Nonasymptotic Guarantees for Spiked Matrix Recovery with Generative Priors

Authors: Jorio Cocola, Paul Hand, Vlad Voroninski

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We further corroborate these findings by proposing a (sub)gradient algorithm which, as shown by our numerical experiments, is able to recover the sought spike with optimal sample complexity.
Researcher Affiliation Collaboration Jorio Cocola Department of Mathematics Northeastern University Boston, MA 02115 cocola.j@northeastern.edu Paul Hand Department of Mathematics and Khoury College of Computer Sciences, Northeastern University Boston, MA 02115 p.hand@northeastern.edu Vladislav Voroninski Helm.ai, Menlo Park, CA 94025 vlad@helm.ai
Pseudocode Yes Algorithm 1 Gradient method for the minizimization problem (4)
Open Source Code No The paper does not provide any specific links or explicit statements about releasing the source code for the methodology described.
Open Datasets No The paper describes generating 'synthetic generative priors' and 'randomly sample the weights of the matrix' and 'consider data Y according the spiked models (1) and (2)', rather than using a publicly available or open dataset. Therefore, no information on public dataset access is provided.
Dataset Splits No The paper performs numerical experiments on synthetically generated data, varying parameters like 'samples N' and 'noise level ν'. It does not describe standard training, validation, and testing splits from a fixed dataset, nor does it specify exact percentages or counts for such splits.
Hardware Specification No The paper mentions running 'numerical experiments' but does not provide any specific details about the hardware (e.g., CPU, GPU models, memory) used for these experiments.
Software Dependencies No The paper does not list specific software dependencies with version numbers (e.g., Python, PyTorch, specific libraries).
Experiment Setup Yes We consider 2-layer generative networks with Relu activation functions, hidden layer of dimension n1 = 250, output dimension n = 1700 and varying number of latent dimension k [10, 30, 70]. We randomly sample the weights of the matrix independently from N(0, 2/ni), which removes that 2d dependence in Theorem 2. We then consider data Y according the spiked models (1) and (2), where x Rk is chosen so that y = G(x ) has unit norm. For the Wishart model we vary the samples N while for the Wigner model we vary the noise level ν so that the following quantities remain constant for the different networks (latent dimension k).