Exploring Patterns in Number Theory with Deep Learning - A Case Study with Möbius and Squarefree Indicator Functions

Watch a 25-minute mathematics lecture from Harvard CMSA's Mathematics and Machine Learning Workshop where David Lowry-Duda from ICERM explores the application of deep learning techniques to number theory patterns, specifically focusing on the Möbius and squarefree indicator functions. Learn about experiments using neural networks and the Int2Int transformer to predict these mathematical functions, with particular attention to how the Möbius function relates to the Riemann zeta function's reciprocal coefficients. Discover how different input representations and model variations affect prediction accuracy and gain insights into the explainable and mysterious aspects of the model's predictive capabilities in understanding these fundamental number theory concepts.