Minkyu Jeon

Hello! I'm a Ph.D. student in Computer Science at Princeton University under the guidance of Prof. Ellen D. Zhong.

My main research focuses on 3D reconstruction, generative models, and their applications to inverse problems encountered in computer vision, with a particular emphasis on scientific imaging modalities such as cryo-EM. Furthermore, I am interested in understanding and developing methods that address the limitations faced by existing machine learning methods in generalizing beyond their training distributions.

Email  /  CV  /  Google Scholar  /  Twitter  /  LinkedIn

Activities

• 09-2024:  Our work on cryo-EM heterogeneous reconstruction datasets, benchmarks, and metrics has been accepted to NeurIPS2024 (spotlight).
• 08-2024:  I present my CryoBench paper at Flatiron Institute.
• 07-2024:  Instructed high school students at AI4ALL Princeton.
• 02-2024:  I gave an invited lecture on k-SALSA at CMU MetaMobility Lab.
• 09-2023:  Starting my Ph.D journey at Princeton!
• 01-2023:  Starting an Associate Computational Biologist role at the Broad Institute of MIT and Harvard.
• 11-2022:  Our work on self-supervised representation learning for localization task has been accepted to Information Science Journal 2023.
• 09-2022:  Our work on saliency-guided point cloud data mixup has been accepted to NeurIPS2022.
• 07-2022:  Our work on GAN-based approach to preserve the privacy of retina images has been accepted to ECCV2022.
• 09-2021:  Starting as a visiting graduate student at the Broad Institute of MIT and Harvard.

Research

*: co-first author

We propose CryoBench, a suite of datasets, metrics, and performance benchmarks for heterogeneous reconstruction in cryo-EM.

We propose a simple and effective self-supervised learning method that uses the positional relationships between the entities in an image by shuffling the convolution kernels.

We propose SageMix, a saliency-guided Mixup for point clouds to preserve salient local structures. We extract salient regions from two point clouds and smoothly combine them into one continuous shape.

We presented an end-to-end pipeline for synthesizing a k-anonymous retinal image dataset given a private input dataset. We leverage local style alignment, our new approach for summarizing source images in a cluster while preserving local texture information.

We propose a general framework that learns to adaptively train each layer of deep neural networks via meta-learning. Our framework leverages the local error signals from layers and identifies which layer needs to be trained more at every iteration.

Teaching

Credit: Jon Baron.