2719: Redefining AI's Learning Curve: The Art of Unlearning

Today, I welcome Zuzanna Stamirowska, CEO and co-founder of Pathway, a pioneering French tech startup. Zuzanna, celebrated for her achievements including the Female Founders Challenge award at VivaTech, joins us to explore the innovative concept of 'unlearning' in AI and its transformative implications. Unlearning in AI, as Zuzanna elucidates, involves the ability of AI models to delete specific data points and their effects without undergoing a complete retraining. This advancement is particularly crucial in the realm of linear models like trees, compared to the complexities of neural networks. The concept goes beyond mere technical innovation; it's a step towards implementing privacy rights like the EU's 'right to be forgotten' and addressing the challenge of catastrophic forgetting in AI. I explore how Pathway leverages unlearning for dynamic operational analytics in sectors such as logistics, legal search, and financial trading. We discuss the unique challenges and solutions in applying unlearning to generative AI, emphasizing its multidimensional nature. An enlightening segment of our conversation focuses on Pathway's mission of enabling real-time learning and data use for data scientists. We explore their recent advancements, including query functionality for AI applications and efforts towards ensuring consistency in real-time systems. This episode also touches on practical applications and client success stories, highlighting how organizations like La Poste have benefited from Pathway's unlearning capabilities, leading to significant improvements in efficiency and data management. As we look ahead, Zuzanna shares insights into Pathway's future directions, aiming to bring streaming data and models to a broader user base, including integrating with tools like Jupyter and expanding conversational AI capabilities.