2013 - 2018(expected)
Ph.D. student, Circuits and Systems, Zhejiang University
Machine Learning, Bayesian Learning, Variational Inference. My adviser is Chunguang Li.
Summer 2012
C/C++ Engineer Internship, State Street, Hangzhou
Maintenance and Development for Princeton Financial Systems.
2009 - 2013
Bachelor's Degree, Zhejiang University of Technology
Double major in Computer Science and Automation. My adviser is Shengyong Chen.
Biography. I am currently a PhD Student at College of ISEE, Zhejiang University, China, working with Chunguang Li on Variational Bayesian Algorithms and their applications in Distributed Systems.

Research interests. My Ph.D research work covers a range of issues: variational Bayesian inference, stochastic/distributed optimization, probabilistic graphical models, transfer learning, multitask learning and sensor networks. Currently, I am interesting in variety of approaches for deep learning (CNN, GAN, Deep Bayesian Learning, etc.) and their applications to computer vision (image processing, video analysis, etc.).

Curriculum Vitae (keep updating). My CV written in English can be found here, and that written in Chinese can be found here (中文简历).

On a side for fun I blog, weibo, and maintain several Projects (e.g. ActionRecognition, AgeEstimation, VBClusterings, UnixFileSystem).

Publications

List of papers on Google scholar.

Manuscripts

Junhao Hua, Chunguang Li,
Distributed Robust Kalman Filtering By Variational Bayesian Approximations,

in preparing.


Junhao Hua, Chunguang Li,
Distributed Jointly Sparse Bayesian Learning with Quantized Communication,
submitted to IEEE Transactions on Cybernetics, 2017.
Distributed Learning of Predictive Structures from Multiple Tasks Over Networks
Junhao Hua, Chunguang Li, Hui-Liang Shen
IEEE Transactions on Industrial Electronics(TIE, ZJU-TOP100), vol. 64, no.5, pp.4246-4256, May 2017.
We concerned with the problem of distributed multitask learning over networks, which aims to simultaneously infer multiple node-specific parameter vectors in a collaborative manner. In this work, we implicitly model the similarity of parameter vectors by assuming that the parameter vectors share a common low-dimensional predictive structure on hypothesis spaces, which is learned using the available data in networks. A distributed structure learning algorithm for the in-network cooperative estimation problem is derived based on the block coordinate descent method integrating with the inexact ADMM technique.
Distributed Variational Bayesian Algorithms over Sensor Networks
Junhao Hua, Chunguang Li
IEEE Transactions on Signal Processing (TSP, SCI-TOP), vol.64, no.3, pp.783-798, Feb. 2016.
We propose two novel distributed VB algorithms for general Bayesian inference problem, which can be applied to a very general class of conjugate-exponential models. In the first approach, the global natural parameters at each node are optimized using a stochastic natural gradient that utilizes the Riemannian geometry of the approximation space, followed by an information diffusion step for cooperation with the neighbors. In the second method, a constrained optimization formulation for distributed estimation is established in natural parameter space and solved by ADMM. An application of the distributed inference/estimation of a Bayesian Gaussian mixture model is then presented, to evaluate the effectiveness of the proposed algorithms.

Talks

2016 Nov: Talk at SIIP 2016 seminar, ZJU: Distributed variational Bayesian Algorithm in Networked System (in chinese) [slides].
2015 Jan: SIIP Group Talk: An Introduction to Transfer Learning [slides].
2014 Oct: SIIP Group Talk: Privacy Preserving Regression [slides].
2014 Oct: SIIP Group Talk: Vertically Partitioned Data [slides].
2014 Jul: SIIP Group Talk: Zero-Determinant Strategies [slides].
2014 Apr: Talk at Course of "Image & Video Analysis": Action Recognition & Categories via Spatial-Temporal Features [slides].
2014 Mar: Talk at csmath (2014-2015) course: A Tutorial on Variational Bayes [slides].
2013 Sep: SIIP Group Talk: Recursive parameter estimation and inference with incomplete data – Recursive EM & VB [slides].
2013 Jun: Undergraduate thesis defense: Brain MRI Segmentation based on Variational Bayesian methods [slides].
2012 Nov: SIIP Group Talk: Distributed Image Processing: Camera Networks, CV algorithms and Decentralized Multicamera Tracking [slides].

Projects

Action Recognition & Categories via Spatial-Temporal Features
Author: Junhao Hua, Shangyao, Lin
2014 April
This project consider this problem of recognizing and localizing multiple actions in long and complex video sequences containing multiple motions. Inspired by the previous works by Juan Niebles et al, 2008, we use the "bag of word" represetations for action recogntion. We first extract the spatio-temporal features (interest points), then construct the codebook (a set of words) by clustering of interest points using k-means algorithm. Then, the action categories can be infered by using unsupervised learning such as pLSA or LDA learned by the MCMC/variational inference. The hierarchal structure can be written as: document (video) - words (by clustering of interest points) - topic( types of actions). In this project, we simply use the supervised algorithms (such as KNN, SVM) to classify each word in every frame. For classifying multiple types of actions in a single video, we propose a simply algorithm called 'voting', to vote the top-N topics each frame/image is likely to have. This simple method can achieve the aim of multiple actions recogintion. Thanks to Piotr's Computer Vision Matlab Toolbox, the project is implemented by MATLAB.
show more
Action Recognition & Categories via Spatial-Temporal Features
Author: Junhao Hua, Shangyao, Lin
2014 April
This project consider this problem of recognizing and localizing multiple actions in long and complex video sequences containing multiple motions. Inspired by the previous works by Juan Niebles et al, 2008, we use the "bag of word" represetations for action recogntion. We first extract the spatio-temporal features (interest points), then contruct the codebook (a set of words) by clustering of interest points using k-means algorithm. Then, the action categories can be infered by using unsupervised learning such as pLSA or LDA learned by the MCMC/variational inference. The hierarchal structure can be written as: document (video) - words (by clustering of interest points) - topic( types of actions). In this project, we simply use the supervised algorithms (such as KNN, SVM) to classify each word in every frame. For classifying multiple types of actions in a single video, we propose a simply algorithm called 'voting', to vote the top-N topics each frame/image is likely to have. This simple method can achieve the aim of multiple actions recogintion. Thanks to Piotr's Computer Vision Matlab Toolbox, the project is implemented by MATLAB.

Misc

Address: Zhejiang University Yuquan Campus, No.38 Zheda Road, Xihu District, Hangzhou, Zhejiang, China.
Office: Administration Building #428
Emails: huajh -at- zju.edu.cn (replace -at- by @), huajh7 -at- gmail.com (replace -at- by @)
Last update:
  • 10 Apr 2017