AI model disgorgement: Methods and choices.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Machine learning models, particularly large language models (LLMs) in generative AI, have grown in size and complexity.
• Training these models requires vast datasets and computational resources, making it impractical to retrain them to remove problematic data.
• Manual inspection of massive training corpora is infeasible, leading to potential data defects like protected or private content.

Purpose of the Study:

• To explore model disgorgement as a viable approach to address defects in machine learning training data.
• To define and categorize model disgorgement techniques applicable to contemporary machine learning systems.
• To investigate methods for removing the influence of specific data from trained models without full retraining.

Main Methods:

• Surveying existing literature on model disgorgement techniques.
• Developing a taxonomy to classify different disgorgement strategies.
• Analyzing the concept of 'removing effects' of data on ML models post-training.

Main Results:

• Identified model disgorgement as a practical solution for mitigating issues arising from training data defects.
• Proposed a framework for understanding and applying various disgorgement methods.
• Demonstrated that data influence can be reduced without the need for complete model retraining.

Conclusions:

• Model disgorgement techniques are crucial for ensuring responsible AI development and addressing data integrity concerns in large models.
• The proposed taxonomy provides a structured overview of methods for model disgorgement.
• Further research into efficient and effective disgorgement methods is warranted for modern machine learning systems.