Learning in the Presence of Low-dimensional Structure: A Spiked Random Matrix Perspective

Authors: Jimmy Ba, Murat A. Erdogdu, Taiji Suzuki, Zhichao Wang, Denny Wu

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we empirically validate the theoretical results presented in previous sections. We construct target functions f with a link of varying degree p and information exponent k and experimentally compute the prediction risk of KRR and that of a two-layer NN.
Researcher Affiliation Collaboration 1University of Toronto, 2Vector Institute, 3x AI, 4University of Tokyo, 5RIKEN AIP, 6University of California San Diego, 7New York University, 8Flatiron Institute
Pseudocode Yes This procedure is summarized as follows (see Algorithm 1 in Appendix for more details).
Open Source Code No The paper does not contain any explicit statements or links indicating that the source code for the described methodology is publicly available.
Open Datasets No We consider the problem of learning a single-index target function f : Rd R under the spiked covariance data: f (x) = σ 1 1+θ x, µ , x N(0, Id + θµµ ), θ dβ for β [0, 1),... where the goal is to estimate the target function (teacher) f , which is a single-index model depending on the signal direction µ Rd, and σ : R R is an unknown link function.
Dataset Splits No The paper operates within a 'proportional asymptotic limit' (n, d -> infinity, n/d -> psi (0, infinity)) and uses synthetically generated data. It does not mention explicit train/validation/test splits by percentages or sample counts, nor does it refer to predefined splits from a standard public dataset. While it evaluates prediction risk, the concept of a dataset split in the conventional empirical sense is not detailed.
Hardware Specification No The paper does not specify any hardware details (e.g., GPU models, CPU types, or memory) used for running the experiments described in Section 5.
Software Dependencies No The paper does not provide specific version numbers for any software dependencies or libraries used in its experiments.
Experiment Setup Yes We set ψ = 5. For KRR we use the Gaussian RBF kernel and λ = 10^-2. For two-layer NN, we set η = 1/√N.