Wyart connects statistical physics with deep learning by treating intelligence as the discovery of useful large-scale variables hidden inside high-dimensional data. When observations have a hierarchical structure, deeper networks can build successive abstractions more efficiently than shallow models that try to represent every interaction at once.
The discussion links this hierarchy to the curse of dimensionality, language and creativity. Wyart argues that useful generalization depends on finding lower-dimensional structure and recombining learned patterns, rather than memorizing every possible configuration of the world.
He proposes predicting learned latent representations instead of raw sensory detail as a route to greater sample efficiency. The theory is still being tested empirically, and reconstructing rich observations from those abstract representations remains an open generative problem.
Watch on YouTube
