Error Bounds for Learning with Vector-Valued Random Features

Authors: Samuel Lanthaler, Nicholas H. Nelsen

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To study how our theory holds up in practice, we numerically implement the vector-valued RF-RR algorithm on a benchmark operator learning dataset. Figure 2a shows the decay of the relative squared test error as M increases (with λ 1/M) for fixed N.
Researcher Affiliation Academia Samuel Lanthaler California Institute of Technology slanth@caltech.edu Nicholas H. Nelsen California Institute of Technology nnelsen@caltech.edu
Pseudocode No The paper contains mathematical proofs and derivations, but no pseudocode or algorithm blocks are provided.
Open Source Code Yes All code used to produce the numerical results and figures in this paper are available at https://github.com/nickhnelsen/error-bounds-for-vv RF.
Open Datasets Yes The data {(un, G(un))}N n=1 is noise-free, the {un} are iid Gaussian random fields, and G : L2(T; R) L2(T; R) is a nonlinear operator defined as the time one flow map of the viscous Burgers equation on the torus T. The initial conditions un ν are sampled iid from a centered Matérn-like Gaussian process according to [23, Sect. 6.3, p. 32].
Dataset Splits No The paper mentions a test set of size N=500 and a training input set, but does not specify the size of the training set or the splitting methodology for the training/validation/test sets, nor does it mention a validation set.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory) used for running experiments are provided.
Software Dependencies No The paper mentions the use of ELU, but does not specify any software names with version numbers for libraries, frameworks, or programming languages used in the experiments.
Experiment Setup Yes The random features are defined in a similar way to the Fourier Space RFs in [32, Sect. 3.1, p. 15]: φ(u(0); θ) = 2.6 ELU F 1{1(|k| kmax)χk (Fu(0))k (Fθ)k}k Z and θ µ , where µ is also a centered Matérn Gaussian measure with covariance operator 1.82( d2 dx2 +152 Id) 3. In the above display, F maps a function to its Fourier series coefficients, and F 1 expresses a Fourier coefficient sequence as a function expanded in the Fourier basis. The filter χ is given by [32, Eqn. 3.6, p. 15] with δ = 0.32 and β = 0.1. We take kmax = 64. In Figures 2a and 3a, the regularization factor is chosen as λ = 7 10 4/M and as λ = 3 10 6/ N in Figures 2b and 3b.