Check back soon for more information on the computer science seminar series. Unless otherwise noted, the seminars meet on Mondays at 4pm in Stanley Thomas 302. If you would like to receive notices about upcoming seminars, you can subscribe to the announcement listserv.
Demi Qin and Akshay Mehra | Computer Science PhD Students, Tulane University
These presentations will be delivered online. You may access the presentations on Monday, October 19th, at 4:00 pm CST via the following link: https://tulane.zoom.us/j/95685637204 . Meeting ID: 956 8563 7204. Please be sure to mute your microphone when you log on.
Abstract: Topological data analysis (TDA) is attracting increasing interest among researchers in machine learning due to the power of capturing shapes and structure in data. In this talk, we particularly consider biopsy image classification of prostate cancer with TDA that can utilize the topological summaries of images in machine learning tasks. We begin with the theoretical background of TDA and show our previous work on prostate cancer diagnosis by applying TDA in machine learning applications. Next, we define two aspects to improve the use of TDA: 1) A parallel computation pipeline of our previous work; 2) Comparing distance metric on topological summaries. Our results give new insights on when topological summaries could be more suitable and can be used to design better feature-based learning models with TDA.
Abstract: Bilevel optimization problems are at the center of several important machine learning problems such as hyperparameter tuning, learning with noisy labels, meta-learning, and adversarial attacks. In this presentation, I will talk about our algorithm for solving bilevel problems using the penalty method and discuss its convergence guarantees and show that it has linear time and constant space complexities. Small space and time complexities of our algorithm make it an effective solver for large-scale bilevel problems involving deep neural networks. I will present results of the proposed algorithm on data denoising, few-shot learning, and data poisoning problems in a large-scale setting and show that it outperforms or is comparable to previously proposed algorithms based on automatic differentiation and approximate inversion in terms of accuracy, run-time and convergence speed.