Generative AI: Paradigm Shift from Learning, Relearning, and Unlearning

An Insight into the BFSI Sector and Beyond

Dr. Magesh Kasthuri, Distinguished Member of Technical Staff, Wipro Limited

Introduction

Generative Artificial Intelligence (AI) has transformed various fields, from creative industries to complex scientific research. The paradigm shift from traditional learning to relearning and unlearning plays a crucial role in the evolution of generative AI. This article explores these concepts and their significance, particularly focusing on practical use cases in the Banking, Financial Services, and Insurance (BFSI) sectors.

Learning in Generative AI

Learning in generative AI involves training models on vast datasets to generate new content or make predictions. This foundational phase is critical, as it determines the model’s initial capabilities. However, the static nature of initial learning can be limiting as new data and scenarios emerge.

Relearning: Adapting to New Information

Relearning involves updating models with new data, enhancing their performance, and ensuring they remain relevant. In the BFSI sector, relearning is paramount due to the dynamic nature of financial markets and regulatory environments. For example, stock market trends and insurance product updates requires relearning.

Practical Use Cases of Relearning in BFSI

· Fraud Detection: Financial institutions constantly face new fraud tactics. Relearning allows AI models to adapt to these new methods, improving fraud detection accuracy.

· Risk Assessment: As market conditions evolve, AI models need to relearn to provide accurate risk assessments for investments, loans, and insurance policies.

· Customer Service: Relearning helps AI-powered chatbots and virtual assistants to adapt to changing customer queries and preferences, enhancing service quality.

Unlearning: The Importance and Implementation

In Machine learning (ML) world, unlearning is popular which helps to remove target datasets from trained datasets which are not likely needed from the context of evaluation. In Generative AI world, Unlearning is the process of removing outdated or irrelevant information from AI models. This step is crucial to avoid biases and improve the model’s accuracy and performance. When we are focussing on AI fairness, Ethical AI and explainable AI, Unlearning plays an important role in redefining these goals in Generative AI solution design.

How Unlearning is Achieved

Unlearning can be achieved through techniques such as:

· Selective Forgetting: Identifying and removing specific data points that contribute to biases or inaccuracies.

· Model Resetting: Partially or fully resetting the model’s parameters and retraining it with updated data.

· Data Pruning: Systematically eliminating data that no longer serves the model’s purpose or is deemed irrelevant.

Implications of Unlearning

The following table outlines the implications of unlearning for various factors in the context of generative AI model development:

The Concept of a ‘Forget Set’ in Generative AI

In the context of generative AI, a ‘forget set’ refers to a collection of data points or information that needs to be removed from the model. This concept is integral to the unlearning process, as it ensures that irrelevant or outdated information does not skew the model’s output.

Relevance of ‘Forget Set’ in Unlearning for Generative AI LLMs

For large language models (LLMs) in generative AI, the ‘forget set’ is crucial to maintaining the model’s accuracy and relevance. By systematically unlearning specific data points, LLMs can avoid perpetuating biases and inaccuracies, thereby improving their overall performance and reliability.

Conclusion

The paradigm shift from learning to relearning and unlearning is essential for the continuous improvement and relevance of generative AI models. In the BFSI sector, these processes enable institutions to stay ahead of emerging threats and opportunities, ensuring robust and reliable AI applications. Understanding and implementing unlearning, along with managing the ‘forget set’, are critical steps in the evolution of generative AI, paving the way for more accurate, secure, and efficient models.