Modulated Neural ODEs
Authors: Ilze Amanda Auzina, Çağatay Yıldız, Sara Magliacane, Matthias Bethge, Efstratios Gavves
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We test Mo NODE on oscillating systems, videos and human walking trajectories, where each trajectory has trajectory-specific modulation. Our framework consistently improves the existing model ability to generalize to new dynamic parameterizations and to perform far-horizon forecasting. To investigate the effect of our proposed dynamics and static modulators, we structure the experiments as follows: First, we investigate the effect of the dynamics modulator on classical dynamical systems, namely, sinusoidal wave, predator-prey trajectories and bouncing ball (section 5.1), where the parameterisation of the dynamics differs across each trajectory. Second, to confirm the utility of the static modulator we implement an experiment of rotating MNIST digits (section 5.2), where the static content is the digit itself. Lastly, we experiment on real data with having both modulator variables present for predicting human walking trajectories (section 5.4). |
| Researcher Affiliation | Collaboration | Ilze Amanda Auzina University of Amsterdam i.a.auzina@uva.nl Ça gatay Yıldız University of Tübingen Tübingen AI Center Sara Magliacane University of Amsterdam MIT-IBM Watson AI Lab Matthias Bethge University of Tübingen Tübingen AI Center Efstratios Gavves University of Amsterdam |
| Pseudocode | No | The paper does not contain any pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | Yes | Our official implementation can be found at https://github.com/Ilze Amanda A/Mo NODE. |
| Open Datasets | Yes | To investigate the effect of our proposed dynamics and static modulators, we structure the experiments as follows: First, we investigate the effect of the dynamics modulator on classical dynamical systems, namely, sinusoidal wave, predator-prey trajectories and bouncing ball (section 5.1)... Second, to confirm the utility of the static modulator we implement an experiment of rotating MNIST digits (section 5.2)... Lastly, we experiment on real data with having both modulator variables present for predicting human walking trajectories (section 5.4). Next, we evaluate our framework on a subset of CMU Mocap dataset, which consists of 56 walking sequences from 6 different subjects. We pre-process the data as described in [Wang et al., 2007]. For data generation, we use the script provided by Sutskever et al. [2008]. |
| Dataset Splits | Yes | Sinusoidal data: The training data consists of N = 300 oscillating trajectories with length T = 50. ... Validation and test data consist of Nval = Ntest = 50 trajectories with sequence length Tval = 50 and Ttest = 150, respectively. (Section 5.1.1). Table 6 also explicitly lists N, Nval, and Ntest for all datasets. |
| Hardware Specification | Yes | All models for sine, PP and rotating MNIST have been trained on a single GPU (GTX 1080 Ti) with 10 CPUs and 30G memory. |
| Software Dependencies | No | We implement all models in Py Torch [Paszke et al., 2017]. The encoder, decoder, differential function, and modulator prediction networks are all jointly optimized with the Adam optimizer [Kingma and Ba, 2014]. For solving the ODE system we use torchdiffeq [Chen, 2018] package. We use the 4th-order Runge-Kutta numerical solver to compute ODE state solutions. While software packages are mentioned, specific version numbers (e.g., PyTorch 1.9) are not provided; only publication years for their initial release/description are cited. |
| Experiment Setup | Yes | Implementation details We implement all models in Py Torch [Paszke et al., 2017]. The encoder, decoder, differential function, and modulator prediction networks are all jointly optimized with the Adam optimizer [Kingma and Ba, 2014]. For solving the ODE system we use torchdiffeq [Chen, 2018] package. We use the 4th-order Runge-Kutta numerical solver to compute ODE state solutions (see App. D for ablation results for different solvers). For the complete details on data generation and training setup, we refer to App.B. Further, we report the architectures, number of parameters, and details on hyperparameter for each method in App. C Table 7. (Appendix B and C provide extensive details on training setup, batch sizes, learning rates, epochs, and hyperparameter tables). |