🔬 Why Transformers Hit a Wall the Moment Physics Shows Up — Anima Anandkumar, Caltech
Source
youtube.com
Author
Latent Space
Date
Context
A physical prediction may need a different spatial resolution from the data used to train it. In this research interview, Anima Anandkumar explains neural operators as models that learn relationships between continuous functions, allowing inputs and outputs to be sampled at different resolutions.
Weather provides the example: coarse observations may need to support finer predictions. Anandkumar explains that extending a prediction to finer detail involves assumptions; additional physical constraints, such as conservation laws, can guide it. The useful capability is flexibility across resolutions, with physical information helping constrain the detail that training data alone does not determine.