Thu, Apr 27, 2023 @ 10:00 AM - 12:00 PM
Thomas Lord Department of Computer Science
PhD Thesis Proposal - Iordanis Fostiropoulos
Committee: L. Itti (Chair), M. Soleymani, S. Nikolaidis, N. Schweighofer (Outside Member)
Title: Towards Learning Generalizable Representations
Abstract: Current work in Machine Learning (ML) research lack systematic tools and methods for evaluating the performance of a ML model on the ability to generalize beyond the train set; where the current accepted practice is on the evaluation of the loss on a test set. Work in ML for defining generalization is abstract and based on anthropocentric measures. While practical metrics in evaluating generalization are poor indicators where there are trade-offs between the metric (such as loss) and the performance of the Deep Neural Network (DNN) to Out-of Distribution examples, such as robustness-accuracy trade-off or hallucinations of transformer models. While algorithmic solutions are often in the form of paradigm shifts that are ad-hoc and domain specific with a lack of consensus in literature. Our work focus on generalization as it pertains on evaluating and improving current ML systems, as opposed to proposing a paradigm shift, where we address three evaluation settings of generalization. First, the generalization of a DNN to learn generalizable representations useful beyond the task it was trained on. Second, the generalization of the learning hyper parameters used to fit a DNN; a meta-model. Third, the learning algorithm generalization, where we evaluate generalization in the context of Continual Learning. We present our work on the analysis and theoretical findings on the short-comings of generalization and provide practical solutions that both confirm and can in-part address the issue. We motivate that the problem of generalization extend well beyond the three areas our work addresses where improvements in algorithms, tools, and methods are required. Finally, based on our empirical observations we discuss several future directions for improving generalization in ML systems.
Audiences: Everyone Is Invited
Contact: Melissa Ochoa