| Jul 22, 2023 | In our new paper for ICCV 2023, we explore machine unlearning in the context of transfer learning. We introduce a transfer learning strategy titled Disposible Transfer Learning (DTL). DTL filters out extra information during fine-tuning, which is useful for limiting the risk of exposing the pre-trained model when publishing the expert model. Our goal is to help continue open-source AI in the era of foundation models via striking a balance between transparency and security of ownership. |