Current browse context:
stat.ML
Change to browse by:
References & Citations
Statistics > Machine Learning
Title: On the inability of Gaussian process regression to optimally learn compositional functions
(Submitted on 16 May 2022 (v1), last revised 27 Sep 2022 (this version, v2))
Abstract: We rigorously prove that deep Gaussian process priors can outperform Gaussian process priors if the target function has a compositional structure. To this end, we study information-theoretic lower bounds for posterior contraction rates for Gaussian process regression in a continuous regression model. We show that if the true function is a generalized additive function, then the posterior based on any mean-zero Gaussian process can only recover the truth at a rate that is strictly slower than the minimax rate by a factor that is polynomially suboptimal in the sample size $n$.
Submission history
From: Matteo Giordano [view email][v1] Mon, 16 May 2022 15:42:25 GMT (31kb)
[v2] Tue, 27 Sep 2022 13:02:49 GMT (30kb)
Link back to: arXiv, form interface, contact.