Machine unlearning for generative AI

This paper introduces a new field of AI research called machine unlearning and examines the challenges and approaches to extend machine unlearning to generative AI (GenAI). Machine unlearning is a model-driven approach to make an existing artificial intelligence (AI) model unlearn a set of data from its learning. Machine unlearning is becoming important for businesses to comply with privacy laws such as General Data Protection Regulation (GDPR) customer’s right to be forgotten, to manage security and to remove bias that AI models learn from their training data, as it is expensive to retrain and deploy the models without the bias or security or privacy compromising data. This paper presents the state of the art in machine unlearning approaches such as exact unlearning, approximate unlearning, zero-shot learning (ZSL) and fast and efficient unlearning. The paper highlights the challenges in applying machine learning to GenAI which is built on a transformer architecture of neural networks and adds more opaqueness to how large language models (LLM) learn in pre-training, fine-turning, transfer learning to more languages and in inference. The paper elaborates on how models retain the learning in a neural network to guide the various machine unlearning approaches for GenAI that the authors hope can be built upon their work. The paper suggests possible futuristic directions of research to create transparency in LLM and particularly looks at hallucinations in LLMs when they are extended to do machine translation for new languages beyond their training with ZSL to shed light on how the model stores its learning of newer languages in its memory and how it draws upon it during inference in GenAI applications. Finally, the paper calls for collaborations for future research in machine unlearning for GenAI, particularly LLMs, to add transparency and inclusivity to language AI.