TRENDy: Temporal Regression of Effective Nonlinear Dynamics
Authors: Matthew Ricci, Guy Pelc, Zoe Piran, Noa Moriel, Mor Nitzan
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our principal contributions are Experiments indicating TRENDy can predict qualitative, topological changes in the fitted system s behavior, i.e., bifurcations. ... Demonstrations that TRENDy can predict and explain the emergence of complex spatial patterns... An analysis of spatiotemporal data taken from high-resolution videos of the development of the ocellated lizard. |
| Researcher Affiliation | Academia | Matthew Ricci, Guy Pelc, Zoe Piran & Noa Moriel School of Computer Science & Engineering The Hebrew University of Jerusalem 9190401 Jerusalem, Israel EMAIL Mor Nitzan School of Computer Science & Engineering Racah Institute of Physics Faculty of Medicine The Hebrew University of Jerusalem 9190401 Jerusalem, Israel EMAIL |
| Pseudocode | No | The paper describes methods and equations in detail but does not present any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | 1The TRENDy codebase is available at https://github.com/nitzanlab/trendy. |
| Open Datasets | Yes | An analysis of spatiotemporal data taken from high-resolution videos of the development of the ocellated lizard. These results indicate a link between body geometry and pattern growth and showcase TRENDy s relevance to real, noisy data in an important biological test case 1. ... We first analyzed data generated by Fofonjka & Milinkovitch (2021); Manukyan et al. (2017), and acquired a dataset of patch evolutions by randomly sampling locations in a high-resolution video of the developing ocellated lizard. |
| Dataset Splits | Yes | Gray Scott. ... We generated 1000 training samples with |k k | > .01 and 250 testing samples with |k k | < .01... Brusselator We used 2000 training and 500 testing samples in the parameter regions described in the main text and controlled by the holdout factor, ϵ. Lizard patterning ...we did an 80-20 split of the a(T) into training and testing sets and fit an SVM to the training samples. |
| Hardware Specification | No | The paper mentions that TRENDy was trained in PyTorch and numerical continuation was run in Julia, but it does not specify any particular hardware components like CPU or GPU models used for these experiments. |
| Software Dependencies | Yes | TRENDy was trained in Py Torch (Paszke et al., 2019) and numerical continuation was run on Bifurcation Kit (Veltz, 2020) in Julia (Bezanson et al., 2017). |
| Experiment Setup | Yes | In all experiments, the NODE is a multilayer perceptron with four layers, with 64 hidden units in each layer, and with zero-rectification nonlinearities. ... For the Brusselator experiment, we set the derivative regularizer to be β = 10 4 and it was 0 otherwise. We used an Adam optimizer (Kingma & Ba, 2014) with learning rate 10 4. ... Gray Scott. ... TRENDy was fit over 2000 epochs, each of which took approximately 31.48 seconds. We used a burn-in period of 10 time steps. ... Numerical continuation experiments were performed using pseudo-arclength continuation with a Newton Raphson correction having threshold 1 10 5. In parallel, a bisection algorithm with 15 steps was used to detect bifurcations using an eigenvalue threshold of 1.0. |