Generalized Teacher Forcing for Learning Chaotic Dynamics
Authors: Florian Hess, Zahra Monfared, Manuel Brenner, Daniel Durstewitz
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We show on several DS that with these amendments we can reconstruct DS better than current SOTA algorithms, in much lower dimensions. Performance differences were particularly compelling on real world data with which most other methods severely struggled. This work thus led to a simple yet powerful DS reconstruction algorithm which is highly interpretable at the same time. |
| Researcher Affiliation | Academia | 1Department of Theoretical Neuroscience, Central Institute of Mental Health, Medical Faculty Mannheim, Heidelberg University, Mannheim, Germany 2Faculty of Physics and Astronomy, Heidelberg University, Heidelberg, Germany 3Cluster of Excellence STRUCTURES, Heidelberg University, Heidelberg, Germany 4Interdisciplinary Center for Scientific Computing, Heidelberg University, Heidelberg, Germany. Correspondence to: Florian Hess <Florian.Hess@zi-mannheim.de>, Daniel Durstewitz <Daniel.Durstewitz@zi-mannheim.de>. |
| Pseudocode | Yes | Algorithm 1 Adaptive GTF |
| Open Source Code | Yes | All code is available at https://github.com/DurstewitzLab/GTF-shPLRNN. |
| Open Datasets | Yes | ECG recordings of a human subject were taken from the open-access PPG-Da Li A dataset (Reiss et al., 2019). ... As another challenging real-world data set we use open-access EEG recordings from a human subject under task-free, quiescent conditions with eyes open (Schalk et al., 2000). |
| Dataset Splits | Yes | For the Lorenz-63 and Lorenz-96 and the empirical ECG data, we used a sequence length of T = 200. For the EEG data, we used only T = 50. ... For our experiments, we use the first T = 100, 000 samples ( ~143s) as the training set and the next T = 100, 000 samples as the test set. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running the experiments (e.g., GPU models, CPU types, or specific cloud instances). |
| Software Dependencies | No | The paper mentions several software components like "Dynamical Systems.jl (Datseris, 2018) Julia library," "Py SINDy package (de Silva et al., 2020)," and "torchdiffeq package." However, it does not specify version numbers for any of these libraries or the Julia/Python environments used, which is required for reproducible software dependencies. |
| Experiment Setup | Yes | In all experiments, we use the RAdam (Liu et al., 2020) optimizer with a learning rate starting at 10^-3 which is exponentially reduced to reach 10^-6 at the end of training. For all datasets, we trained for 5000 epochs, where one epoch is defined as processing of 50 batches of size S = 16. ... We fixed k = 5, α0 = 1, and γ = 0.999. |