57 min
Seminar, Hosted by Toyota Technological Institute at Chicago (TTIC)
Immunization and Test-Time Augmentation for Pre-Trained Computer Vision Models
Purdue University
Recording
Abstract
With the spread of open-source pre-trained models, computer vision systems are now built largely by starting from such models and then fine-tuning or adapting them. This talk examines two aspects of pre-trained models: mitigating their potential misuse, and improving their performance at test time. It presents two recent works addressing these topics: IMMA, which immunizes text-to-image models against malicious adaptation, and a study showing that deep networks with subsampling layers unintentionally discard useful activations at test time.