Derek Wang

Styled Music Generation with Conditional Generative Models (presentation) (report)

Generative models such as Conditional Variational Autoencoders (CVAE) are widely used for music generation. Such tasks include styled melody generation and accompaniment generation. However, generating styled music through analogy making and swapping different conditions does not always work well. In our project, we propose several methods that may enhance conditional effects in the music generation process. The constrained CVAE models we propose show promising results for composing minor styled melodies by changing the model’s condition from major chord to minor chord. We conduct an extensive analysis about the pros and cons of the algorithms, and we introduce two unique perspectives that may be useful for future research.

Quinny He and Zheng Zhang

Sensing Acrobatic Human Movement (presentation) (report)

Approaches based on deep-neural-networks (DNNs) have dominated the field of human pose estimation in terms of both efficiency and performance. Numerous pre-trained models, such as COCO and BODY_25, are capable of accurate body tracking in real-time. However, since most models are trained with datasets containing only upright, everyday poses, these models often fail in the case of inverted, acrobatic movements. In this paper, we implement a rotational post-processing algorithm, that is able to augment any existing pose estimation models to a performance gain of 381% AP-score and 408% AR-score. Evaluation is carried out using the standard OKS metrics proposed by the COCO 2019 Keypoint Detection Task on a custom dataset featuring two professional circuit performers and five handstand variations. Furthermore, we build on the dataset and the algorithm to train a model through transfer learning, which will open doors for real-time tracking.

Nathalia Lin and Yufeng Zhao.

Improved Merging and Rendering for the NYUSH IMA Telewindow Project (presentation) (report)

Initiated in the IMA department at NYU Shanghai, “Telewindow” is an ongoing art and research project with the goal of enhancing 1:1 live teleconferencing using state-of-the-art VR/AR techniques, such as eye-tracking and volumetric video. Implemented with four Intel RealSense depth cameras, the pipeline produces an overall immersive single-user experience, where a volumetric representation of the user is presented through an interactive stereoscopic viewport. However, there are still improvements to be made in the quality of the graphic pipeline by mitigating both the noise in the raw point cloud from the depth cameras and the artifacts introduced by imperfect merging from different views. To improve the visual outcome of the real-time volumetric video under the current hardware setup, our approach includes merging and rendering respective point clouds by performing marching cubes mesh reconstruction in real-time. As a result, we implemented a different rendering technique into the current pipeline and introduced a possible approach that could be expanded for the “Telewindow” project.

Wujie Duan and Wenxin Feng

Study of ArcGIS - 3D Modeling for Urban Buildings and Improvement of Deep Learning Algorithms for Object Detection (presentation) (report)

This project aims to contribute to urban study and ecology by means of developing an analytical tool and improving the accuracy of object detection. It can be divided into two parts. First, a 3D map of urban buildings is created using ArcGIS web API, which is also able to do statistical analysis and export specified data. Second, a palm tree detection system is implemented based on deep convolutional neural networks as an improvement of the palm tree detector used on ArcGIS pro. The new model increases the accuracy of detection from 63% to 74% and reduces the time of 10 minutes.

Casey Pan and Tianyu Zhang.

TeleWindow Compression & Streaming (presentation) (report)

Volumetric data has become more popular in various industries across field such as arts, sports, geography and medicine, which creates an urging need for a robust and feasible solution for volumetric data streaming. Our proposed solution is a volumetric video compression and streaming system built with Socket.IO as a base protocol, using ZFP compression algorithm, featuring receiver-end frame rate expansion with key frame registration on the receiver machine’s frame buffer. We conducted experiments comparing multiple compression methodologies including DRACO and ZFP, tested WebRTC and Socket.IO as base protocols, and implemented key frame registration method for receiver-end frame rate expansion. Our current streaming system achieved real-time streaming with latency under 100ms and viewer frame rate of 30FPS.

