I am a PhD student in UC Berkeley Vision Science program, on computational vision track, advised by Prof. Stella Yu. Previously I was a research intern at Sensetime. I received my B.E in Electronic Engineering from Xi'an Jiaotong University in 2018, where I worked with Prof. Jinjun Wang at Institute of Artificial Intelligence and Robotics.
New paper! A light-cost regularizer orthogonal convolution that reduces the feature redundancy and improves CNN performance and robustness under attack. Code to release very soon!
Orthogonal convolutional neural networks is a light-cost regularizer which reduces the feature redundancy and improves network performance and robustness under attack.
This work develops a novel deep learning architecture for naturally complex-valued data, with improved results and only 10% of the parameters as the baseline model.
A Deep Learning Approach for Meibomian Gland Atrophy Evaluation in Meibography Images
Jiayun Wang, Thao N. Yeh, Rudrasis Chakraborty, Stella X. Yu, Meng C. Lin.
Translational Vision Science and Technology 2019. [paper] [code]
This work develops a deep learning approach to digitally segmenting meibomian gland atrophy area and computing percent atrophy in meibography images.
This work formally defines the problem of open long-tailed recognition (OLTR) as optimizing for the overall accuracy of a naturally-distributed dataset with the presence of open classes.
Insights and Approaches Using Deep Learning to Classify Wildlife
Zhongqi Miao, Kaitlyn M Gaynor, Jiayun Wang, Ziwei Liu, Oliver Muellerklein, Mohammad S. Norouzzadeh, Alex McInturff, Rauri C. K. Bowie, Ran Nathon, Stella X. Yu, Wayne M. Getz.
Nature - Scientific Reports 2019. [Paper]
This work aims to interpret the concepts behind the convolutional neural networks (CNNs) in classifying wildlife.
Successive Embedding and Classification Loss for Aerial Image Classification
Jiayun Wang, Patrick Virtue, Stella X. Yu.
arXiv 2017. [paper] [code]
This work aims to address the overfitting problem in remote sensing image classification.
This work designs a dynamically adaptive loss function to overcome the drawbacks of conventional loss functions for person re-identification.
We use a hierarchical approach for long-term video prediction, aiming at estimating high-level structure in the input frame first, then predicting how that structure grows in the future.
Perceptual Metric for Realistic
Jiayun Wang, Linda F. Liu.
VS 260A (Optical and Neural Limits to Vision) Final Project. [slides]
We use a nerual network based perceptual metric for realistic DiffuserCam imaging.
Our robotic system (Baxter) is able to recognize a number unique objects on the workspace and move them to their desired positions with relatively low rates of failure.