Hoyong Kim

Seoul National University

Seoul, Republic of Korea

hoyong.kim@snu.ac.kr

hoyong.kim.96@gmail.com

I am a postdoctoral researcher at the Visual & Geometric Intelligence (VGI) Lab at Seoul National University, working with Prof. Jaesik Park. I received my Ph.D. in Artificial Intelligence from GIST in 2026, advised by Prof. Kangil Kim at the IRR Lab.

My research asks how deep networks organize information in their representation space, and how that geometry can be understood, diagnosed, and shaped. My approach follows three steps: identify a model’s failure from the perspective of representation learning, verify it theoretically and empirically, and propose a theory-grounded remedy.

During my Ph.D., I studied Neural Collapse and the collapse phenomena that degrade representations when idealized training assumptions break down. This line of work includes angle-only representations with the Spherization Layer (NeurIPS 2022), zero-mean Neural Collapse with a fixed non-negative orthogonal classifier (ICLR 2024), and minority collapse caused by imbalanced mixed labels (NeurIPS 2026). With colleagues at the IRR Lab, I have also studied disentangled and object-centric representations (ICML 2026, TMLR 2026).

I am now extending this geometric view toward embodied AI. My current interests are:

Mechanistic interpretability of Vision-Language-Action (VLA) models, using tools such as sparse autoencoders to study latent space geometry
World models and World Action Models for robotics, with a focus on 3D geometry-aware representations
Representation learning theory, including Neural Collapse and how representations align across modalities

See my publications, research, and CV for details.

news

Sep 24, 2026	Our paper Does Mixed Label Imbalance Matter to Minority Collapse in Imbalanced Learning? was accepted to NeurIPS 2026! See you in Sydney.
Jun 30, 2026	Our paper on representational consistency of identical transformations for disentangled representation was accepted to TMLR.
May 17, 2026	I joined the VGI Lab at Seoul National University as a postdoctoral researcher, working with Prof. Jaesik Park.
May 13, 2026	I was selected as a Gold Reviewer for ICML 2026.
May 10, 2026	I joined the DGIST InnoCORE Research Group for Bio-Embodied Physical AI.
Apr 30, 2026	Two papers with the IRR Lab, Factor-Wise Homogeneity of Slot-Attention and the Orthogonal Subspaces Projection layer, were accepted to ICML 2026.
Feb 19, 2026	I received my Ph.D. in Artificial Intelligence from GIST. Dissertation: From Theory to Remedy: Addressing Representation Collapse for Enhanced Class-Discriminative Clarity.
Jun 30, 2024	Two papers on Asymptotic Midpoint Mixup and Softmax Masking for continual learning were accepted to the CoLLAs 2024 Workshop Track.
Jan 15, 2024	Our paper on the Fixed Non-negative Orthogonal Classifier (FNOC) was accepted to ICLR 2024!
Sep 14, 2022	Our paper Spherization Layer: Representation Using Only Angles was accepted to NeurIPS 2022!

selected publications

NeurIPS
Does Mixed Label Imbalance Matter to Minority Collapse in Imbalanced Learning?

Hoyong Kim and Kangil Kim

In Advances in Neural Information Processing Systems (NeurIPS), 2026

Abs Bib HTML Website

Minority collapse, where minority classes become indistinguishable, is a significant challenge in imbalanced learning. This challenge is addressed by methods such as Mixup with class-balanced sampling. Although minority collapse has been mathematically analyzed using the layer-peeled model alongside Neural Collapse, no prior work has analyzed minority collapse under Mixup, particularly from the perspective of mixed labels. We investigate this overlooked factor and raise the question: Is mixed label balance important for alleviating minority collapse? Our analysis reveals that (i) mixed labels should be balanced, and (ii) in this setting, interpreting mixed labels as singletons is beneficial. Motivated by this analysis, we propose a Balanced Mixed Label Sampler and a Mixed-Singleton classifier, which balance mixed labels and treat them as singleton labels. Through theoretical analysis and experimental results, we highlight the importance of balancing mixed labels in imbalanced learning.
@inproceedings{kim2026mixedlabel, title = {Does Mixed Label Imbalance Matter to Minority Collapse in Imbalanced Learning?}, author = {Kim, Hoyong and Kim, Kangil}, booktitle = {Advances in Neural Information Processing Systems (NeurIPS)}, year = {2026}, }
ICLR
Fixed Non-negative Orthogonal Classifier: Inducing Zero-mean Neural Collapse with Feature Dimension Separation

Hoyong Kim and Kangil Kim

In International Conference on Learning Representations (ICLR), 2024

Abs Bib HTML PDF Video Slides

Fixed classifiers in deep networks induce neural collapse toward a simplex equiangular tight frame. We study what happens when the fixed classifier is not a simplex and introduce the fixed non-negative orthogonal classifier, which achieves global optimality in the layer-peeled model while inducing zero-mean neural collapse and feature dimension separation, with benefits for continual and imbalanced learning.
@inproceedings{kim2024fixed, title = {Fixed Non-negative Orthogonal Classifier: Inducing Zero-mean Neural Collapse with Feature Dimension Separation}, author = {Kim, Hoyong and Kim, Kangil}, booktitle = {International Conference on Learning Representations (ICLR)}, year = {2024}, }
NeurIPS
Spherization Layer: Representation Using Only Angles

Hoyong Kim and Kangil Kim

In Advances in Neural Information Processing Systems (NeurIPS), 2022

Abs Bib HTML PDF Video Slides

We propose the spherization layer, which represents features using only angles on a hypersphere so that the norm carries no information. The layer preserves the expressiveness of conventional layers while making angular similarity the sole carrier of information, and we show its benefits across vision and language tasks.
@inproceedings{kim2022spherization, title = {Spherization Layer: Representation Using Only Angles}, author = {Kim, Hoyong and Kim, Kangil}, booktitle = {Advances in Neural Information Processing Systems (NeurIPS)}, volume = {35}, pages = {16398--16410}, year = {2022}, }