Generalizing Outside the Training Distribution through Compositional Generation

Explore a cutting-edge lecture on compositional generative modeling and its applications in AI generalization. Delve into the concept of energy-based models and their role in enabling compositional generative modeling. Discover how these models can synthesize complex plans for unseen tasks at inference time, pushing the boundaries of AI capabilities. Learn about the application of compositionality to multiple foundation models trained on diverse Internet data, and how this approach enables the construction of decision-making systems capable of hierarchical planning and solving long-horizon problems in a zero-shot manner. Gain insights from Yilun Du, a final year PhD student at MIT CSAIL, as he shares his research spanning machine learning, computer vision, and robotics, with a focus on generative models.

Generalizing Outside the Training Distribution through Compositional Generation: Yilun Du (MIT)