ECE PhD Dissertation Defense: Kai Li

PhD Dissertation Defense: Robust Visual Learning with Limited Labels

Kai Li

Location: Zoom Link

Abstract: The recent flourish of deep learning in various tasks is largely credited to the rich and accessible labeled data. Nonetheless, massive supervision remains a luxury for many real-world applications: It is costly and time-consuming to collect and annotate a large amount of training data. Sometimes it is even infeasible to get large training datasets because for certain tasks only a few or even no examples are available, or annotating requires expert knowledge.
In this dissertation research, I investigate techniques systematically addressing the problem of learning with limited labels from the following three aspects. The first aspect is learning to generalize from limited label supervision. I develop few-shot learning algorithms that perform data augmentation in the feature space and that generate task-specific networks based on the limited supervision provided. The second aspect is learning to reuse label supervision from a relevant but different task. I propose domain adaptation algorithms that adapt label supervision from a richly-labeled source domain to a scarcely-labeled target domain with consistency learning, data augmentation and adversarial learning. The last aspect is learning representations without label supervision. I develop algorithms that learn semantic-rich representations that allow to reliably establish relations among high-dimensional data. This is achieved by explicitly modeling the intrinsic relationship among data points during the representation learning process.