Overcoming the curse of dimensionality with Laplacian regularization in semi-supervised learning

Authors: Vivien Cabannes, Loucas Pillaud-Vivien, Francis Bach, Alessandro Rudi

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Figure 3: (Left) Comparison between our kernelized Laplacian method ... and graph-based Laplacian... We report classification error as a function of the number of samples n. The error is averaged over 50 trials... Our method discovers the structure of the data much faster than graph-based Laplacian...
Researcher Affiliation Academia Vivien Cabannes ENS INRIA PSL Paris, France ... Loucas Pillaud-Vivien EPFL Lausanne, Switzerland ... Francis Bach ENS INRIA PSL Paris, France ... Alessandro Rudi ENS INRIA PSL Paris, France
Pseudocode Yes Algorithm 1: Empirical estimates based on spectral filtering.
Open Source Code Yes Our code is available online at https://github.com/Vivien Cabannes/partial_labelling.
Open Datasets No We fixed the ratio nℓ/n to one tenth, and generated the data according to two Gaussians in dimension d = 10 with unit variance and whose centers are at distance δ = 3 of each other. Explanation: The paper describes how data was generated but does not provide concrete access information (link, citation, repository) for a publicly available dataset, nor does it specify that this generated data is made publicly available.
Dataset Splits No We report classification error as a function of the number of samples n. The error is averaged over 50 trials, with errorbars representing standard deviations. Explanation: The paper does not specify percentages or counts for training, validation, or test sets. It only mentions the total number of samples and the ratio of labeled to unlabeled points.
Hardware Specification Yes When dealing with 1000 points, our algorithm, as well as graph-based Laplacian, can be computed in about one tenth of a second on a 2 GHz processor, while the naive kernel implementation requires 10 seconds.
Software Dependencies No The paper mentions 'Implemented with Lapack' but does not provide any specific version numbers for software dependencies.
Experiment Setup Yes Figure 3: (Left) Comparison between our kernelized Laplacian method (Tikhonov regularization version with λ = 1, µn = n 1, p = 50) and graph-based Laplacian based on the same Gaussian kernel with bandwidth σn = n 1 d+4 log(n) as suggested by graph-based theoretical results [24].