ETD

• computer vision
• image-free
• machine unlearning
• unified framework
• weight learning
• zero-shot learning

This thesis investigates model weight learning as a data modality to support different downstream tasks in the computer vision context. By treating
model weights as data, we explore their potential to adapt knowledge for
specific tasks. We first apply this perspective to the zero-shot learning problem, showing how model weights can be used to produce new weights for
an image classifier, enabling the classification of new/unseen classes. Next,
we explore the problem of machine unlearning, where weight-based strategies allow for efficient and targeted forgetting of specific classes without
retraining from scratch. Finally, we draw a future direction for a general
framework to jointly address both zero-shot learning and unlearning.