Google, IBM & Meta Certificates — All 10,000+ Courses at 40% Off
One annual plan covers every course and certificate on Coursera. 40% off for a limited time.
Get Full Access
Explore groundbreaking research that challenges fundamental assumptions about how artificial intelligence systems store and retrieve information. This 35-minute video examines a newly published ArXiv pre-print from researchers at Carnegie Mellon University and Google Research that investigates the competition between associative and geometric memory during AI model optimization. Discover how factors like training time, learning rate, and weight decay influence which memory type dominates in deep sequence models. Learn about the implications for choosing between parametric versus contextual memory approaches and understand the relevance to generative retrieval versus dual encoder retrieval models. Gain insights into why deep sequence models tend to memorize geometrically and what this means for the future development of AI systems, as researchers call for more careful empirical research to make geometric memory views more broadly applicable across different AI architectures.
