Tuesday, November 19, 2024, 9am

This dissertation addresses the problem of enhancing the efficiency and trustworthiness of neural models.

On the efficiency side, my work focuses on the problem of Hyper-parameter tuning (HPT), a crucial but expensive step to optimize the performance of Neural Networks (NN). In this dissertation, I propose two new HPT methods, HyperJump and TrimTuner. Both methods aim to maximize model’s quality while reducing training and optimization time. Despite using different techniques to solve the optimization problem, both rely on low-fidelity observations (e.g., training with sub-sampled datasets) to efficiently identify promising configurations to be then tested via high-fidelity observations (e.g., using the full dataset).

On the trustworthiness side, the focus of this work is on adversarial robustness and uncertainty estimation. On the adversarial robustness front, I investigate the challenges that arise when performing HPT for models that are adversarially trained, showing that, although the complexity of the HPT problem is exacerbated in adversarial settings, tuning the HPs independently for standard and Adversarial Training (AT) can improve accuracy. To reduce HPT costs, I propose leveraging cheap AT methods to obtain inexpensive, yet highly correlated, estimations of the quality of more robust methods. This approach, combined with a recent multi fidelity optimizer, enhances significantly the efficiency of the HPT process. On the uncertainty estimation front, I introduce Error-Driven Uncertainty Aware Training (EUAT), a method that strives to ensure that the model is highly uncertain when making inaccurate predictions and confident when making accurate ones. During training EUAT selectively employs two loss functions based on the model’s performance on training examples: i) it reduces uncertainty for correct predictions and ii) increases uncertainty for mispredictions, while preserving the accuracy.

This current proposal presents the research results already achieved so far and plans for future work in order to complete my dissertation.

Thesis Committee
David Garlan (Chair)
Paolo Romano (Co-Chair, Instituto Superior Técnico, University of Lisbon)
Matt Fredrikson
Steven Wu
Chrysoula Zerva (Instituto Superior Técnico, University of Lisbon)

Additional Information

In Person and Zoom Participation. See announcement.

Event Type: Thesis Proposals
Room Number: In Person and Virtual - ET
Building: TCS Hall 358 and Zoom
Speaker's Name: PEDRO GONÇALO BRAVO MENDES
Speaker Website: web.tecnico.ulisboa.pt…
Speaker's Professional Title: Ph.D. Student, Ph.D. Program in Software Engineering, Software and Societal Systems Department, Carnegie Mellon University
Talk Title: Techniques for Enhancing the Efficiency and Trustworthiness of Neural Networks
Event Poster Title: Poster
Event Poster URL: www.cs.cmu.edu…
For More Information: aroudebu@andrew.cmu.edu
Affiliations: Software and Societal Systems Department (S3D)
Organization(s): School of Computer Science