top of page


Project | 01

Multi-dimensional Count Sketch

                             Appeared in NeurIPS 2018 workshop

When dealing with large-scale multi-dimensional data, it is crucial to design efficient compression schemes that remove redundancy while preserving relevant information.  In this work, we propose multi-dimensional count sketch (MS), an unbiased estimator of the given tensor. 

[Paper-Long Abstract] [Poster]

Project | 02

Question Type Guided Attention in Visual Question Answering

                                                    Appeared in ECCV 2018

We propose Question Type-guided Attention. It utilizes the information of question type to dynamically balance between bottom-up and top-down visual features, respectively extracted from

ResNet and Faster R-CNN networks.

[Paper]  [Poster] [Codes[Tutorial on VQA]

Project | 03

 ​We consider the problem of robust PCA in the streaming setting with space constraints. Our result is the first to obtain finite-sample guarantees while having the weakest assumption on the sparse perturbation, namely, deterministic support, and a standard identifiability assumption on the low-rank component, namely, incoherence.


Streaming Robust PCA

                           Appeared in NIPS 2017 OPT Workshop

In this project, we find a intuitive and efficient way to do tensor contraction with extended BLAS kernels (Bateched BLAS). We also benchmarked with state-of-art library: BTAS.

[Paper]  [Codes]  [Slides]

Project | 04

Tensor Contraction using Extended BLAS Kernels

                                                     Appeared in HiPC 2016

We applied matrix robust Principle Component Analysis(PCA) to tensor cases: separate a tensor as a low rank tensor and a sparse tensor. This is done by using nonconvex optimization. We comapred our method with matrix robust PCA.

[Paper] [Codes]

Project | 04

Project | 05

Tensor Robust PCA

                                               Appeared in AISTATS 2016​

Project | 06

Using tensor convolutional decomposition to find out usefull features for unsupervise/supervise learning.

Convolutional Tenor Decomposition in Unsupervise/Supervise Learning

                                                             Under preparation

Considering the physical water network as a graph, a CRF model is built and learned by the Structured Support Vector Machine (SSVM) . After that, we form high order CRF system by enforcing twitter leakage detection information.


Project | 07

Leak Event Identification in Water Systems Using High Order CRF 

                                                                    Course project​

To see more or discuss possible work let's talk >>
bottom of page