SPRING 2020 PROJECTS
Working with Austin Pets Alive! (APA!), an Austin, Texas-based no-kill animal shelter, AdOptimize is revolutionizing the pet adoption process. Namely, one of the limiting factors in adoption rates is the poor quality of the animals' photos -- from a lack of photos, bad angles, lighting, etc. By using Machine Learning, AdOptimize greatly improves the photo process. Empirically, great success has been shown by capturing photos of dogs (increased adoption rates). A natural extension is to focus on making similar progress for cats.
Kensho, a private finance technology firm, aims to understand and predict financial markets, ideally by using the wealth of text data on the Internet. In attempt to understand any text, it is often crucial to know which entities (e.g., people, locations, organizations) are in the text.
Independent of this, our world knowledge can be represented as a structured format called a knowledge graph, where each node represents a real-world entity, and every node is connected to others via their real-world relationships (e.g., Michael Jordan was a teammate_with Scotty Pippen, and was coached_by Phil Jackson). Given a knowledge graph constructed from all of Wikipedia's data, can we use this structured information to better understand and identify entities within other non-structured text documents (e.g., financial reports).
Harvard Law School's Caselaw Access Project aims to greatly improve public access to U.S. law, in part by publishing all U.S. course decisions in a structured, easy-to-access format. With this wealth of text data comes the power to glean incredible insights into not only law but also the ability to create useful tools to assist in summarizing court rulings, citing relevant sources, or predicting court outcomes. Further, via longitudinal analysis, one could extract meaning insights corresponding to social impact (e.g., the war on drugs, gender inequality, etc).
FALL 2019 PROJECTS
Working in collaboration with the Associated Press (AP), this capstone group built a Text-to-Image recommendation system to recommend a set of images using headline captions.
Since Machine Learning methods cannot optimize text directly, the team converted text to a numerical representation using word embeddings, which are means by which a word can be represented as a vector of numbers.
In April 2019, the Event Horizon Telescope (EHT) Collaboration released the first image of a black hole. To accomplish this, the EHT used radio dishes across the globe simultaneously recording radio waves from near the black hole, synchronized by Global Positioning System (GPS) timing and referenced to atomic clocks for stability.EHT observations typically take place during a 10-12 day window with 5-6 days to be triggered when conditions are optimal. This project's goal is to use machine learning and/or prediction methods to help the EHT determine which nights should be triggered for global observations. This is an opportunity for students to work with EHT scientists and engineers on various aspects of black hole science in order to assess the probability that observations will lead to breakthrough results.
Deep learning frees us from feature engineering, but creates a new problem of “architecture engineering”. Numerous neural network architectures have been invented, but the design of architectures often feels more like an art than science. In this project, we investigate an efficient gradient-based search method called DARTS (Differentiable Architecture Search). DARTS is shown to require ~100x fewer GPU hours than previous methods like NASNet and AmoebaNet, and is competitive to the ENAS approach from Google Brain. We will compare DARTS to random search and state-of-the-art, hand-designed architectures such as ResNet.
Named Entity Disambiguation (NED), or Named Entity Linking, is a natural language processing (NLP) task which assigns a unique identity to entities mentioned in text. This can be helpful in text analysis. For example, a financial company may want to identify all companies mentioned within a news article, and subsequently investigate how the relations between the companies might affect the markets.
Deteriorating roads plague areas with highly volatile weather and budgetary constraints. It’s a constant challenge for municipal governments to keep ahead of the wear and tear as they catalogue and target hot spots to fix. In the U.S., most states only employ semi-automated methods for keeping track of road damage, and in other parts of the world, the process is completely manual, or foregone altogether. The costly and time-consuming procedure for collecting these data is only compounded by the fact that it must be done with relatively high frequency to ensure the data are up to date. This begs the question: can computer vision help?
If everything continues as planned, Somerville, Massachusetts — a city just outside of Boston — will be getting a new subway line in 2021. Though the new line is exciting, it may cause issues for the existing citywide resident on-street parking program. To address transportation planning questions, Somerville is conducting an audit of their parking supply. They have a good estimate of on-street parking capacity, but they have much less data about off-street parking. Their question is deceptively simple: how many residential units in Somerville have off-street parking?
One of the main challenges for Spotify is to recommend the right music to each user. Users' satisfaction can be monitored based on whether they skip the recommendation. Therefore the goal of a good recommender system is to show users content they like, and to minimize the probability that they will skip a song. In this project, we present the problem of sequential music recommendations.
The original question of our project was whether we could incorporate information from a knowledge base such as WikiData to improve performance on NER. We explore several methods for constructing type-specific vocabularies compiled from the knowledge base and show the non-triviality of compiling and cleaning this data. We then explore several methods of incorporating these vocabularies to learn an NER classifier trained on Wikipedia articles in a weakly supervised way. We demonstrate the challenges of incorporating non-contextual information in a setting where context is key. Lastly, we show how we can incorporate ideas from low resource neural machine translation to improve the generalizability of NER classification.