LEADS: Learning Dynamical Systems that Generalize Across Environments
Authors: Yuan Yin, Ibrahim Ayed, Emmanuel de Bézenac, Nicolas Baskiotis, Patrick Gallinari
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We instantiate this framework for neural networks and evaluate it experimentally on representative families of nonlinear dynamics. We show that this new setting can exploit knowledge extracted from environment-dependent data and improves generalization for both known and novel environments. Our experiments are conducted on three families of dynamical systems described by three broad classes of differential equations. |
| Researcher Affiliation | Collaboration | 1Sorbonne Université, Paris, France 2There SIS Lab, Thales, Paris, France 3Criteo AI Lab, Paris, France |
| Pseudocode | No | The paper describes the framework and optimization problem mathematically but does not include a pseudocode block or an algorithm. |
| Open Source Code | Yes | Code is available at https://github.com/yuan-yin/LEADS. |
| Open Datasets | No | LV and GS data are simulated with the DOPRI5 solver in Num Py [10, 13]. NS data is simulated with the pseudo-spectral method as in [19]. The authors simulated their own data and do not provide a link or citation to a public dataset. |
| Dataset Splits | No | The paper specifies training and test data sets and their sizes but does not explicitly mention a distinct validation set split. |
| Hardware Specification | Yes | All experiments are performed with a single NVIDIA Titan Xp GPU. |
| Software Dependencies | No | The paper mentions software like Num Py and optimizers like Adam, and numerical methods (RK4, Euler), but it does not specify version numbers for any of these software dependencies. |
| Experiment Setup | Yes | We used 4-layer MLPs for LV, 4-layer Conv Nets for GS and Fourier Neural Operator (FNO) [19] for NS. We apply an exponential Scheduled Sampling [17] with exponent of 0.99 to stabilize the training. We use the Adam optimizer [15] with the same learning rate 10 3 and (β1, β2) = (0.9, 0.999) across the experiments. For the hyperparamters in Eq. 8, we chose respectively λ = 5 103, 102, 105 and α = 10 3, 10 2, 10 5 for LV, GS and NS. |