Latent Space Energy-based Neural ODEs
Authors: Sheng Cheng, Deqian Kong, Jianwen Xie, Kookjin Lee, Ying Nian Wu, Yezhou Yang
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on oscillating systems, videos and real-world state sequences (Mu Jo Co) demonstrate that our model with the learnable energy-based prior outperforms existing counterparts, and can generalize to new dynamic parameterization, enabling long-horizon predictions. ... 4 Experiments This section investigates the performance and adaptability of our proposed ODE-LEBM in various scenarios. |
| Researcher Affiliation | Collaboration | Sheng Cheng EMAIL School of Computing and Augmented Intelligence, Arizona State University Deqian Kong EMAIL Department of Statistics and Data Science, University of California, Los Angeles Jianwen Xie EMAIL Akool Research Kookjin Lee EMAIL School of Computing and Augmented Intelligence, Arizona State University Ying Nian Wu EMAIL Department of Statistics and Data Science, University of California, Los Angeles Yezhou Yang EMAIL School of Computing and Augmented Intelligence, Arizona State University |
| Pseudocode | Yes | Algorithm 1 Learning algorithm for our ODE-LEBM ... Algorithm 2 Test-phase algorithm for our ODE-LEBM |
| Open Source Code | No | The paper does not provide an explicit statement about open-source code availability or a link to a code repository. |
| Open Datasets | Yes | Rotating MNIST The data is generated following the implementation by Casale et al. (2018); Auzina et al. (2024), with the total number of rotation angles set to 16. We include all ten digits from MNIST dataset... Mu Jo Co physics simulation (Todorov et al., 2012), widely used for training reinforcement learning models, is employed in our experiments with three physical environments: Hopper, Swimmer, and Half Cheetah using the Deep Mind control suite (Tunyasuvunakool et al., 2020). |
| Dataset Splits | Yes | Irregularly-sampled time series ... The dataset is split into 80% for training and 20% for testing. ... Rotating MNIST ... The training data consists of N = 1000 trajectories... The validation and test data consist of Nval = Ntest = 100 trajectories... Mu Jo Co physics simulation ... The data was split into training and testing sets at an 80/20 ratio. |
| Hardware Specification | Yes | Computing resources for training All experiments can be done within an 11GB GPU like GTX 1080Ti. However, to speed up the training process, we use a single V100 for training. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers. It mentions using 'MLP' and 'GELU' but no framework or library versions. |
| Experiment Setup | Yes | Model Architecture For the emission model and neural ODE model, the network architecture designs are the same as the baseline model in Auzina et al. (2024); Rubanova et al. (2019). For EBM prior, all experiments use 3 layers MLP with activation GELU. ... Training details of prior model and MCMC-based inference The number of steps in Langevin sampling is 20 for training. We list the remaining training details in Table 9. (Table 9 includes: hidden dim, step size, learning rate for Irregular sampling, Bouncing balls, Rotating MNIST, Mu Jo Co). |