ResNet with one-neuron hidden layers is a Universal Approximator

Authors: Hongzhou Lin, Stefanie Jegelka

NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We begin by empirically exploring the difference between narrow fully connected networks, with d neurons per hidden layer, and Res Net via a simple example: classifying the unit ball in the plane. Figure 2 shows the results.
Researcher Affiliation Academia Hongzhou Lin MIT Cambridge, MA 02139 hongzhou@mit.edu Stefanie Jegelka MIT Cambridge, MA 02139 stefje@mit.edu
Pseudocode No The paper describes mathematical constructions and proof steps but does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code No The paper does not contain any explicit statement about releasing source code or provide a link to a code repository for the described methodology.
Open Datasets No The training set consists of randomly generated samples (zi, yi)i=1 n R2 { 1, 1} with yi = 1 if zi 2 1; 1 if 2 zi 2 3.
Dataset Splits No The paper describes a 'training set' for its motivating example but does not provide specific details on training/validation/test dataset splits.
Hardware Specification No The paper describes empirical observations in Section 2 but does not provide any specific hardware details (e.g., GPU/CPU models, memory) used for the experiments.
Software Dependencies No The paper does not specify any software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions) that would be needed to replicate the experiments.
Experiment Setup Yes As loss, we use the logistic loss 1 n P log(1 + e yi ˆ yi), where ˆyi = f N (zi) is the output of the network on the i-th sample. Decision boundaries obtained by training fully connected networks with width d = 2 per hidden layer (top row) and Res Net (bottom row) with one neuron in the hidden layers on the unit ball classification problem.