Machine Unlearning (MUL)

Source: The post Machine Unlearning (MUL) has been created, based on the article “Teaching computers to forget” published in “The Hindu” on 30th July 2024

UPSC Syllabus Topic: GS Paper 3-science and technology-Awareness in the fields of IT, Space, Computers, robotics

Context: The article discusses Machine Unlearning (MUL), a method to make AI systems forget incorrect or sensitive data. It explores different ways this can be implemented: privately by companies, publicly by governments, or internationally through global standards. MUL helps manage AI issues like bias and privacy breaches.

What is Machine Unlearning (MUL)?

1. Machine Unlearning (MUL) is a method to make AI systems selectively forget certain types of data, particularly false, discriminatory, outdated, or sensitive information.
2. Origin: The concept was introduced by Cao and Yang in their work “Towards Making Systems Forget with Machine Unlearning.”
3. Purpose: MUL aims to address the complexities and risks involved with AI managing vast amounts of data, which can lead to privacy breaches, misinformation, and AI bias.
4. Application Example: IBM is actively testing MUL models to achieve better accuracy, intelligibility, and cost-efficiency in deleting unnecessary or harmful data from AI systems.

Why is MUL Important?

1. Complexity Management: MUL helps in managing the overwhelming complexity of AI systems that process vast amounts of data, making it difficult to maintain data integrity.
2. Privacy and Bias Issues: It addresses increases in AI bias, misinformation, and privacy breaches, which are particularly problematic during sensitive times like elections.
3. Cost and Efficiency: Implementing MUL can be more cost-effective and efficient than deleting entire datasets and retraining AI models, which is costly and reduces accuracy, as discussed in IBM’s tests on MUL models.

How Can MUL be Implemented?

1. Private Approach:
2. Companies like IBM are testing MUL models to improve how quickly and accurately data can be deleted, while keeping costs low.
3. This approach is voluntary and allows companies to adapt their AI systems without government intervention.
4. However, it might be less feasible for smaller companies due to high costs and required expertise.
5. Public Approach:
6. Governments can legislate the use of MUL. For instance, the European Union’s AI Act uses a soft-law approach to address data poisoning as a cybersecurity issue.

2.Governments could also develop their own MUL models to be used across different platforms, helping especially in developing countries by making such technology more accessible.

1. International Approach:
2. This involves creating international standards for MUL to ensure consistent practices across different countries.
3. This could help in managing the global implications of AI technology. However, geopolitical tensions could complicate this approach.

Way forward

MUL is still in early development, facing technical and regulatory challenges. Stakeholders need to address these issues to ensure that MUL can be effectively implemented to tackle the problems of generative AI and protect users’ rights in the digital age.

Question for practice:

Discuss the importance of Machine Unlearning (MUL) in managing AI systems and how it addresses issues like privacy breaches and bias.