Foundation Models: Shifting Unlearning from Data to Knowledge Tracing

This position paper proposes a shift from data-tracing to knowledge-tracing for machine unlearning in foundation models. The authors argue that current data-tracing methods are impractical for foundation models, as users often lack access to the massive training datasets. Instead, they suggest unlearning based on the knowledge or capabilities a model should not possess, which aligns better with human cognitive processes of forgetting. The paper also outlines the significant challenges associated with this knowledge-tracing approach and presents a case study using a vision-language foundation model. AI

IMPACT Proposes a new paradigm for unlearning in foundation models, potentially simplifying compliance and aligning with human cognitive processes.

RANK_REASON This is a research paper proposing a new methodology for machine unlearning in foundation models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Data-Tracing Machine Unlearning
• Foundation Models
• Knowledge-Tracing Machine Unlearning
• Machine Unlearning

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →