Taesun Yeom

I'm a second-year M.S.-Ph.D. student at EffL (Efficient Learning Lab), POSTECH (advisor: Prof. Jaeho Lee). For my research, I primarily focus on understanding various phenomena that arise in deep neural networks from a theoretical perspective.

My recent research interests include (1) Implicit biases and learning dynamics of neural networks, (2) Deep learning with infinite-dimensional functions (e.g., neural fields), and (3) Efficient deep learning.

Before joining EffL, I received my bachelor's degree in Mechanical Engineering from Chung-Ang University and worked closely with Prof. Minhyeok Lee and Prof. Seokwon Lee.

Email  /  GitHub  /  Google Scholar  /  LinkedIn  /  CV

Publications

Over-Alignment vs Over-Fitting: The Role of Feature Learning Strength in Generalization

We establish the existence of an optimal feature learning strength in classification and provide a theoretical explanation.

Activation Quantization of Vision Encoders Needs Prefixing Registers

We propose simple post-training quantization (PTQ) methods for ViTs that can be applied on top of existing approaches.

On the Internal Representations of Graph Metanetworks

What do graph metanetworks learn?

Fast Training of Sinusoidal Neural Fields via Scaling Initialization

We propose a simple yet effective method to accelerate neural field training.

DuDGAN: Improving Class-Conditional GANs via Dual-Diffusion

We propose a dual-diffusion process to mitigate overfitting in class-conditional GANs.

Superstargan: Generative adversarial networks for image-to-image translation in large-scale domains

This work enables GANs for image-to-image translation across large-scale domains.

Education

Design and source code from Jon Barron's website