Granular Concept Circuits: Toward a Fine-Grained Circuit Discovery for Concept Representations
Dahee Kwon, Sehyun Lee and Jaesik Choi
Interpretability, XAI, Concept Discovery, Computer Vision
We introduce an automatic circuit discovery method in which each circuit represents a concept relevant to a specific query.
Enhancing Creative Generation on Stable Diffusion-based Models
Jiyeon Han, Dahee Kwon and Jaesik Choi
Creative Generation, Representation Learning, Text-to-Image Generation, Diffusion
C3 is a training-free approach designed to improve creative generation in Stable Diffusion-based models.
Understanding Distributed Representations of Concepts in Deep Neural Networks without Supervision
Wonjoon Chang, Dahee Kwon and Jaesik Choi
Oral presentation | Image Classification, Representation Learning, Explainable AI
RDR identifies visual concept representations learned by image classifiers without requiring supervision.
Breaking the Lock-in: Diversifying Text-to-Image Generation via Representation Modulation
Dahee Kwon, Haeun Lee and Jaesik Choi
Generative AI, Representation Learning, Diverse Image Generation, Computer Vision
DAVE is a training-free intervention that mitigates lock-in behavior and improves sample diversity.
Causal Path Tracing in Transformers
Won Jo, Dahee Kwon, Jongeun Baek, Cheongwoong Kang and Jaesik Choi
Causal Path Discovery, Actual Cause, LLM, Transformers
We propose a framework to trace how information causally flows through transformer internals for a given decision.
SPICE: Simple Polysemantic Feature Interpretation via Clustering-based Explanations
Sehyun Lee, Dahee Kwon, Damin Lee and Jaesik Choi
Polysemanticity, Representation Learning, Interpretability, Vision
We study how to disentangle visual concept clusters across diverse recognition architectures and analyze polysemanticity systematically.
MAC: Memory-Augmented Contrastive Learning for Time Series Anomaly Detection
Dahee Kwon, Enver Menadjiev, Qin Xie and Jaesik Choi
Unsupervised Anomaly Detection, Contrastive Learning, Memory Augmentation, Multivariate Time-Series
MAC improves anomaly detection with dynamic memory augmentation, contrastive learning, and end-to-end reconstruction.
Dual Masking for Domain Generalization
Dahee Kwon and Jaesik Choi
Domain Generalization, Representation Learning, Image Classification
DMDG improves robust representation learning by attenuating non-generalizable domain-specific features.