Current Research and Working papers
Large-scale optimization algorithms and its applications in machine learning. More specifically,
- Distributed second-order methods in Deep Neural Network
- Stochastic Gradient Methods (SGD, SDCA, SVRG, etc.) for Large-scale Nonlinear Optimization
- Distributed Restarting NewtonCG Method for Large-Scale Empirical Risk Minimization with Chenxin Ma, Mudigere Dheevatsa, Alejandro Ribeiro, etc.
- An Inexact Regularized Stochastic Newton Method for Non-convex Optimization with Ioannis Akrotirianakis, Amit Chakraborty, Martin Takáč, etc.
- An Inexact Regularized Newton-type Method for Non-convex Optimization, Phd Seminar, Lehigh University, 2017.
- Steps to Success in Training Neural Networks by Using Second-order Algorithms, INFORMS Annual Meeting, Houston, TX, 2017. slides
- Budgeted Category Offer Assignment Optimization, A Distributed Lagrangian Relaxation Approach, Precima, Chicago, IL, 2017. slides
- Distributed Hessian-Free Optimization for Deep Neural Network, AAAI, San Francisco, CA, 2017. slides
- Learning Deep Neural Networks by Inexact Newton-CG Optimization using Indefinite Stochastic Hessian, Phd Seminar, Lehigh University, 2016. slides
- Large Scale Distributed Hessian-Free Optimization for Deep Neural Network, MOPTA, Lehigh University, 2016. slides
- Dual Free Adaptive Mini-batch SDCA for Empirical Risk Minimization, INFORMS Annual Meeting, Nashville, TE, 2016. slides
- Dual Free SDCA for Empirical Risk Minimization with Adaptive Probabilities, NIPS, Montréal, Canada, 2015. poster
- Estimating Portfolio Loss Probabilities with Optimal Risk Loading Coefficients and Fixed Dependency among Obligors, Siemens Corporation, Princeton, NJ, 2015. slides
- Random Coordinate Descent Method on Large-scale Optimization Problems, Coral Seminar, Lehigh University, 2015. slides
Selected Course Projects
- Optimization Method in Machine Learning
Explored second-order methods to training fully connected deep neural network on handwritten digits classification problem (
- Massive Data Mining
Designed competitive Q&A system to attain up to 39.5% accuracy by using Apache Lucene and Natural Language Toolkit, etc.. (
- Computation Methods
Used l1-regularized lasso model to recovery pictures with missing pixels. Multiple algorithms (ISTA, FISTA, GRPS) are implemented in
- Pattern Recognition
Matlabsoftware package to compare various of classifier technologies (SVM, Artificial Neural Network, Decision Tree, KKN) for character-image classification problem.
- Integer Programming
Pythonsoftware package to address mixed binary programming problem with branch and cut method. (Group project)
- Nonlinear Programming
Matlabsoftware package for unconstrained nonlinear optimization using various of Nonlinear Optimization methods (GD, Newton, CG, BFGS/DFP, TR and Amijo/Wolfe).
- Credit Portfolio Risk
Proposed model and developed a
Rsoftware package to estimate large-loss probability of a Portfolio with derived optimal loading coefficient.
Department of Industrial and Systems Engineering, Lehigh University, Bethlehem, PA.
ISE410/Industrial Automation and Robotics (Spring 2018), with Prof. Derya Pamukcu
ISE324/Design of Experiments (Fall 2017), with Prof. Eugene Perevalov
ISE112/Applied Engineering Statistics (Spring 2015), with Prof. Alexander (Sasha) Stolyar
ISE112/Applied Engineering Statistics (Fall 2014), with Prof. Eugene Perevalov
Intel Corporation, Parallel Computing Lab, Santa Clara, CA.
Adaptive sample-size sub-sampled Newton-CG algorithms, with Mudigere Dheevatsa
Distributed Hessian-free methods for deep neural network, with Mudigere Dheevatsa
Business Analytics and Monitoring, Siemens Corporation, Corporate Technology, Princeton, NJ.
Deep learning optimization, with Dr. Ioannis Akrotirianakis
Portfolio credit risk, with Dr. Amit Chakraborty
Department of Mathematics, Nankai University, Tianjin.
Theory of Optimization (Spring 2014), with Prof. Qingzhi Yang
Linear Algebra (Fall 2013), with Prof. Qingzhi Yang