Learning single-index models with shallow neural networks

Authors: Alberto Bietti, Joan Bruna, Clayton Sanford, Min Jae Song

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

Reproducibility Variable Result LLM Response
Research Type Experimental We illustrate our theoretical results with experiments in Section A.
Researcher Affiliation Academia Alberto Bietti New York University Joan Bruna New York University Clayton Sanford Columbia University Min Jae Song New York University
Pseudocode Yes The overall approach is described in Procedure 1. ... Procedure 1 Gradient Flow ... Procedure 2 Fine-Tuning
Open Source Code Yes Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] Supplemental zip file.
Open Datasets No We focus on regression problems under a single-index model with Gaussian input data. Specifically, we assume d-dimensional inputs x γd := N(0, Id), and labels y = F (x) + = f (h , xi) + , where 2 Sd 1 and N(0, σ2) is an independent, additive Gaussian noise.
Dataset Splits No The paper discusses the theoretical sample size 'n' and uses synthetic data, but does not specify explicit training, validation, or test dataset splits needed for reproduction.
Hardware Specification Yes We ran all experiments on an Intel Core i9-9900K CPU @ 3.60GHz with 64GB RAM and a NVIDIA GeForce RTX 2080 Ti.
Software Dependencies No We generate synthetic data as specified in Section 3 and train our models using PyTorch. (No specific version numbers are provided for PyTorch or other software.)
Experiment Setup Yes The network has N = 1000 hidden units with biases drawn from N(0, 2) with = 1.2. We use the Adam optimizer with learning rate = 0.01, and default PyTorch parameters for momentum. We train for 100 epochs, with batch size 100. For the experiments in Figure 1, we set the dimension d = 50. For Figure 2, we set d = 100, N = 2000, batch size = 200.