Gated Inference Network: Inference and Learning State-Space Models
Authors: Hamidreza Hashempoorikderi, Wan Choi
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To substantiate our claims, we conduct five experiments. First, we simulate a nonlinear dynamic system using the pendulum sequence video... The model is then trained end-to-end on videos depicting a diverse range of simulated and real-world physical systems, and outperforms its counterparts RNNs, autoregressive models, and variational approaches in state estimation and missing data imputation tasks. |
| Researcher Affiliation | Academia | Hamidreza Hashempoor Seoul National University Department of Electrical and Computer Engineering hamidreza.hashemp@snu.ac.kr Wan Choi Seoul National University Department of Electrical and Computer Engineering wanchoi@snu.ac.kr |
| Pseudocode | Yes | Algorithm Training the GIN |
| Open Source Code | Yes | The entire code to reproduce the experiments are available in Github repository. |
| Open Datasets | Yes | Visual Odometry of KITTI Dataset. We also evaluate the GIN with the higher dimensional observations for the visual odometry task on the KITTI dataset [40]. |
| Dataset Splits | No | The paper does not explicitly provide training/test/validation dataset splits by percentages or counts, nor does it reference predefined splits with citations for reproducibility. |
| Hardware Specification | Yes | In all experiments, we used the Adam optimizer [46] on an NVIDIA Ge Force GTX 1050 Ti with 16GB RAM. |
| Software Dependencies | No | The paper mentions software like Adam optimizer, Backpropagation through time, layer normalization, and Tensorflow library, and Python 3.6+, but does not specify exact version numbers for the libraries or frameworks used (e.g., TensorFlow 2.x, PyTorch 1.x). |
| Experiment Setup | Yes | We chose an initial learning rate of 0.006 with an exponential decay rate of 0.9 every 10 epochs. Backpropagation through time [47] was employed to compute gradients, given the use of GRU cells in our structure. The gradients are applied to GRU cells with the constraint explained in theorem 3, where we use the spectral theorem mentioned in the main script . We applied the layer normalization technique [48] to stabilize the dynamics within the recurrent structure and normalize the filter response. ... In our experiments, we set K = 15 to accommodate the complexity of the latent space... |