LyaNet: A Lyapunov Framework for Training Neural ODEs

Authors: Ivan Dario Jimenez Rodriguez, Aaron Ames, Yisong Yue

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

Reproducibility Variable Result LLM Response
Research Type Experimental Relative to standard Neural ODE training, we empirically find that Lya Net can offer improved prediction performance, faster convergence of inference dynamics, and improved adversarial robustness.
Researcher Affiliation Collaboration 1Department of Computational and Mathematical Sciences, California Institute of Technology 2Argo AI.
Pseudocode Yes Algorithm 1 Monte Carlo Lya Net Training Algorithm 2 Path Integral Lya Net Training
Open Source Code Yes Our code is available at https://github. com/ivandariojr/Lyapunov Learning.
Open Datasets Yes We evaluate primarily on three computer vision datasets: Fashion MNIST, CIFAR-10 and CIFAR-100.
Dataset Splits Yes We found this by performing a grid search on learning rates and batch sizes over (0.1, 0.001, 0.001) (32, 64, 128), validated on a held out set of 10% of training data.
Hardware Specification Yes Our experiments ran on a cluster 6 GPUs: 4 Ge Force 1080 GPUs, 1 Titan X and Titan RTX. All experiments were able to run on less than 10GB of VRAM.
Software Dependencies No The paper mentions 'Nero (Liu et al., 2021)' and 'PGD as implemented by Kim (2020)' but does not specify version numbers for these software components or libraries.
Experiment Setup Yes To simplify tuning, we trained our models using Nero (Liu et al., 2021) with a learning rate of 0.01 with a batch size of 64 for models trained with Lya Net and 128 for models trained with regular backpropagation. We found this by performing a grid search on learning rates and batch sizes over (0.1, 0.001, 0.001) (32, 64, 128), validated on a held out set of 10% of training data. All models were trained for a total of 120 epochs. For our adversarial attack we used PGD as implemented by Kim (2020) for 10 iterations with a step size α = 2 255.