and Bernt Schiele. My work focuses on building reliable and responsible AI systems.

My research interests span multiple aspects of Trustworthy AI, including uncertainty quantification, out-of-distribution detection, explainability/interpretability, and privacy/security of machine learning systems. I aim to understand the key limitations of current AI systems and develop systematic solutions to make them more reliable and trustworthy. I've published extensively at top computer vision and machine learning venues like CVPR, ICCV, NeurIPS, and ICLR.

In recent years, I've been particularly interested in uncertainty quantification and probabilistic deep learning. I develop methods to make neural networks aware of their own limitations and communicate them effectively to users. I also work on training data attribution and understanding model behaviors under distribution shifts. Through my research, I hope to contribute to making AI systems that we can confidently deploy in real-world applications.

Publications

CLIP Behaves like a Bag-of-Words Model Cross-modally but not Uni-modally

Darina Koishigarina, Arnas Uselis, Seong Joon Oh

ABS HTML PDF

Scaling Up Membership Inference: When and How Attacks Succeed on Large Language Models

Haritz Puerto, Martin Gubri, Sangdoo Yun, Seong Joon Oh

arXiv.org 2024

ABS HTML PDF

Overcoming Domain Limitations in Open-vocabulary Segmentation

Dongjun Hwang, Seong Joon Oh, Junsuk Choe

arXiv.org 2024

ABS HTML PDF

Scalable Ensemble Diversification for OOD Generalization and Detection

Alexander Rubinstein, Luca Scimeca, Damien Teney, Seong Joon Oh

arXiv.org 2024

ABS HTML PDF

Towards User-Focused Research in Training Data Attribution for Human-Centered Explainable AI

Elisa Nguyen, Johannes Bertram, Evgenii Kortukov, Jean Y. Song, Seong Joon Oh

arXiv.org 2024

ABS HTML PDF

Studying Large Language Model Behaviors Under Context-Memory Conflicts With Real Documents

Evgenii Kortukov, Alexander Rubinstein, Elisa Nguyen, Seong Joon Oh

Seong Joon Oh

Publications

CLIP Behaves like a Bag-of-Words Model Cross-modally but not Uni-modally

Scaling Up Membership Inference: When and How Attacks Succeed on Large Language Models

Overcoming Domain Limitations in Open-vocabulary Segmentation

Scalable Ensemble Diversification for OOD Generalization and Detection

Towards User-Focused Research in Training Data Attribution for Human-Centered Explainable AI

Studying Large Language Model Behaviors Under Context-Memory Conflicts With Real Documents

Do Deep Neural Network Solutions Form a Star Domain?

Calibrating Large Language Models Using Their Generations Only

Benchmarking Uncertainty Disentanglement: Specialized Uncertainties for Specialized Tasks

Pretrained Visual Uncertainties

TRAP: Targeted Random Adversarial Prompt Honeypot for Black-Box Identification

Mitigating Shortcut Learning with Diffusion Counterfactuals and Diverse Ensembles

Exploring Practitioner Perspectives On Training Data Attribution Explanations

Trustworthy Machine Learning

URL: A Representation Learning Benchmark for Transferable Uncertainty Estimates

ProPILE: Probing Privacy Leakage in Large Language Models

A Bayesian Approach To Analysing Training Data Attribution In Deep Learning

Playing repeated games with Large Language Models

Neglected Free Lunch – Learning Image Classifiers Using Annotation Byproducts

Probabilistic Contrastive Learning Recovers the Correct Aleatoric Uncertainty of Ambiguous Inputs

SelecMix: Debiased Learning by Contradicting-pair Sampling

Scratching Visual Transformer's Back with Uniform Attention

ID and OOD Performance Are Sometimes Inversely Correlated on Real-world Datasets

Dataset Condensation via Efficient Synthetic-Data Parameterization

ECCV Caption: Correcting False Negatives by Collecting Machine-and-Human-verified Image-Caption Associations for MS-COCO

Weakly Supervised Semantic Segmentation using Out-of-Distribution Data

ALP: Data Augmentation using Lexicalized PCFGs for Few-Shot Text Classification

Which Shortcut Cues Will DNNs Choose? A Study from the Parameter-Space Perspective

Keep CALM and Improve Visual Feature Attribution

Neural Hybrid Automata: Learning Dynamics with Multiple Modes and Stochastic Transitions

Rethinking Spatial Dimensions of Vision Transformers

Region-based dropout with attention prior for weakly supervised object localization

Re-labeling ImageNet: from Single to Multi-Labels, from Global to Localized Labels

Probabilistic Embeddings for Cross-Modal Retrieval

VideoMix: Rethinking Data Augmentation for Video Classification

Evaluation for Weakly Supervised Object Localization: Protocol, Metrics, and Datasets

AdamP: Slowing Down the Slowdown for Momentum Optimizers on Scale-invariant Weights

Slowing Down the Weight Norm Increase in Momentum-based Optimizers

An Empirical Evaluation on Robustness and Uncertainty of Regularization Methods

Reliable Fidelity and Diversity Metrics for Generative Models

Evaluating Weakly Supervised Object Localization Methods Right

On Recognizing Texts of Arbitrary Shapes with 2D Self-Attention

Learning De-biased Representations with Biased Representations

CutMix: Regularization Strategy to Train Strong Classifiers With Localizable Features

What Is Wrong With Scene Text Recognition Model Comparisons? Dataset and Model Analysis

Modeling Uncertainty with Hedged Instance Embedding

Understanding and Controlling User Linkability in Decentralized Learning

Gradient-Leaks: Understanding and Controlling Deanonymization in Federated Learning

Sequential Attacks on Agents for Long-Term Adversarial Goals

Natural and Effective Obfuscation by Head Inpainting

Whitening Black-Box Neural Networks

Towards Reverse-Engineering Black-Box Neural Networks

Person Recognition in Personal Photo Collections

Generating Descriptions with Grounded and Co-referenced People

Adversarial Image Perturbation for Privacy Protection A Game Theory Perspective

Exploiting Saliency for Object Segmentation from Image Level Labels

Faceless Person Recognition: Privacy Implications in Social Media

I-Pic: A Platform for Privacy-Compliant Image Capture

Person Recognition in Personal Photo Collections

CYP1A2 activity as a risk factor for bladder cancer.

Studying Large Language Model Behaviors Under Realistic Knowledge Conflicts

SelecMix: Debiased Learning by Mixing up Contradicting Pairs

Gradient-Leaks: Understanding Deanonymization in Federated Learning

Modeling Uncertainty with Hedged Instance Embeddings

Image manipulation against learned models: privacy and security implications

From Understanding to Controlling Privacy against Automatic Person Recognition in Social Media

Exercise 3 : Deep Neural Networks and Backpropagation

Demo: I-Pic: A Platform for Privacy-Compliant Image Capture