ML Projects

Stochastic Optimization of Reinforcement Learning

Currently working on developing novel optimization method for RL in the average reward setting.

Superresolution

Advisors: Dr. Cynthia Rudin and Dr. Aayush Bansal

Teammate: Jerry Liu

Worked on patch-based method for portrait superresolution using interpretable representation learning. Focused on clustering patches of face (eyes, mouth, nose, etc.) to train a model on each cluster to learn more patch-specific details.

Automating 2D Classification in Cryo-Em Single-Particle Reconstruction Pipeline

Advisor: Dr. Alberto Bartesaghi

Teammate: Jeevan Tewari

Course: Duke CS 590: Computational Cryo-EM 3D Imaging

Research project for final assignment. In the Computational Cryo-EM 3D Imaging pipeline, software programs, such as cisTEM, create classes of images based on the particle orientation and provide class averages. A human picks classes with high-quality averages and the class images are used for reconstruction. I devised an approach using unsupervised representation learning to automate the selection of high-quality images.

Neural Style Transfer

Instructor: Dr. Yiran Chen

Course: Duke ECE 590: Computer Engineering Methods for Deep Learning

Final class project where I implemented “Texture Networks: Feed-forward Synthesis of Textures and Stylized Images” by Ulyanov et al. Link to my code on Github.

Neural Network Dimension Reduction with Topological Constraint

Advisor: Dr. Xiuyuan Cheng

Program: DoMath Summer 2020

Using MNIST and synthetic datasets with known manifolds, compared Variatinal Autoencoders to spectrel embedding in preservation of topology in the latent space. Summary of work here

Interpretable vs Black-box Recidivism Models

Advisor: Dr. Cynthia Rudin

Teammates: Caroline Wang and Bin Han

Published in Journal of Quantitative Criminology, 2022

Processed criminal history data from Broward County, FL and the state of Kentucky. My team then assessed the performance and fairness of various interpretable and black-box algorithms. Our findings show that they perform comparably to each other and we thus advocate against the use of black-box and privatized algorithms in pre-trial risk assessments.

Recommended citation: Caroline Wang, Bin Han , Bhrij Patel, Cynthia Rudin (2022). “In pursuit of interpretable, fair and accurate machine learning for criminal recidivism prediction.” Journal of Quantitative Criminology. https://link.springer.com/article/10.1007/s10940-022-09545-w