I’m a CS PhD candidate (supervised by Dr. Derek Doran) at Wright State University, Dayton OH. My research interests are broadly in machine learning, with a focus on explainable and reliable deep learning, computer vision, and interaction between vision and natural languages. My academic research studies deep neural network mechanisms toward interpretability, that aid in interpreting how they process input data in a human-understandable way. I am also interested in multimodel inference and reasoning and have proposed a novel task Visual Entailment to explore challenging problems on interaction between vision and natural languages.
Ph.D. in Computer Science and Engineering • GPA 4.0 • Jan. 2016 - Aug. 2020(expected)
Courses Taken: Computer Vision, Algorithm Design and Analysis, Network Science, Advanced Programming Languages, Machine Learning, Smart Cities:Devices & Methods, Distributed Computing, Computer Organization, OS Internals and Design
Selected course projects:
B.S. in Pure and Applied Mathematics • Sep. 2010 - Jun. 2014
Research Intern • Machine Learning Department • May. - Aug. 2018, Jan. - May. 2019
Graduate Research Assistant • Jan. 2016 - Present
Algorithm Engineer • Jun. 2014 - Dec. 2015
In view of limitations from VQA datasets (VQA-v1.0, VQA-v2.0, CLEVR) on the pursuit of visual intelligence in the ML community, we propose a noval visual reasoning task Visual Entailment, along with a new dataset SNLI-VE.
A short paper on this project has been accpeted by NeurIPS 2018 Workshop: Visually Grounded Interaction and Language (ViGIL). The full paper is available in arxiv.
*Work performed as a NEC Labs intern.
When a person's future, health, finances, and safety is on the line, human decision makers relying on an AI system need to rationalize the recommendations the system offers. Yet present day AI is hampered by our ability to interpret or ''explain'' the rationale AI uses to arrive at a conclusion. This project investigates new approaches for reaching ''explanations'' of an AI's decision making process, specifically applied to deep learning, for computer vision tasks.
A paper on this project has been accpeted by NIPS 2017 Workshop: Interpreting, Explaining and Visualizing Deep Learning, NIPS IEVDL 2017
*Work supported by the Ohio Federal Research Network.
The majority of requests seen by web servers and clouds are caused by web robots and IoT(Internet of things) agents. Prefetching web resources for caching and content management systems is a common technique to anticipate and pre-load the resources likely to be requested next for fast, low latency access. This project introduces a novel soft computing resource prefetcher, which marries a deep recurrent neural network(RNN) with a Bayesian network, to predict future requests made by web robots. Experiments with traffic logs from web servers across two universities demonstrate an improvement in precision and recall over a traditional dependency graph-based approach.
*Work supported by the National Science Foundation under grant #1464104.
Autonomous driving is an active research topic these years in both academia and industry. In this work, I designed a novel end-to-end augmented convolutional neural network (AUG-CNN) to address this problem. Based on raw image pixels from the front-facing cameras equipped on the car and the car’s current status on speed and throttle, AUG-CNN autonomous driving system is able to directly make decisions to control the steering wheel and the throttle. The training dataset and testing simulator are released by Udacity Self-Driving Car Nanodegree Program. Although the AUG-CNN system is trained only on the daytime-plain-driving dataset, the performance is surprisingly good in the simulator of a nightfall-mountain-driving scenario that the model has never seen. This work is the final project for Smart Cities:Devices & Methods.
The spread of misinformation in media has become an increasingly relevant problem in society. However, automatically identifying whether a news story is an incorrect claim or factual from the content of a story is an extremely difficult machine learning problem that has yet to be solved by the research community. Our approach is to examine dynamic social networks on the discussion of news for patterns which can distinguish between disreputable claims and factual stories. Using a dynamic stochastic block model (DSBM), we demonstrate the ability to infer structural behavior such as the emergence of new connected components in social networks from temporal patterns. We also explore the potential and difficulties of using network based features to train a classifier over the truth or falseness of news. This work is the final project for Network Science
It is well known that U.S. university students often graduate with copious amounts of financial debt. The tuition rates for some colleges are often far above a reasonable threshold of being considered practical, which often limits many aspiring students' enrollment choices to educational institutions that seem more affordable. Even worse, students who do graduate often find themselves facing repayment obligations that far outweigh their potential income earnings. These notations collectively spark a number of questions.