Grace Gao and Xinyi Wang

A study of Migrant Pattern in New York City through Digitized Historical Data of the City’s Culinary Scene (presentation) (report)

By the end of the 19th century, a large percentage of the restaurants in New York City were owned by foreign-born citizens, which have left significant records for us to learn about the intriguing mobility pattern of immigrants in New York around that period. Prof.Heather Lee at NYU Shanghai has been conducting a project calledNew York Restaurant Keepers that aims to explore the migrant experience in New York City through its culinary scene. Along the way, there has been a problem that has been holding back the research process. The research group needs to digitize data from images of scanned documents through OCR (Optical Character Recognition),but both the tabular structure and the noises of these images affect the recognition accuracy significantly. Digitized data from these documents are crucial to the project because all of the further data analysis depends on it. Our research project, therefore, aims to bring up the OCR recognition accuracy rate and provide visualization based on the extracted data. Our solution proposes a tailored preprocessing and recognition process customized for the documents used in NewYork Restaurant Keepers. Images are categorized into groups by their needs of preprocessing steps, and we introduced various data clean-ing methods into the post-processing phase. This process depicted more than 85.28% accuracy on the subset of data from the borough of Manhattan, and we published the visualization of map and charts onto the project website for demonstration and future reference.

Xinyue Chen and Wenxin Feng

Automatic Subreddit Recommendation System (presentation) (report)

This project aims to improve the current subreddit recommendation system via building a classification model that suggests a subreddit to the user that based on his or her post title. Word2Vec is used as the major embedding method. Different machine learning models and deep learning models are implemented to solve this classification problem, including Naive Bayes, support vector machine, random forest, logistic regression, LeNet-5, TextCNN, and TextRNN. The baseline is 0.71 achieved by Naive Bayes classifier, and the highest accuracy is 0.88 achieved by TextRNN model.

Osborn Chen and Yunxiao Song

Customer Retention Modelling (presentation) (report)

The problems we are addressing are the prediction of customer attrition and what factors contribute the most to it. The prediction of customer attrition can help HSBC come up with targeted approaches to retain customers. Studying what factors contribute the most to customer attrition can help provide signals for staff in HSBC and identify scenarios when a customer is highly likely to leave HSBC. The issues we are facing include missing values, value transformation, and the identification of what factors contribute to customer attrition. Our approaches include analyzing and visualizing the distribution of attrition and no attrition for each feature, feature engineering, machine learning modeling, and performing ensemble methods on machine learning models. The most important and relevant features that the machine learning models produce basically aligns with our data analysis result. Moreover, our models achieve over 0.95 AUC score.

Pengyu Lu

Deep Classification of Reddit Posts (presentation) (report)

This project aims at developing a classification model for Reddit post titles using deep learning methods. This model can help users choose a subreddit to create a post and improve their user experience. In this project, different CNN models are created and examined. The best model significantly outperforms the baseline and has an accuracy of over 0.78.

Yuxin Zhang

Dynamic Pricing Model for Commercial Aircraft (presentation) (report)

Efficient aircraft valuation and prediction are vital for Aircraft Appraisers, as well as other participants of the aviation industry. However, the monopolistic industry feature, pre-established theoretical valuation approaches, and lack of data transparency hindered the application of a quantitative modeling approach towards aircraft valuation. This project targets at solving those issues by pooling data from multiple sources such as reports from FAA, OEMs, Airlines, and Appraisers. Through a quantitative exploration of multiple models, including the Ridge Regression Model, the Random Forest Model, the XGBoost Model, this project eventually attains a Stacking Model, with Training_mse of 0.02145 and Testing_mse of 0.02159 (onLog(Base Value)/1 million dollar).

Alexander Konrad Bogdanowicz

Dynamic Topic Modeling: COVID-19 (presentation) (report)

COVID-19 has led to unprecedented changes in global politics, economics, and social interactions. Amongst its other impacts, this work aims to study, through the use of advanced Natural Language Processing and Dynamic Topic Modeling, social sentiment and reactions to COVID-19 through topics expressed on social media. We deploy an end-to-end data processing pipeline and implement a Sequential Latent Dirichlet Allocation model to track the daily growth and changes in topic composition of 8 million tweets over the period March 31st to April 13th. We successfully identify 12 diverse topics which exhibit strong macro-trends over time while experiencing micro topic variations, covering domains such as healthcare, politics, community, and economics.

