Select a calendar:
Filter October Events by Event Type:
University Calendar
Events for October
-
PhD Thesis Proposal - Umang Gupta
Mon, Oct 03, 2022 @ 03:00 PM - 04:30 PM
Thomas Lord Department of Computer Science
University Calendar
PhD Candidate: Umang Gupta
Title: Controlling Information for Fairness and Privacy in Machine Learning
Committee: Greg Ver Steeg, Paul Thompson, Bistra Dilkina, Kristina Lerman, Fred Morstatter.
Abstract:
With the increasing ubiquity of machine learning models in everyday life, a critical issue occurs when these models capture unintended information. This leads to unintended biases and memorization of training data, resulting in unfair outcomes and risking privacy. These phenomena are especially troublesome in applications where data privacy needs to be upheld, such as medical imaging, or where unfairness can lead to disparate outcomes, such as hiring decisions. To this end, we study this underlying problem of capturing unintended information in various domains. Specifically, we discuss ways to ensure fairness in decision-making by learning fair data representations and controlling unfair language generation by correctly modulating information in neural networks. Finally, we demonstrate that releasing neuroimaging models can reveal private information about the individuals participating in the training set and discuss ongoing work on learning with privacy.
WebCast Link: https://usc.zoom.us/j/96698045892?pwd=UTRDZUNHRTVFS1dieW1URmtEWXZydz09
Audiences: Everyone Is Invited
Contact: Lizsl De Leon
-
PhD Thesis Proposal - Shichen Liu
Wed, Oct 05, 2022 @ 10:00 AM - 11:00 AM
Thomas Lord Department of Computer Science
University Calendar
PhD Candidate: Shichen Liu
Title: Fast and Accurate Geometry Inference with Learned Recurrent Optimizers
Time: Oct. 5th, 2022, 10am - 11am
Committee: Yajie Zhao, Randall Hill, Stefanos Nikolaidis, Andrew Nealen, Aiichiro Nakano
Abstract:
Geometry inference requires accuracy, efficiency, and robustness in various use cases, such as AR/VR, autonomous driving, etc. Compared to traditional methods, deep convolutional neural networks have achieved noticeable improvement on a wide range of geometry inference tasks in terms of robustness. However, these models either produce inaccurate predictions or have a large computation overhead due to specially designed structures, which are hard to be deployed into a real-world system. In this thesis proposal, I introduce a framework that can largely improve the inference speed with high accuracy for geometry inference by taking the vanishing point detection task as a case study.
WebCast Link: https://usc.zoom.us/j/3154287574?pwd=c2ZjeExXR2NpVEtvZjJBd0QrRGVRZz09
Audiences: Everyone Is Invited
Contact: Lizsl De Leon
-
PhD Thesis Proposal - George Papadimitriou
Fri, Oct 07, 2022 @ 12:00 PM - 01:30 PM
Thomas Lord Department of Computer Science
University Calendar
PhD Candidate: George Papadimitriou
Title: Cyberinfrastructure Management for Dynamic Data Driven Applications
Committee:
Ewa Deelman
Rafael Ferreira da Silva
Ramesh Govindan
Victor Prasanna
Aikiro Nakano
Abstract:
Computational science today depends on many complex, data-intensive applications operating on distributed datasets that originate from a variety of scientific instruments and data repositories. Moreover, modern cyberinfrastructure has a highly distributed and heterogeneous character, with compute sites in multiple locations offering access to both commodity hardware and accelerators, but they are not close to the data.
Workflow management systems and scientific workflows are great tools in leveraging the infrastructure given to them and making data available to where a computation will take place. They automate science, making it portable and reproducible. However, until now, even though workflow management systems are aware of the infrastructure they don't control elements of it. Software defined federated cyberinfrastructure has paved the way for workflow management systems to control how the cyberinfrastructure is configured allowing for data access and network optimisations based on the needs of a science application.
In this work we explore approaches and techniques to programmatically configure and control the cyberinfrastructure based on the requirements of an application and we provide solutions in efficiently sharing the allocated network resources among workflow ensembles.
WebCast Link: https://usc.zoom.us/j/94528899681
Audiences: Everyone Is Invited
Contact: Lizsl De Leon
-
PhD Thesis Proposal - Chuizheng Meng
Wed, Oct 19, 2022 @ 03:30 PM - 05:00 PM
Thomas Lord Department of Computer Science
University Calendar
Phd Candidate: Chuizheng Meng
Title: Trustworthy Spatiotemporal Prediction Models
Committee:
Prof. Yan Liu (chair)
Prof. Salman Avestimehr
Prof. Aram Galstyan
Prof. Greg Ver Steeg
Prof. Craig Knoblock
Abstract:
With the great success of data-driven machine learning methods, concerns with the trustworthiness of machine learning models have been emerging in recent years. From the modeling perspective, the lack of trustworthiness amplifies the effect of insufficient training data. Purely data-driven models without constraints from domain knowledge tend to suffer from over-fitting and losing the generalizability on unseen data. Meanwhile, concerns with data privacy further obstruct the availability of data from more providers. On the application side, the absence of trustworthiness hinders the application of data-driven methods in domains such as spatiotemporal forecasting, which involves data from critical applications including traffic, climate, and energy. My thesis proposal constructs spatiotemporal prediction models with enhanced trustworthiness from both the model and the data aspects. For model trustworthiness, the proposal focuses on improving the generalizability of models via the integration of physics knowledge. For data trustworthiness, the proposal proposes a spatiotemporal forecasting model in the federated learning context, where data in a network of nodes is generated locally on each node and remains decentralized. Future works towards the completion of the thesis will target at amalgamating the trustworthiness from both aspects and combine the generalizability of knowledge-informed models with the privacy preservation of federated learning for spatiotemporal modeling.
WebCast Link: https://usc.zoom.us/j/99153030181?pwd=ZGJHK1Zha1VHa2ZVNjRUcUNXaFdPZz09
Audiences: Everyone Is Invited
Contact: Lizsl De Leon