Provable Benefit of Orthogonal Initialization in Optimizing Deep Linear Networks

Authors: Wei Hu, Lechao Xiao, Jeffrey Pennington

ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we provide empirical evidence to support the results in Sections 4 and 5. To study how depth and width affect convergence speed of gradient descent under orthogonal and Gaussian initialization schemes, we train a family of linear networks with their widths ranging from 10 to 1000 and depths from 1 to 700, on a fixed synthetic dataset (X, Y ).
Researcher Affiliation Collaboration Wei Hu Princeton University huwei@cs.princeton.edu Lechao Xiao Google Brain xlc@google.com Jeffrey Pennington Google Brain jpennin@google.com
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any statement or link indicating the release of open-source code for the methodology described.
Open Datasets No We choose X R1024 16 and W R10 1024, and set Y = W X. Entries in X and W are drawn i.i.d. from N(0, 1).
Dataset Splits No The paper mentions a 'fixed synthetic dataset' and 'training loss' but does not provide explicit details about training, validation, or test splits.
Hardware Specification No The paper does not specify any details about the hardware used for the experiments.
Software Dependencies No The paper does not specify any software dependencies or their version numbers.
Experiment Setup Yes To study how depth and width affect convergence speed of gradient descent under orthogonal and Gaussian initialization schemes, we train a family of linear networks with their widths ranging from 10 to 1000 and depths from 1 to 700, on a fixed synthetic dataset (X, Y ). Each network is trained using gradient descent staring from both Gaussian and orthogonal initializations.