Interpretable Concept Decomposition Enables Precise Machine Unlearning in VLMs

A novel approach known as ICED (Interpretable Concept-level Unlearning via Interpretable Concept Decomposition) tackles the issue of machine unlearning in Vision-Language Models (VLMs). Existing unlearning techniques typically function at the image or instance level, often inadequately eliminating specific knowledge without disrupting unrelated semantics, particularly when a single image encompasses multiple intertwined concepts. ICED utilizes a multimodal large language model to create a concise, task-specific vocabulary from the forgetting set. It then breaks down visual representations into sparse, nonnegative combinations of these semantic concepts, allowing for precise knowledge adjustments. The unlearning process is framed as concept-level optimization, focusing on the concepts to be discarded while maintaining contextual integrity. The paper can be found on arXiv with the reference 2605.14309.