bibkey: “jacot2018neural” authors: “Arthur Jacot; Franck Gabriel; Clément Hongler” year: 2018 title: “Neural Tangent Kernel: Convergence and Generalization in Neural Networks” doi: null claim: “The neural tangent kernel describes output gradient-flow dynamics; a constant limiting kernel requires infinite-width and regularity assumptions.” strata_touched: [] license: “citation-only” triage: “anchor” url: “https://arxiv.org/abs/1806.07572”

Neural Tangent Kernel: Convergence and Generalization in Neural Networks

Verified locator

arXiv:1806.07572v4, Theorem 2, pages 6 and 13. First submitted 20 June 2018; version 4 is 10 February 2020. The theorem uses a Lipschitz, twice differentiable activation with bounded second derivative, control of the time integral of the training direction, sequential width limits, and a finite time interval. It does not make an arbitrary finite network kernel constant or prove closure of output and kernel together.

Declared identifiers: https://arxiv.org/abs/1806.07572.