Fast structure learning with modular regularization

Authors: Greg Ver Steeg, Hrayr Harutyunyan, Daniel Moyer, Aram Galstyan

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments on synthetic data demonstrate that our approach is the only method that recovers modular structure better as the dimensionality increases. We also use our approach for estimating covariance structure for a number of real-world datasets and show that it consistently outperforms state-of-the-art estimators at a fraction of the computational cost.
Researcher Affiliation Academia Greg Ver Steeg Information Sciences Institute University of Southern California Marina del Rey, CA 90292 gregv@isi.edu Hrayr Harutyunyan Information Sciences Institute University of Southern California Marina del Rey, CA 90292 hrayrh@isi.edu Daniel Moyer Information Sciences Institute University of Southern California Marina del Rey, CA 90292 moyerd@usc.edu Aram Galstyan Information Sciences Institute University of Southern California Marina del Rey, CA 90292 galstyan@isi.edu
Pseudocode Yes Algorithm 1 Linear Cor Ex. Implementation is available at https://github.com/hrayrhar/T-Cor Ex.
Open Source Code Yes Algorithm 1 Linear Cor Ex. Implementation is available at https://github.com/hrayrhar/T-Cor Ex.
Open Datasets Yes We considered the weekly percentage returns for U.S. stocks from January 2000 to January 2017 freely available on http://quandl.com. ... To avoid cherry-picking, we selected datasets from Open ML [17] according to the following criteria: between 100 and 11000 numeric features, at least twenty samples but fewer samples than features (samples with missing data were excluded), and the data is not in a sparse format. ... applying it with 100 latent factors on the resting-state f MRI of the first session (session id: 014) of the publicly available My Connectome project [19].
Dataset Splits Yes In all experiments hyper-parameters are selected from a grid of values using a 3-fold cross-validation procedure. ... We use an 80-20 train-test split, learning a covariance matrix from training data and then reporting the negative log-likelihood on test data.
Hardware Specification Yes The experiment was done in the setting of Sec. 4.1 on an Intel Core i5 processor with 4 cores at 4Ghz and 64Gb memory. We used Nvidia RTX 2080 GPU when running the proposed method on a GPU.
Software Dependencies No We implement the optimization problem (3) in Py Torch and optimize it using the ADAM optimizer [8]. No specific version numbers for software dependencies (e.g., libraries, frameworks) were provided in the main text.
Experiment Setup Yes In all experiments hyper-parameters are selected from a grid of values using a 3-fold cross-validation procedure. ... The annealing procedure consists of rounds, where at each round we pick a noise amount, η ∈ [0, 1], and in each iteration of that round replace X with its noisy version, X, computed as follows: X = (1 − η2)X + ηE, with E ∼ N(0, Ip). ... We do 7 rounds with the following schedule for η, [0.61, 0.62, . . . , 0.66, 0]. ... For factor-based methods including our own, we chose the number of factors from the set m ∈ {5, 20, 50, 100} using 3-fold cross-validation.