Zhuoer Wang

Misinformation and Fake-news Diagnostic Software (presentation) (report)

With the development of Internet and social media, the speed by which misinformation and fake news spread has reached an unprecedented level. Many researchers and organizations have worked hard on combating the spread of misinformation, but few of the techniques are available to the average information consumers. Therefore, my project builds a user-oriented software helping people check the reliability of information and providing a series of related materials for their reference. The information reliability is predicted by several popular machine learning and deep learning models, and the references are based on a real-time search.

Fang Cao, Zheng Li, and Qiyun Zhang

Spending Patterns of High Net Worth Individuals (presentation) (report)

High net worth individuals, or HNWI, are targeted by HSBC, as this group of customers will bring larger profit to the bank through personalized financial services customers. By analyzing the spending patterns of the customers, we summarize the characteristics of HNWI customers, and ultimately classify the customers accurately. With the help of machine learning algorithms, useful features representing the characteristics are found, and the classification model is also accessed. Meanwhile, other findings during the process of modeling also see the limitation of classification through spending patterns, showing the necessity of comprehensive researches combining spending patterns with other features.

Joseph Hensersky

Automated Network Mapping and Virtualization (presentation) (report)

Configuration changes and updates on a live environment can introduce unpredictability where mistakes and unforeseen problems cost real time, money, and effort for corporations and institutions. To combat this, testing such changes in an emulated environment before live deployment can have huge benefits. Our aim is to evaluate and compare a subset of relevant technologies that are well-equipped for automating the creation of test environments based on preexisting live networks and hosts. Our research has identified relevant technologies, established a testing methodology, and compared several solutions.

Jonathon Haley

Building a GUI for a voice-recognition demo system (presentation) (report)

My objective for this project was solve both of the problems by creating a high-level Python GUI library for embedded systems, and then use this library to create a GUI for Syntiant’s neural network testing system. This not only solves Syntiant’s problem of demoing its chips, but also serves as a test case for the GUI library, to demonstrate its effectiveness both as a graphical touchscreen interface and as an easy-to-use programming API. To create the Python GUI library, I examined existing open-source embedded GUI solutions in other languages and chose one of them to extend with a Python API. Enabling the GUI’s use in Syntiant’s system also required adding a touchscreen to their testing system and writing device-specific code to allow the GUI to communicate between the microprocessor and the screen.

Masaki Kagesawa

CourseLink - A Degree Requirement and Course Management Platform (presentation) (report)

Degree requirements at NYU Shanghai are managed and shared with students through multiple Google Sheets tables. This has made the course selection and academic advising process both unnecessarily time-consuming and error-prone. CourseLink will display the course requirements in an easy to understand UI, show the student’s current degree progress, recommend courses to students, and allow search for study away sites based on major. The recommender system is built with item-based collaborative filtering and user-based collaborative filtering algorithms.

Owen England

Cybersecurity Learning Platform (presentation) (report)

Among the general public there is a large knowledge gap surrounding cyber security skills, and best practices. This is a complicated issue, as cyber security is a highly technical topic, but also important as security mishaps and problems have the ability to impact any, and every user of a service. In order to address the knowledge gap and tackle the challenges of teaching a technical topic such as cyber security, I propose a platform that uses a Capture the Flag style format, a popular competition among security enthusiasts and professionals, to "gamify" the learning and education process. My platform addresses the problem by encouraging the users to solve relevant challenges in combination with reviewing supplementary resources to comprehend various issues and threats to enable them to make informed decisions to better protect themselves online.

Anpu Li, Wenhe Li, and Jinzhong Yu

Hyper.media (presentation) (report)

Low-latency information transmitting is crucial to all fields including finance, entertainment, and education. Therefore, we want to propose a low-latency streaming solution that has little requirements on the network quality and give an auto-benchmark solution for latency. In the low-latency streaming solution, we proposed on-device deep learning to reduce the usage of bandwidth and used HTTP2 & 3 to reduce overhead while transmitting. Finally, we found that with current on-device computational power, we can not utilize the Deep Learning model to reduce latency but modern HTTP protocols can reduce it. Besides, with the auto-benchmark, we found the major factors (jitter & bandwidth) for live streaming latency and how they affect latency.

