We will delve into the question: How do we enhance the robustness of Deep Neural Networks amidst shifting data distributions in online learning systems? Jehanzeb will present several approaches for test-time tuning for various domains such as images, videos, and point clouds. He will also touch upon methods developed for adapting vision-language models to improve their zero-shot classification abilities, bringing them closer to dedicated (closed category set) classifiers trained with supervised fine-tuning.
Bio: “I am a third year Computer Vision PhD. student advised by Professor Horst Bischof, at TU Graz, Austria. My PhD. research is mainly focused on designing self-supervised and unsupervised representation learning techniques for making deep neural networks robust to distribution shifts on-the-fly, at test-time. During numerous research projects, I have worked with different types of data including images, point clouds, videos, radar signals and most recently natural language. Along with the main research focus on online learning, recently I have also started working with large language models (LLMs) and I am particularly interested in multi-modal (vision-language) models.”
Add comment