Jayeon Yoo

I am currently a Ph.D. candidate majoring in Intelligence Information Convergence Studies at Seoul National University. I am advised by Prof. Nojun Kwak, as a member of Machine Intelligence and Pattern Analysis Lab (MIPAL). Prior to my graduate studies, I earned my Bachelor's degree in Industrial and Management Engineering from POSTECH.

Email  /  CV  /  Scholar  /  Github

Research

I am deeply passionate about advancing deep learning models in demanding real-world scenarios, distinct from conventional training environments. With a solid background in Unsupervised Domain Adaptation, Test-Time Adaptation, and Open-set tasks, I tackle key challenges such as data distribution variations and exploiting unlabeled data. My goal is to improve the adaptability and versatility of deep learning models, ensuring they can effectively tackle new and evolving tasks with efficiency.

Leveraging an adaptor module and an adaptation skipping strategy, our approach enables efficient object detector adaptation to dynamic test environments, improving detection performance, preventing catastrophic forgetting and maintaining high efficiency.

Enhancing open-vocabulary object detection through the proxy of novel classes via a convex combination in the CLIP embedding space.

By hypothesizing that suppressing the Hessian enhances gradient-based meta-learning, an algorithm is introduced that improves few-shot learning performance with minimal added computational complexity, demonstrating broad applicability and efficacy.

By introducing refined gradients for pruned weights, our Dynamic Collective Intelligence Learning approach significantly enhances performance and training stability in deep neural networks at high sparsity levels.

Transferring inter-pixel correlations from source to target domain using a self-attention module, which captures domain-invariant properties, significantly improves adaptation for semantic segmentation, notably enhancing performance for rare or small-region classes.

By analyzing a unique characteristic of object detection—features distributed by both object class and regression values—we effectively adapt object detectors to the target domain, taking into account these regression values to mitigate negative transfer and improve adaptation performance.

Leveraging temporal information to discern frames with minimal change, a novel dynamic network is proposed for Semi-supervised Video Object Segmentation, significantly boosting inference speed with negligible accuracy loss by selectively bypassing intensive mask generation for static or slow-moving scenes across various datasets.