○ unseen · kind paper · level 1 · 0h
Book by Daniel A. Roberts, Sho Yaida & Boris Hanin (Cambridge University Press, 2022). Develops an effective theory approach to deep neural networks: predictions of trained networks are nearly-Gaussian, with the depth-to-width aspect ratio controlling deviations from the infinite-width description. Introduces representation group flow (RG flow) to characterize signal propagation, gives a practical solution to exploding/vanishing gradients by tuning networks to criticality, and classifies activation functions into universality classes.