- Understanding Generative Modeling Principles: Simple and general methods for generative modeling, especially over timeseries data.
- World Models in the Loop for Decision Making: The ability to predict what will happen in the future implies understanding the world itself. This capability is necessary to make complex decisions in a complex world. How can world models learn from data within a virtuous loop, optimal for downstream decision making?
- Modeling Uncertainty About the Future: Generative models should explicitly model inherent uncertainty about the future, instead of just sampling one possible future. A simple example: if my opponent will move left with 3/4 chance, and right with 1/4 chance, then I can make a decision based on that distribution. Learning to do this from data is an open challenge.
- Memory-Conditioned Generation: Simple principles of object permanence and persistent information are elusive to today’s generative models, due to a lack of sophisticated memory formation and recall capabilities. How can generative models learn a representation space optimal for storing only what is necessary to remember, and learn to fill their own memories?
- Compositional generative modeling: Humans understand the world’s factorial components, able to recombine them to solve new examples never seen before. From a raw dataset, how can these factors be discovered and then composed to solve out of distribution generalization problems?
- Other topics: Emergent hierarchy, visual representations, evolution, efficient coding, theoretical neuroscience.