Latent Field Discovery in Interacting Dynamical Systems with Neural Fields

Authors: Miltiadis (Miltos) Kofinas, Erik Bekkers, Naveen Nagaraja, Efstratios Gavves

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our proposed method, Aether, on settings that include static as well as dynamic fields. First, we explore 2D charged particles that evolve under the effect of a static electrostatic field, as well as 3D particles that evolve under a Lorentz force field [10]. Then, we evaluate our method on a subset of in D [4] that contains a single location, and thus a static field as well. Finally, we explore 3D gravitational n-body problems [5] with dynamic fields. Our code, data, and models will be open-sourced online1. In most experiments, we compare our method against d NRI [14] and Lo CS [24] , two state-of-the-art networks for sequence-to-sequence trajectory forecasting, as well as G-Lo CS.
Researcher Affiliation Collaboration Miltiadis Kofinas University of Amsterdam m.kofinas@uva.nl Erik J. Bekkers University of Amsterdam e.j.bekkers@uva.nl Naveen Shankar Nagaraja BMW Group Naveen-Shankar.Nagaraja@bmw.de Efstratios Gavves University of Amsterdam egavves@uva.nl
Pseudocode No The paper provides detailed equations and descriptions of the model architecture and training process in Appendix A, but it does not include explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes Our code, data, and models will be open-sourced online1. 1https://github.com/mkofinas/aether
Open Datasets Yes We use in D [4], a dataset with real-world traffic scenes that comprises trajectories of pedestrians, vehicles, and cyclists.Du et al. [10] introduced a dataset of 3D charged particles evolving under the influence of a Lorentz force field.We extend the charged particles dataset from Kipf et al. [23] by adding a number of immovable sources.
Dataset Splits Yes We generate 50,000 simulations for training, 10,000 for validation and 10,000 for testing.The subset corresponds to 12 recordings; we use 8 for training, 2 for validation, and 2 for testing.
Hardware Specification Yes All experiments were performed on single GPUs. We used 2 different GPU models, namely the Nvidia RTX 2080 Ti, and Nvidia GTX 1080 Ti.
Software Dependencies Yes Our source code was written in Py Torch [32], version 1.4.0, and CUDA 10.0.
Experiment Setup Yes Unless specified differently, our neural field has a hidden size of 512. In charged particles and in n-body problems, we only use positions as input to the neural field, while in traffic scenes we also use orientations.Unless stated otherwise, in all experiments, we use a learning rate of 5e 4.