A unified geometric and thermodynamic account of deep linear network training, combining invariant manifolds, an exact Boltzmann entropy, and Riemannian Langevin dynamics.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.NE 1years
2024 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
The geometry of the deep linear network
A unified geometric and thermodynamic account of deep linear network training, combining invariant manifolds, an exact Boltzmann entropy, and Riemannian Langevin dynamics.