Importance Sampling with LSTM Networks - Aiwen Xu (poster) (report)
Importance sampling is one of the methods for estimating multidimensional sums and integrals. However, it is often not easy to find an importance sampling distribution that yields an estimator with a low variance. In this project, we aim to improve the importance sampling distribution by training an LSTM network with the gradient descent algorithm, because recent advances in deep learning has shown that neural networks are powerful function approximators. After implementing this method to calculate the normalization constant of the queuing problem, which is a class of multidimensional sums, we see that this approach is very effective in a simple setup of the queuing problem and reasonably effective in a more complicated setup of the queuing problem.
Neural Music Generation - Weilin Zhang (poster) (report)
In this project, we explore the problem of neural music generation. Music generation is a challenging problem since it needs both to control the generation quality and to maintain a long time dependency. We explore and implement a Wavenet-like structure for symbolic music generation. The model is able to generate high-quality long sequences and multi-track sequences. Finally, we discuss and analyze the results and draw future perspectives
Incremental Object Recognition with Tree-Structured Image Encoding - Kate Feng & Zixin Yao (poster) (report)
We address the object recognition problem. Our goal is to build an adaptive and robust system that more resembles human visual system, and that improves performance with more computation. We build upon existing work that takes glimpses at the image and introduce a tree-structured image encoding method. Besides, we use a teacher network to assist our model’s learning process. Our model achieves comparable accuracy with the baseline model and converges faster, and the visualization validates the usefulness and interpretability of glimpses choices.
Isla: the Infinitely Scalable Learned Assistant - Alex Grigoras (poster) (report)
As technology is permeating every aspect of our lives, the only tools designed to manage our technological environment, virtual assistants, are confined to narrow ecosystems and cannot interact with the vast majority of software running on our computers, smartphones, or IoT devices. I propose a solution to this challenging issue in the form of a virtual assistant which does not rely on software integrations or APIs, but which instead interacts with other software using emulated human actions and gestures. This would allow for infinite horizontal and vertical scalability, and would enable any end user to further expand the capabilities of the assistant in order to apply it to any technological task achievable by a human-being using software.
Coffee Break
Community Based Grayware Detection - Ruihan Yang & Shi Zeng (poster) (report)
Grayware, proposed by Andow et al. [1], is a new concept of Android software. It refers to unmalicious programs that have negative impacts on performance, user experience of Android applications(apps) or security of system, such as mobile adware or rooting tools. Existing industrial use of Android mobile app classification focuses on malware detection while grayware without any malicious actions are unable to be detected. That is to say, the industrial standard detection systems can only detect malware or benign- ware on the market. Those unclassified apps are called undecidable apps and require manual labeling. In this paper, we propose a community-based algorithm that builds a similarity network from mixed datasets and generates app clusters for different categories — malware, benign-ware and undecidable software. We seek to identify the existence of grayware among undecidable apps and prove that the similarities shared by some grayware enable them to form clusters together.
Lockchain: Blockchain Locking System for Revocable Access - Sam Arellano (poster) (report)
Revocable access is a problem that many locking systems do not solve effectively. This is a difficult problem because of the tradeoff between security and performance in distributed systems. I propose an efficient locking system based on blockchain smart contracts. My system’s scale is more fitting for this specific problem than existing alternatives. Smart contracts implement a user hierarchy which controls user access permissions and records all actions in the system. I also implement a user interface for interacting with these smart contracts, and a physical smart lock to give a tangible example of the system in action.
Frequent Itemset Mining in Action: A Treebased Distributed Solution - TianYi Zheng (poster) (report)
Given the number of transactions happening in today's ecommerce sites, an efficient frequent itemset mining algorithm can significantly enhance the accuracy and the speed of retailers' recommendation systems. Such algorithm often requires a distributed system that handles transactions in parallel. This paper presents an implementation of the FRENO algorithm which uses a compact tree structure to store frequent itemsets. The implementation ensures the uniform distribution of the requests to the servers and achieves automatic loadbalancing when the tree generates new nodes.
Single-Agent Pathfinding for a SLAM Robot - Johan Renzy Yao (poster) (report)
Implementing an optimal admissable path-finding algorithm for a simple mapping robot can be of great challenge, depending on the specifications and limitations of the robot. The issues lie with the nuances that come with employing a realistic implementation of this algorithm, such as the format of the graph input, the ability to simulate the limited movements of a robot, and the choice of a simulation testbed in the absence of actual hardware. My approach simulates a robot that travels multiple “shortest” paths given a starting point and destination point on a 2D grid map, whereby each path is provided by an A* path-finding algorithm that uses a different heuristic function. I implemented a simulation which takes into account the limited movements of a real life robot and allows for a performance comparison of different heuristics on various grid map specifications.
Mapping Indoor Spaces with Robots - Kevin Ni (poster) (report)
This project seeks to use an automated vehicle to explore and map an indoor space. This problem is usually referred to as the Simultaneous Localising and Mapping (SLAM). Particle filter (PF) and Extended Kalman Filter (EKF) are the two most studied and most popular SLAM solutions: PF is more widely used with visual- based sensors such as RGB and infrared cameras; EKF is more adaptive and has been proved to work well on various sensors. This project utilizes LIDAR to perceive its surroundings. Therefore, PF is a better match. I implemented both algorithms and conducted a comparative study: while both algorithms are able to map the given environment, on average PF consumes more computational power, takes more time to finish, and its output map is noisier.
