Trustworthy Deep Learning: Explainable AI (2024)

CPSC471 Project Report by Jeb Cui

"Ensemble explanations, also known as composite or aggregate explanations, offer a promising approach to enhancing the explainability of machine learning models. In this context, I analyze and evaluate the work of Hryniewska-Guzik et al. I propose and test new aggregation functions and pre-aggregation modifications to enhance the quality of ensemble explanations. Results indicate that the median aggregation function produces some of the most robust explanations, while aggregating using the minimum absolute value function gives the best complexity scores. However, many of the approaches strike a balance among the metrics, which is often more suitable for general use..." Check out this project's Github!

iQuHack (2024)

1^st place Moody's Analytics Challenge

By Zachary Zitzewitz (Yale '26), Victor Cai (Harvard '26), Jeb Cui (Yale '26), Christian Kang (Yale '26), and Mason Abrell (Yale '26)

“Team <Y|🦆|H> competed in Moody’s Analytics’ challenge, which was to build a quantum portfolio optimizer algorithm, from start to finish, and prepare a pitch on why companies should invest in quantum today. The final algorithm used quantum annealing to solve QUBO-formulated mean-VaR portfolio optimization problems. This team’s algorithm was tested using simulated annealing and on real quantum hardware using D-Wave Systems’ quantum annealer 2000Q. Their algorithm’s optimized portfolios lead to projected returns of up to 6% per business quarter.” Check out this project's news article! Check out this project's Github! Check out Moody's hackathon recap!

Bosonic Qiskit Textbook (2022-23)

Supervised by Prof. Steven M. Girvin and Dr. Kevin Smith.

“This textbook presents a zero-to-sixty introduction to continuous-variable quantum computing... This work was produced by the Bosonic Qiskit project group under the Yale undergraduate Quantum Computing club, consisting of Ben McDonough, Jeb Cui, and Gabriel Marous. Writing this was an incredible learning experience. We found the richest and most helpful source for quantum optics to be Introductory Quantum Optics by Gerry and Knight, which we recommend if you are interested in learning more.” Check out this project's Github!

Digital Humanities: A Gallery of the Coins of Dura-Europos (2023)

Paper by Jeb Cui & project by the Numismatics group in Yale's Digital Humanities course (2022-23) under the guidance of Prof. Holly Rushmeier and Prof. Anne Chen

"There are over 2800 coins from Dura-Europos in the Yale University Art Gallery presented on their website, but it might be difficult to determine that by using their search interface. Numismatics is the study of currency, including coins, and at the site Dura-Europos, numismatics can reveal much about the history of the site, the people who resided at Dura, and their lives. Even though coins can provide such a unique lens into the site and are an important part of the research there, there is no specific tool to search and look through the coins..." Check out this project's Github!

Motion Capture Analysis of Diabolo Juggling (2022)

At TJHSST's Computer Systems Lab under Dr. Gabor

This project analyzes data from the motion capture of diabolo (Chinese yoyo) performances. This builds on the current state-of-the-art diabolo motion predictors using improved analytical models and machine learning. Check out this project's Github!

Novel OVO Strategy for SVMs: Neural Network Classifer (2022)

By Jeb Cui, Raffu Khondaker, and Andy Wang in Dr. Yilmaz's Machine Learning Class at TJHSST

Published in TJHSST Teknos

Abstract: Both support vector machines and neural networks are popular machine learning algorithms for classification. However, a limitation of support vector machines is that it is a “shallow” rather than a “deep” algorithm. “Deep” and “shallow” refers to a model’s number of layers, and “deep” algorithms have more layers and are able to extract more complex relations from data. We sought to improve support vector machine multiclass classification by incorporating neural networks to support vector machines at the classification stage when using the one-versus-one strategy. Our algorithm, the neural network classifier support vector machine, can perform as well and better than other algorithms depending on the structure, size, and characteristics of various classification datasets. Check out this project's Github!