Hongyi Li and Hongyi Liu

Popularizing Distributed Systems (presentation) (report)

Many beginners find the abstractions and jargon in the study of distributed systems hard to comprehend. This is not conducive for beginners to form an intuitive understanding of the problems and algorithms. In this study, we propose dramatization as a pedagogical tool to help beginners understand distributed algorithms. We embed the concepts of distributed algorithms in a play set in the universe of Gulliver’s Travels [1]. We present the drama in the form of a role-playing game (RPG). Our preliminary experiment shows that dramatization can improve beginners’ understanding of distributed problems. However, we believe that there is still much to improve regarding the presentation of the drama.

Amy Chen

ROOM Transportation Charge Estimates (presentation) (report)

For many companies, predicting how much they will be spending in the future is an important metric for budgeting. In an attempt to relieve the stress of waiting for the end of the month to calculate the total spending, the company ROOM, which specializes in delivering and assembling its work booths, has requested a program to help in predicting their total expenditures for insights. As they work with different carriers, a created database of sourced data will be gathered to ease the pain of meticulously looking for carriers that deliver and assemble at the best cost. Instead of analyzing each different cost from each carrier individually, the program can make predicted cost with transportation optimization methods to find the optimal cost from the different warehouses ROOM works with. Since this application needs to be unique to their orders, no current application exists so a website has been built to estimate and summarize a minimum for the future months’ total with the use of the Least Cost Method and Modified for Distribution Method.

Qihang Zeng, Chen Zhao, and Chongling Zhu

A Comparative Study on Frequent Itemset Mining Algorithms (presentation) (report)

With greater availability of data nowadays, the analysis of big data becomes increasingly essential. One of the most important tasks is Association Rule Mining, of which identifying frequent itemsets is the key part. Over the last two decades, a lot of research work has focused on developing and optimizing Frequent Itemsets Mining (FIM) algorithms. They come in four categories: centralized versus distributed algorithms, static versus dynamic algorithms. Several surveys do a great job of introducing FIM algorithms. However, most of them focus on MapReduce implementations and only consider static FIM algorithms. In this paper, we set up metrics for comparing both static and dynamic FIM algorithms, and conduct an experimental comparison on Spark.

Abby Feehan

A Reputation System for Distributed Mobile Payments (presentation) (report)

User-to-user mobile payments are already a reality in everyday life. Cashless mobile payments offer a convenient, secure alternative to cash or debit cards. In China alone there are estimated to be 700 million unique mobile payment users by 2022 [1]. A huge limitation of current mobile payment services is their need to gain real time access to banking information. Users must have stable network connectivity in order to carry out a transaction. An alternative design for mobile payments verification is a reputation based system. By allowing nearby users to take part in a verification process, the transaction may be completed without a need for centralized processing. Distributed processing enables secure mobile payments without the need for a strong internet connection. This paper implements a reputation system for mobile payments which can be combined with a delay tolerant network to provide service in low connectivity areas.

Chuyi Zhang

Mobility Models to Assess the Feasibility of Mobile Payments (presentation) (report)

This project looks into the use of distributed ledger on top of a delay-tolerant network to allow mobile payments to happen with intermittent network connectivity. A study of ensuring the trustworthiness in delay-tolerant network with a reputation system is also included.

Ariana Cai, Yixiang Xiao, and Yijie Zou

Parallelization of Monopoly Game Simulations (presentation) (report)

The Monopoly game can be a tool for economists to study real-world economic equality. In this project, researchers need a program to simulate numerous Monopoly games with changeable parameters; and the interface should be aimed at people with no computer science background. This project builds a Monopoly game simulation system which can create parallel simulation jobs on HPC(High-Performance Computing) and a user interface to connect to the system remotely. Users can acquire statistics about a large number of Monopoly game simulations on HPC by filling a simple form.

