Computation and Cognition

John Lafferty will welcome attendees and open the conference.

Zhang will discuss the ability of time-series foundation models to zero-shot forecast dynamic systems. The findings reveal that LLMs utilize context parroting as a key mechanism. Exploring zero-shot forecasting strategies beyond parroting can provide insights into how artificial and natural intelligence learn from limited data.

Neural networks optimized with generic learning objectives acquire representations that support remarkable behavioral flexibility. Grant will illustrate how interpreting and analyzing neural networks lays the groundwork for a productive account of how cognitive capacities arise in neural systems.

Light refreshments and coffee will be available in the lounge.

For a learning algorithm to generalize to unseen data, it must choose one of many possible ways to extrapolate beyond existing data, a property known as the inductive bias of the learner. Bordelon examines how neural representation geometry, learning rule, and architecture influence the inductive bias of neural circuits in solvable mathematical models. The findings can be used to guide scaling and design principles for deep learning systems.