Jiyeon Han

I’m Jiyeon Han, a researcher in AI, currently beginning my postdoctoral position in the GrUVi Lab at the School of Computing Science, Simon Fraser University, under the supervision of Prof. Hao (Richard) Zhang. I recently completed my Ph.D. at KAIST, Korea, advised by Prof. Jaesik Choi. My dissertation, “Computational Creativity in AI: Assessing and Enhancing the Creative Output of Generative Models,” reflects my core research interests in developing reliable methods for evaluating creativity and enhancing the creative capabilities of generative models.

Email  /  CV  /  Scholar  /  LinkedIn  /  Github

Research

I’m interested in building creative AI through generative models, with a broader focus on pushing the boundaries of what AI can achieve. I’m also deeply interested in understanding and analyzing the internal mechanisms of AI models to gain insights into how they learn, generate, and reason.

We present a training-free approach to enhance creative generation on Stable-Diffusion-based models. We achieve this by amplifying low-frequency features in the shallow blocks in the pretrained models.

This work explores a method for generating diverse rare samples that remain faithful to the original outputs of a pretrained GAN. By leveraging Normalizing Flows for density estimation, we enable end-to-end optimization to produce samples with lower data likelihood.

We present a new metric to assess rareness of an individual generated sample from pretrained generative models. Rarity score computes the (inverse) density of the sample estimated by k-NN based data manifold of real training data.

We investigate internal neurons that are closely associated with defective generations in generative neural networks. Our analysis reveals that neurons corresponding to artifact regions often exhibit abrupt local activation changes. Based on this observation, we propose the Local Activation Score to identify artifact-related neurons.

This work proposes a method for automatically identifying internal feature maps responsible for defective generations. By training an external artifact classifier and applying the explainable AI technique Grad-CAM, we highlight the feature maps most strongly associated with artifact regions. Building on this, we introduce a sequential correction algorithm that progressively refines the generation by suppressing artifact-related activations.

This work proposes a method for automatically identifying internal feature maps responsible for defective generations. By training an external artifact classifier and applying the explainable AI technique Grad-CAM, we highlight the feature maps most strongly associated with artifact regions. Building on this, we introduce a sequential correction algorithm that progressively refines the generation by suppressing artifact-related activations.

Miscellanea

This website is built upon Jon Barron's source code.