Declarative nets that are equilibrium models

Authors: Russell Tsuchida, Suk Yee Yong, Mohammad Ali Armin, Lars Petersson, Cheng Soon Ong

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

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically show that this initialisation scheme improves training stability and performance over random initialisation.
Researcher Affiliation Academia Data61, CSIRO Canberra, Australia Space & Astronomy, CSIRO Epping, Australia Machine Learning & Artificial Intelligence Future Science Platform
Pseudocode No The paper describes methods mathematically and textually but does not include structured pseudocode or algorithm blocks.
Open Source Code Yes We have included code and the instructions required to reproduce our results. The code is publically available at https://github.com/Russell Tsuchida/deq-glm.
Open Datasets Yes Image denoising. We consider a convolutional architecture... applied to an image denoising task using the CIFAR10 dataset.
Dataset Splits No The paper specifies training and test set sizes (e.g., 'training and test set of size N = 20, 000 and 2, 000') but does not explicitly mention a separate validation set split or its proportions.
Hardware Specification Yes Empirically measured cost (in seconds) for initialisation schemes measured on a DELL Laptop (16GB RAM, Intel Core TM i7-8665U CPU), averaged over 100 runs.
Software Dependencies No The paper mentions 'Pytorch default' for initialization and 'Adam' for training but does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes We train for 400 epochs using Adam with default hyperparameters. We repeat this for 100 trials using seeds 0 to 99, see Figure 5(a) and (b).