Principled Weight Initialization for Hypernetworks

Authors: Oscar Chang, Lampros Flokas, Hod Lipson

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

Reproducibility Variable Result LLM Response
Research Type Experimental We derive novel weight initialization formulae for hypernetworks in Section 4, empirically evaluate our proposed methods in Section 5, and finally conclude in Section 6.
Researcher Affiliation Academia Oscar Chang, Lampros Flokas, Hod Lipson Columbia University New York, NY 10027 {oscar.chang, lf2540, hod.lipson}@columbia.edu
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any concrete access information (e.g., specific repository link, explicit code release statement) for source code related to the methodology described.
Open Datasets Yes As an illustrative first experiment, we train a feedforward network with five hidden layers (500 hidden units), a hyperbolic tangent activation function, and a softmax output layer, on MNIST across four different settings
Dataset Splits No The paper specifies training and testing, but does not provide explicit details about dataset validation splits (e.g., percentages, sample counts, or methodology for a separate validation set).
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts) used for running its experiments.
Software Dependencies No The paper mentions "Py Torch and Chainer" but does not provide specific version numbers for these or any other ancillary software components.
Experiment Setup Yes The networks were trained on MNIST for 30 epochs with batch size 10 using a learning rate of 0.0005 for the hypernets and 0.01 for the classical network.