Events Calendar

PhD Thesis Defense - Umang Gupta

PhD Thesis Defense - Umang Gupta

Committee Members: Greg Ver Steeg (Chair), Paul Thompson, Bistra Dilkina, Fred Morstatter

Title: Controlling Information in Neural Networks for Fairness and Privacy

Abstract: As machine learning becomes more prevalent in mission critical domains, the harms of unintended information captured by these models are becoming more apparent. These models can inadvertently introduce biases and memorize training data, leading to potential unfairness, inequitable outcomes, or risking privacy. These phenomena are especially alarming 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. This thesis examines ways to control and limit information in deep learning models, focusing on fairness and privacy. Specifically, we discuss ways to ensure fairness in decision making by learning fair data representations and preventing unfair language generation by correctly modulating information in neural networks. Concerning privacy, we demonstrate that releasing neuroimaging models may reveal private information about the individuals participating in the training set and discuss ways to mitigate these privacy leakages. Among these methods, differential private training is promising as it protects against all possible privacy attacks. However, differential private training can drastically hurt utility since the magnitude of noise in the outputs scales with the model parameters. To this end, we explore techniques to reduce effective model parameters during training.

Audiences: Everyone Is Invited

Contact: Melissa Ochoa

Event Link: https://usc.zoom.us/j/7354464916