Hengjia Zhang at the University of Michigan

My Experience

Deep Learning Intern at The Mathworks

May 2019 - Aug 2019, Natick, MA

Here I was working on the interoperation between MATLAB and other deep learning frameworks, mainly about the interoperation between MATLAB and Keras:

  • Implemented the translation from CuDNNLSTM on Keras to MATLAB Deep Learning Toolbox LSTM Layer.
  • Refactored the architecture design of Keras Model Transformer in MATLAB to be more organized, maintainable and scalable.
  • Implemented the nested Sequential Keras Model Transformation to MATLAB to achieve full Keras Support.
  • Implemented the translation of TimeDistributed Wrapper in Keras where DNN can be applied to the temporal dimension for video inputs.
  • Wrote RFA files, created unit tests and regression tests.

Translation between Real Person and Anime Character by using CycleGAN

Jan 2019 - Apr 2019, Ann Arbor, MI

Anime-to-Real-TransformationGAN is about translation between a real person and an anime character based on CycleGAN. The final result was impressive. We manually collected and cleaned about 15000 data includes both real person images and anime character images for training and the data was trained on my NVIDIA1070 for almost one week. Since we couldn’t map every detail from real person to anime character or anime character to real person, the original design of CycleGAN may not perform very well in this case. Therefore, in order to let the discriminator learn global features about the images better, I added some dilated convolution layer to the discriminator so that it could classify the images more broadly. In addition, to preserve the images details, skip connections was also added on the generator.

ARTGAN architecture The Architecture of ARTGAN

character to person Anime to Person Transformation

character to person Person to Character Transformation

Jan 2019 - Apr 2019, Ann Arbor, MI

This project is to find out whether community detection methods will help link prediction as the graph is growing larger and lager. Link Prediction has wide applications from friend recommendation in Facebook to connection recommendation in LinkedIn. The early proposed approach is based on network statistics such as common neighbours, Jaccard and Adar. As the graph becomes larger and larger, it can become unrealistic to process the whole graph when we try to predict potential links within the graph using these ap- proaches. One solution is to detect multiple communities in a graph, and predict links within each relatively small communities. How- ever, does community detection really capture the essence of the relationships and help downstream tasks such as link prediction? We design an empirical study framework and our proposed method is to use five different community detection methods to partition the graph into multiple communities and use each community as a subgraph to run classic link prediction methods. Through our experiments, we observe that BIGCLAM and METIS perform the best and LEMON performs the worst in terms of link prediction ac- curacy. Some community detection methods can help downstream tasks such as link prediction to achieve better accuracy in terms of denoising but not all methods are equally good.

Graduate Student Instructor for EECS 486 Information Retrieval

Jan 2019 - Apr 2019, Ann Arbor, MI

I was one of Graduate Student Instructor for EECS 486 Information Retrieval in Winter 2018. Here I led discussions, held office hours, gave exam papers and graded them. I taught students from text classification methods like Naive Bayes and Neural Networks, to some theories like fuzzy set, and technical web crawler techniques like BeautifulSoup.

Smart Fiction Search Engine (Boogle)

Sep 2018 - Dec 2018, Ann Arbor, MI

This project is to develop a Smart Fiction Search Engine which searches books based on plot and context. About 5000 books for their descriptions and relative reviews were crawled, about 30 reviews for each book and build an Inverse Index database for it. Based on that, I implemented Okapi BM25 for books and used it as the ranking function for fictions retrieved. In order to test the performance of the search engine, we created 400 queries ourselves and compared the performance between our own book search engine and Google Books Search Engine. The result shows that our top 10 accuracy was 72.5% while Google’s accuracy was 48.3%.

boogle User Interface of Boogle

Gesture Recognition System under Complicated Background

May 2018 - Aug 2018, Shanghai, China

This project is to develop a gesture recognition system by using a USB camera by using OpenCV and Keras. We applied three different masks from OpenCV to achieve hand segmentation from complicated background and a 16-layer Convolutional Neural Network to classify segmented hands.The system classify hands in real time with accuracy of 100% and response time of 60 mini-seconds per frame. Finally we use the recognition result as game inputs to show it’s quickness, accuracy and robust.

Demo of Gesture Recognition System Demo of Gesture Recognition System

Data-driven Programming System on Java Code Prediction

Jan 2017 - Apr 2018, Michigan Database Research Group, Ann Arbor, MI

In this project, I aimed to use algorithm to refine and eventually achieve automate code writing. The end goal is to develop a tool that can enable users to put in demands and get the code they need without actually writing it. Currently, our system can already predict the function or methods that a user would want to call. I also developed a visualized tool for the core machine learning algorithm, which is a plug-in tool for the editor ATOM so that people can easily access it and use it. During the research, what we did first was to design a RESTful API, which could send contents in the editor to the server and returns the predicted results, and second is that once we got the predicted result from the server, to exhibit a list of suggestions in the editor to help the users.

tarocchi_demo

Multi-Network Representation Learning

Sep 2017 - Jan 2018, GEMS Lab, Ann Arbor, MI

For this project, I leveraged Local Sensitive Hashing to scale up representation learning in an attempt to better solve network alignment problems. On top of that, I also implemented and analyzed empirical variants of network alignment algorithms.

IMDb Python Package

Jan 2018 - Apr 2018, Ann Arbor, MI

This is a package which is trying to fetch information needed from IMDb since there’s no official API from IMDb right now. This package can easily fetch all important information one may need about any movie. Basically here I’m using python package beautifulsoup to crawled all the data from IMDB website and it’s on real time!