Iterative Hard Thresholding with Adaptive Regularization: Sparser Solutions Without Sacrificing Runtime

Authors: Kyriakos Axiotis, Maxim Sviridenko

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section we present numerical experiments in order to compare the performance of IHT and regularized IHT (Algorithm 1) in training sparse linear models.
Researcher Affiliation Collaboration 1MIT 2Yahoo! Research.
Pseudocode Yes Algorithm 1 Regularized IHT
Open Source Code Yes In Figure 2 we have python implementations of the IHT and regularized IHT algorithms that we use for our experiments. As can be seen, both implementations are pretty short.
Open Datasets Yes We first experiment with real data, specifically the year regression dataset from UCI (Dua & Graff, 2017) and the rcv1 binary classification dataset (Lewis et al., 2004), which have been previously used in the literature.
Dataset Splits No The paper describes using real and synthetic data for experiments and preprocessing steps, but does not provide specific details on train/validation/test dataset splits (e.g., percentages, sample counts, or explicit mention of validation sets).
Hardware Specification Yes All the experiments were run on a single 2.6GHz Intel Core i7 core of a 2019 Mac Book Pro with 16GB DDR4 RAM using Python 3.9.10.
Software Dependencies Yes All the experiments were run on a single 2.6GHz Intel Core i7 core of a 2019 Mac Book Pro with 16GB DDR4 RAM using Python 3.9.10.
Experiment Setup Yes For our experiments we use ρ = 0.1. For the weight step size of regularized IHT, we set the weight step size to c = s /T... for each algorithm we pick the best fixed step size of the form 2i/s for integer i 0, where s is the fixed sparsity level. The best step sizes of IHT and regularized IHT end up being 2/s, 4/s respectively for the linear regression example, and 8/s, 16/s respectively for the logistic regression example.