Anqi Luo and Ting Luo

Parallelization of Particle Physics Simulations (presentation) (report)

The recursive centered T-matrix algorithm (RCTMA) is a common computational technique for solving multiple light scattering problems. This method involves a huge number of computations, which largely restricts the range of numerical studies. Therefore, this project seeks to optimize the computation speed through code refactoring and parallel computing. Due to the strong data and logic dependencies among the computation jobs, how to improve resource utilization and the degree of parallelism exerts a great challenge for optimization. We revised the data storage manner and the logic of matrix manipulations in the original implementation, and adopted parallel computing to carry out the original serialized process simultaneously. As a result, we achieve better performance and largely shorten the running time.

Zhiye Xie

Revisiting the Windows 10 Random Number Generator (presentation) (report)

We introduce the wheel coloring problem which is abstracted away from the rotation condenser of Win10 RNG and derive a new approach to compute an upper bound for the minimal number of samples that make the internal state of Win10 RNG fully random. The new approach, however, only works for samples from a uniform distribution over the first k bits. For general min-entropy distribution, we show that O ̃(n*n/(d+k−n)) samples from distribution with min-entropy (n − d + 1) can be condensed into Ω(n − d + 1)-bit entropy, given a prime number n. Finally, we show computational results for the soundness of different rotation numbers. We also propose two permutations and show they have a faster rate of convergence in accumulating entropy than rotation.

Frederick Morlock

Exploring the Limitations of t-SNE (presentation) (report)

Introduced in 2008, t-SNE is a dimensionality reduction algorithm that has attracted the attention of many researchers due to its remarkable performance. However, there does not exist a clear consensus in the research community as to how t-SNE achieves its impressive empirical results. One of the important questions that remain is: What are the limitations of t-SNE? In this paper, we intend to explore the limitations of this popular algorithm. In doing so, we have constructed datasets based on the challenges that other researchers have faced in their utilization of t-SNE. By exposing the weak parts of t-SNE, we are opening a door to further improvement on dimensionality reduction algorithms.

Dingsu Wang & Zijian Zhou

Meta Reinforcement Learning (presentation) (report)

Deep Reinforcement learning algorithms usually require large amount of data to learn a task well, which is extremely inefficient under many scenarios in real world problems such as robotics. Meta reinforcement learning (Meta-RL) algorithms aim to solve this problem by helping the agent adapt quickly to new tasks with the prior knowledge about past tasks, which is very sample efficient. The difficulty of Meta-RL is to spot the correlation between visited tasks and unseen tasks, and current algorithms have not reached decent results for some complex tasks. Therefore, we propose a novel algorithm from the perspective of diverse skills learning which have reached the SOTA under some experimental settings. Moreover, we formulate a novel variation of Meta-RL problem named Batch Meta-RL and propose two Batch Meta-RL algorithms which achieve fast adaptations using historical data from past tasks.

Kelly Marshall

Multi-Level Meta Reinforcement Learning (presentation) (report)

Meta Reinforcement Learning challenges an agent to use experience from previous tasks to rapidly learn to solve new ones. Doing so requires reasoning over a distribution of tasks and discovering shared structure. This paper proposes a novel approach to Meta-RL which uses a hierarchical agent to promote efficient learning. We provide a theoretical motivation for this approach as well as an implementation with initial results.

Wanli Hong

Comparing Traditional Methods with Graph Neural Networks for Solving the Community Detection Problem (presentation) (report)

Community detection has long been a studied problem in the field of data science. There have been many proposed traditional methods that can obtain decent results. With the recent development in deep learning and network analysis, Graph Neural Network (GNN) is proposed as a new approach to tackle community detection problem. Our paper adopts the Line Graph Neural Network(LGNN) model and compares its performance with two traditional methods, modularity maximization and spectral clustering on solving community detection problem on two real-world network datasets. Two adjustments of LGNN are also proposed in terms of the loss function and input feature to enhance its performance. Especially, the LGNN model achieves the best result among all the methods on the second dataset.

Yuntian Ye

Nearest Neighbor for Recommendation System: From Neural Network and Graph Signal Perspectives (presentation) (report)

The building of a recommendation system can be viewed as a matrix completion problem from a mathematical perspective. The nearest neighbor method addresses such completion problems. This paper aims at understanding the similarity construction and the nearest neighbor aggregation in the predicting process.