Disentangled Generative Models for Robust Prediction of System Dynamics
Authors: Stathi Fotiadis, Mario Lino Valencia, Shunlong Hu, Stef Garasto, Chris D Cantwell, Anil Anthony Bharath
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experimental results in phase-space and observation-space dynamics, demonstrate the effectiveness of latent-space supervision in producing disentangled representations, leading to improved longterm prediction accuracy and out-of-distribution robustness. |
| Researcher Affiliation | Academia | 1Department of Bioengineering, Imperial College London, UK 2Department of Aeronautics, Imperial College London, UK 3School of Computing and Mathematical Sciences, University of Greenwich, London, UK. Correspondence to: Stathi Fotiadis <s.fotiadis19@imperial.ac.uk>. |
| Pseudocode | No | The paper does not contain any pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | Yes | All the necessary code to reproduce our experiments is provided at https://github.com/stathius/sd-vae. The repository contains code and instructions for generating all the datasets and training all the models presented in this work using the hyperparameters that are clearly presented in the paper. This should significantly help others reproduce our experiments. |
| Open Datasets | No | The paper describes how the datasets were created for the experiments (e.g., "To create the datasets, we use an adaptive Runge-Kutta integrator with a timestep of 0.01 seconds.", "For every simulated sequence we draw a different combination of parameters.", and specific parameter ranges in Table 3 and Appendix A.3). It also states: "The repository contains code and instructions for generating all the datasets...". This indicates the data is generated via provided code, rather than being a pre-existing, publicly accessible dataset with a direct link or citation. |
| Dataset Splits | Yes | For each system, we draw the parameters 𝝃from a uniform distribution which is the same for the training, validation and test sets. These three datasets are considered in-distribution. ... Number of sequences Train/Val/Test 8000/1000/1000. |
| Hardware Specification | No | The paper mentions "5, 000 CPU-hours" but does not specify any particular CPU models, GPU models, or other hardware components used for running the experiments. It only refers to general computing resources. |
| Software Dependencies | No | The paper mentions using an "Adam optimizer" and states that "The repository contains code and instructions...", but it does not specify any particular software libraries, frameworks (like PyTorch or TensorFlow), or their version numbers. For instance, it does not state "PyTorch 1.x" or "Python 3.x". |
| Experiment Setup | Yes | We tune the hyperparameters of each method using grid-search and train the same number of models for each method to avoid favouring one over the others by chance. ... Details for the hyperparameters and number of experiments can be found in Appendix D.1. (Tables 4, 5, 6, 8 within Appendix D.1 explicitly list hyperparameters like Input Size, Output Size, Hidden Layers, Latent Size, Nonlinearity, Learning Rate, Batch size, Sched. patience, Sched. factor, Gradient clipping, Layer norm (latent), Decoder 𝛾, Sup. scaling, Supervision 𝛿). |