Split-kl and PAC-Bayes-split-kl Inequalities for Ternary Random Variables

Authors: Yi-Shan Wu, Yevgeny Seldin

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We present an extensive set of experiments, where we first compare the kl, Empirical Bernstein, Unexpected Bernstein, and split-kl inequalities applied to (individual) sums of independent random variables in simulated data, and then compare the PAC-Bayes-kl, PAC-Bayes-Unexpected-Bersnstein, PAC-Bayes-split-kl, and, in some of the setups, PAC-Bayes-Empirical-Bennett, for several prediction models on several UCI datasets.
Researcher Affiliation Academia Yi-Shan Wu University of Copenhagen yswu@di.ku.dk Yevgeny Seldin University of Copenhagen seldin@di.ku.dk
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes New code in the supplementary.
Open Datasets Yes We evaluate the performance of the PAC-Bayes-split-kl inequality in linear classification and in weighted majority vote using several data sets from UCI and Lib SVM repositories [Dua and Graff, 2019, Chang and Lin, 2011].
Dataset Splits Yes If we split S into two equal parts, S = S1 S2, we can use S1 to train both a reference prediction rule h S1 and a prior πS1, and then learn a PAC-Bayes posterior on S2, and the other way around. By combining the 'forward' and 'backward' approaches we can write Eρ[L(h)] = 1 2Eρ[ L(h, h S1)] + 1 2Eρ[ L(h, h S2)] + 1 2 (L(h S1) + L(h S2))
Hardware Specification Yes All experiments were performed on a local server equipped with an Intel Core i9-9900K CPU and an NVIDIA GeForce RTX 2080 Ti GPU.
Software Dependencies No The paper mentions software like TensorFlow and optimization algorithms like Adam and Rprop, but does not provide specific version numbers for these software dependencies (e.g., 'TensorFlow 2.x' or 'Rprop vX.Y.Z').
Experiment Setup Yes In the experiments we take δ = 0.05, and truncate the bounds at 1. For the Unexpected Bernstein bound we take a grid of γ {1/(2b), , 1/(2kb)} for k = log2( p n/ ln(1/δ)/2) and a union bound over the grid, as proposed by Mhammedi et al. [2019]. For the split-kl bound we take µ to be the middle value, 0, of the ternary random variable.