Lunch Break
Visualization of Philippine Metro Rail Systems - Jose Reyes (poster) (report)
As the Philippine economy continues to grow at an increasing rate and as technology advances exponentially, the infrastructure for traffic, and more specifically public transit, falls farther and farther behind what would be deemed appropriate. Unfortunately, there are no technological platforms currently in place to monitor the service provided by public transit systems in the metropolitan area of Manila. My project provides a survey of various platforms that have been implemented to monitor metro transit systems in different cities around the world. Based on this survey, I propose MMViz (Metro Manila Visualizer), a web application that utilizes existing Philippine GTFS data to visualize the metro transit systems in the Philippine metropolitan area.
NYUSH Class Attendance Tracker - Daniela Oh (poster) (report)
Many classes in NYU Shanghai have attendance as a requirement, yet the school does not provide a tool for professors to take attendance. There are online solutions, but they do not address NYU Shanghai’s specific requirements for an attendance tracker. My project offers a systematic tool that is easy and simple to use, and complies with the school’s database format: professors can import the rosters on NYU Classes, take attendance effortlessly, and save attendance calls for later reference. According to survey results, my implementation achieves its goal: professors stated that they prefer to use this product over their current methods, that it is “easy to understand”, “very convenient”, and saves them “a lot of time”.
Class and Exam Time Scheduling System - Ziming Sheng, Weiyu Wang, & Yuxin Xia (poster) (report)
At NYU Shanghai, it takes the Office of the Registrar a long time to schedule exams and courses by hand, and sometimes they are not able to find a class/exam schedule that works for every student because of time conflicts. As students ourselves, we want to design final exam and course schedules that both meet our needs as well as the school’s requirements. While avoiding time conflicts for students and professors, we also considered some soft constraints, such as reducing early morning classes and spreading out the exams within the finals week. We implemented both a greedy algorithm and a genetic algorithm, and compared their performance based on time usage. We designed an efficient regeneration process to avoid performing the whole scheduling process again when users incorporate a small change. To give users easy access to our system, we developed a website that offers user interactions and visualizations of the schedules. Our solution generates a schedule within several seconds, and our regeneration process is 1000 times faster than the initial generation process. User tests show that our webpage is easy to use, even without any previous knowledge of our design.
Automated Grading Server for Programming Assignments - Jianqiao Shao (poster) (report)
This project aims to establish an automated assessment server for coding assignments in Python. This project is mainly programmed in PHP and uses a mysql database. It provides a solution for automatic grading in Introduction to Programming at New York University Shanghai .
Coffee Break
Design of a Transcription Application - Ziyuan Huang (poster) (report)
Transcription is a very time-consuming process: users spend lots of time not only typing down the texts, but also trying to find the right positions in the audio. I try to speed up the process by building a transcription application that allows the user to post edit computer-generated transcription. Performance tests show that my application speeds up the transcription process significantly.
Computer Architecture Probe - Yuxia Yao & Yi Zhao (poster) (report)
In our Computer Architecture class, we turned to graphical simulators to learn how real computers work “behind the scene”. But such tools only focus on a particular layer of representation, and fail to show the transition between layers. In order to bridge levels of abstraction, our Computer Architecture Probe offers a simultaneous multi-layer simulation, covering the microarchitecture, intermediate, gate, and transistor layers. It allows users to switch between layers instantly and observe the simulation step by step, helping them understand the connections between the different layers.
Real-Estate Recommendation System - Kalkidan Eteffa & Quinn McHale (poster) (report)
The process of buying or renting real-estate is a stressful and inefficient process. Renters find it extremely difficult to find exactly what they are looking for and usually end up choosing an apartment that is not ideal. In this project, we aim to make this entire process more efficient by using a personalized recommendation system. Our paper looks into random algorithms, collaborative filtering algorithms and content-based filtering algorithms. Based on our findings, Pearson’s Correlation Collaborative Filtering gives the best matches by minimizing pitfalls of recommendation systems, while Content-Based filtering can be used to complement collaborative filtering recommendations. We also built a prototype web application that implements these algorithms.
CityConnect: A Map-Based Communication Tool For Local Government - Colton Paul (poster) (report)
A critical component of community participation is the communication between a government and its constituents, and currently some of these channels for communication on the local and city level are unengaging or inefficient. My project highlights the shortcomings of existing systems of communication about city infrastructure, discusses current technologies which could be implemented in a new system, and implements those ideas in an application called CityConnect.
Slime Mold Computer - Han Su (poster) (report) (video)
Researchers have successfully used slime mold for route optimization. What is the boundary of slime mold’s computing power? In today’s era of intelligent machines, we want to explore slime mold’s potential as a living black box, a new direction in the field of natural computing. Our approach mimics the training and testing process of machine learning algorithms at large. Our result shows that slime mold, as a single-celled organism, can grow in specific patterns after training, while slime mold’s capability to learn conditioned responses is not salient.