Correlation flow governs Learning at criticality Source:
On the importance of initialisation Source:
Network setup and dataset Source: Beamer: Set-up and preliminaries; article: Dataset and Figure 1
Quantities of interest Source: Beamer: Quantities of Interest; article: covariance, correlation, Jacobians and NTK
Contributions Source:
The mean-field limit Source: Beamer: Mean-field limit; article: Gaussian pre-activations and covariance recursion
Propagation regimes Source: Phase boundary for the sine network; ordered and chaotic experiments
Vanishing variance and criticality Source: Correlation_flow: Taylor expansion and critical variance decay
Variance flow at criticality Source: Correlation_flow/figures/NNGP.pdf: pre-activation distribution and kernel propagation
Flow of correlations at criticality Source: Correlation_flow/sections/correlation.tex: correlation map and activation condition
Correlation flow at criticality Source: Correlation_flow/sections/correlation.tex: Fig. 6b regimes and parameters
The neural tangent kernel Source: Beamer: The NTK, equations (13)--(15); Correlation_flow: NTK basics
Conclusion Source: Correlation_flow: NTK similarity structure conclusion
Discussion Source: Correlation_flow: scope and open questions
Acknowledgments Source:
Vanishing variance: derivation Source: Beamer: Universal description for the vanishing case; Correlation_flow: activation Taylor expansion and covariance recursion