On the reproducibility of fully convolutional neural networks for modeling time-space evolving physical systems
Abstract
Reproducibility of a deep-learning fully convolutional neural network is evaluated by training several times the same network on identical conditions (database, hyperparameters, hardware) with non-deterministic Graphics Processings Unit (GPU) operations. The propagation of two-dimensional acoustic waves, typical of time-space evolving physical systems, is studied on both recursive and non-recursive tasks. Significant changes in models properties (weights, featured fields) are observed. When tested on various propagation benchmarks, these models systematically returned estimations with a high level of deviation, especially for the recurrent analysis which strongly amplifies variability due to the non-determinism. Trainings performed with double floating-point precision provide slightly better estimations and a significant reduction of the variability of both the network parameters and its testing error range.
Keywords
Cite
@article{arxiv.2105.05482,
title = {On the reproducibility of fully convolutional neural networks for modeling time-space evolving physical systems},
author = {Wagner Gonçalves Pinto and Antonio Alguacil and Michaël Bauerheim},
journal= {arXiv preprint arXiv:2105.05482},
year = {2021}
}