Differentiable and Stable Long-Range Tracking of Multiple Posterior Modes
Authors: Ali Younis, Erik Sudderth
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our theory and experiments expose dramatic failures of existing reparameterization-based estimators for mixture gradients, an issue we address via an importance-sampling gradient estimator. Unlike standard recurrent neural networks, our mixture density particle filter represents multimodal uncertainty in continuous latent states, improving accuracy and robustness. On a range of challenging tracking and robot localization problems, our approach achieves dramatic improvements in accuracy, while also showing much greater stability across multiple training runs. |
| Researcher Affiliation | Academia | Ali Younis and Erik B. Sudderth {ayounis, sudderth}@uci.edu Department of Computer Science, University of California, Irvine |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks (clearly labeled algorithm sections or code-like formatted procedures). |
| Open Source Code | No | The paper does not include an unambiguous statement about releasing the source code for the described methodology, nor does it provide a direct link to a code repository. |
| Open Datasets | Yes | In the Deepmind-Maze tracking task, adapted from [5], we wish to track the 3D state of a robot as it moves through modified versions of the maze environments from Deepmind Lab [48]... This 3D state tracking task is adapted from [7], where a robot navigates single-level apartment environments from the SUNCG dataset [49] using the House3D simulator [50]... |
| Dataset Splits | Yes | We use 5000 training and 1000 validation trajectories of length T = 17, and 5000 evaluation trajectories of length T = 150. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., exact GPU/CPU models, processor types, or memory amounts) used for running its experiments, only general mentions of computation. |
| Software Dependencies | No | The paper mentions the Adam optimizer but does not provide specific version numbers for any key software components or libraries used in the experiments. |
| Experiment Setup | Yes | We use the Adam [52] optimizer with default parameters, tuning the learning rate for each task. We use a batch size of 64 during training and decrease the learning rate by a factor of 10 when the validation loss reaches a plateau. We use Truncated-Back-Propagation-Through-Time (T-BPTT) [47], truncating the gradients every 4 time-steps. For all tasks we use the Gaussian distribution for the positional components of the state and the Von Mises distribution for the angular components when applying KDE. Here β is a dimension-specific bandwidth corresponding to the two standard deviations for the Gaussians and a concentration for the Von Mises. When training MDPF using the Negative Log-Likelihood (NLL) loss, we set the learning rate (LR) of the neural networks to be 0.0005 and the bandwidth learning rate to 0.00005. When training using MSE loss, we use 0.0001 and 0.00001 for the neural network and bandwidth learning rates respectively. |