Machine Learning at the Flatiron Institute Seminar: Mengye Ren

Title: Lifelong and Human-like Learning in Foundation Models

Abstract: Real-world agents, including humans, learn from online, lifelong experiences. However, today’s foundation models primarily acquire knowledge through offline, iid learning, while relying on in-context learning for most online adaptation. It is crucial to equip foundation models with lifelong and human-like learning abilities to enable more flexible use of AI in real-world applications. In this talk, I will discuss recent works exploring interesting phenomena in foundation models when learning in online, structured environments. Notably, foundation models exhibit anticipatory and semantically-aware memorization and forgetting behaviors. Furthermore, I will introduce a new method that combines pretraining and meta-learning for learning and consolidating new concepts in large language models. This approach has the potential to lead to future foundation models with incremental consolidation and abstraction capabilities.