Invariance-based Learning of Latent Dynamics
Authors: Kai Lagemann, Christian Lagemann, Sach Mukherjee
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We study behavior through simple theoretical analyses and extensive empirical experiments. The latter investigate the ability to predict the trajectories of complicated systems based on finite data and show that the proposed approaches can outperform existing neural-dynamical models. and Experiments are structured into four different series that shed light on the performance of La DID. We provide a short overview of the experimental set-up in the following. |
| Researcher Affiliation | Academia | Kai Lagemann* Statistics and Machine Learning, German Center for Neurodegenerative Diseases (DZNE), Bonn, Germany kai.lagemann@dzne.de, Christian Lagemann* Department of Mechanical Engineering, University of Washington, Seattle, USA, Sach Mukherjee Statistics and Machine Learning, German Center for Neurodegenerative Diseases (DZNE), Bonn, Germany MRC Biostatistics Unit, University of Cambridge, Cambridge, UK sach.mukherjee@dzne.de |
| Pseudocode | No | The paper describes the model architecture and training process in prose within the main text and appendix, but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | *A reference implementation is available under https://github.com/kl844477/La DID. |
| Open Datasets | Yes | Source code of the simulation can be found in Champion et al. (2019). (for the Reaction-Diffusion Equation dataset) |
| Dataset Splits | Yes | The training, validation and test dataset is split into 400, 50 and 50 trajectories, respectively. (from I.1 Swinging Pendulum, similar statements are repeated for other datasets in Appendix I) |
| Hardware Specification | Yes | All computations are run on a single GPU node equipped with one Nvidia A100 (40 GB) and a global batch size of 16 is used. and All tests are performed on a NVIDIA A100 40 Gb with a AMD EPYC 7742 processor. (from Table H.2) |
| Software Dependencies | No | All network architectures are implemented in the open source framework Py Torch (Paszke et al., 2019). (A specific version number for PyTorch is not provided, and no other software dependencies with version numbers are listed). |
| Experiment Setup | Yes | Training is carried out in a multi-phase schedule w.r.t. the multiple shooting loss in eq. 9. and All network architectures are implemented in the open source framework Py Torch (Paszke et al., 2019). Further training details and hyperparameters can be found in Appendix F. and Table F.1: Training hyperparameters which lists Initial LR 3e-4, Weight Decay 0.01, Global batch size 16, Number of epochs per subpatch length 3000, Latent dimension 32, Number of attention blocks 8, Number of attention heads 4. |