Switching Autoregressive Low-rank Tensor Models
Authors: Hyun Dong Lee, Andrew Warrington, Joshua Glaser, Scott Linderman
NeurIPS 2023 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We empirically demonstrate quantitative advantages of SALT models on a range of simulated and real prediction tasks, including behavioral and neural datasets. |
| Researcher Affiliation | Academia | Hyun Dong Lee Computer Science Department Stanford University EMAIL Andrew Warrington Department of Statistics Stanford University EMAIL Joshua I. Glaser Department of Neurology Northwestern University EMAIL Scott W. Linderman Department of Statistics Stanford University EMAIL |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Source code is available at https://github.com/lindermanlab/salt. |
| Open Datasets | Yes | Wiltschko et al. [2015] collected videos of mice freely behaving in a circular open field. and We analyzed neural recordings of an immobilized C. elegans worm from Kato et al. [2015]. |
| Dataset Splits | Yes | The likelihood on a held-out validation set shows that the ARHMM overfitted quickly as the number of lags increased, while CP-SALT was more robust to overfitting (Figure 4B). We compared loglikelihoods of the best model (evaluated on the validation set) on a separate held-out test set and found that CP-SALT consistently outperformed ARHMM across mice (Figure 4C). |
| Hardware Specification | Yes | using 100 iterations of EM on a single NVIDIA Tesla P100 GPU. and using 100 iterations of EM on a single NVIDIA Tesla V100 GPU. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers. |
| Experiment Setup | Yes | We used H = 50 discrete states and fitted ARHMMs and CP-SALT models with varying lags and ranks. We imposed stickiness on the discrete state transition matrix via a Dirichlet prior with concentration of 1.1 on non-diagonals and 6 104 on the diagonals. We trained each model 5 times with random initialization for each hyperparameter, using 100 iterations of EM on a single NVIDIA Tesla P100 GPU. and We fitted both single and multi-subspace CP-SALT models with ranks D {8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18}. Similarly, SLDSs were fitted with the same range of latent dimension size. For both CP-SALT models and ARHMMs, we used L {1, 3, 6, 9} and the number of discrete states was set to H = 7 |