On Learning High Dimensional Structured Single Index Models

Authors: Ravi Ganti, Nikhil Rao, Laura Balzano, Rebecca Willett, Robert Nowak

AAAI 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that the proposed method enjoys superior predictive performance when compared to generalized linear models, and achieves results comparable to or better than single layer feedforward neural networks with significantly less computational cost. We now compare and contrast our method with several other algorithms, in various high dimensional structural settings and on several datasets.
Researcher Affiliation Collaboration 1 Walmart Labs, San Bruno, CA gmravi2003@gmail.com 2 Technicolor Research and Innovation, Los Altos, CA 3 EECS, University of Michigan, Ann Arbor, MI 4 ECE, University of Wisconsin, Madison, WI
Pseudocode Yes Algorithm 1 CSI
Open Source Code No The paper does not explicitly state that the source code for the proposed CSI method is available, nor does it provide a link to it. It only mentions the use of third-party libraries like L1-General, MALSAR, and TensorFlow.
Open Datasets Yes We tested the algorithms on several datasets: link and page are datasets from the UCI machine learning repository. We also use four datasets from the 20 newsgroups corpus: atheism-religion, autos-motorcycle, cryptography-electronics and mac-windows. For multilabel learning, the flags dataset... For multitask learning, the atp7d dataset... We consider two datasets, Epinions and Flixster
Dataset Splits Yes We always perform a 50 25 25 train-validation-test split of the data, and report the results on the test set. Out of the training set, we randomly set aside 10% of the measurements for validation. We perform a random 80 10 10 split of the data for training, validation and testing.
Hardware Specification No The paper does not explicitly describe the specific hardware (e.g., GPU/CPU models, memory) used to run the experiments. It mentions TensorFlow but without hardware details.
Software Dependencies No The paper mentions software like MATLAB, L1-General library, MALSAR package, TensorFlow, and Adam optimizer, but does not provide specific version numbers for these software components.
Experiment Setup Yes We varied the regularization parameter over {2 10, 2 9, , 29, 210}. ... We varied the sparsity of the solution as {d/4, d/8, d/16, , d/1024} ... number of hidden units were varied between 5 and 1000. ... The settings used are: learning rate=0.1, beta1=0.9, beta2=0.999, epsilon=1e-08, use locking=False ... We varied the step size η [2 6, 22] on a log scale for our method, setting the group sparsity parameter to be 5 for both datasets.