SIXO: Smoothing Inference with Twisted Objectives
Authors: Dieterich Lawson, Allan Raventós, Andrew Warrington, Scott Linderman
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We experimentally explore our claims that: 1. The SIXO bound can become tight while FIVO cannot. 2. DRE-learned twists enable better posterior inference than filtering SMC. 3. Model learning with SIXO provides better parameter estimates than FIVO. We conclude by comparing FIVO and SIXO for parameter recovery in Figure 4 and Table 3. We see that FIVO-bs converges to a poor parameter estimate, and recovers a poor variational bound. |
| Researcher Affiliation | Academia | {jdlawson, aravento, awarring, scott.linderman}@stanford.edu Stanford University |
| Pseudocode | Yes | Algorithm 1 SIXO-DRE. See Algorithm 2 in Appendix B.1. |
| Open Source Code | No | Code is available publicly at the URL listed. (However, no URL is provided within the document). |
| Open Datasets | Yes | We now apply SIXO to a stochastic volatility model (SVM) of monthly foreign exchange rates for N = 22 currencies, over the period 9/2007 to 8/2017 [40]. |
| Dataset Splits | No | Figure 4a: Validation set L256 Method over training. While the paper states that 'Extensive written configurations are included in the supplement' which implies data splits, these specifics are not detailed within the provided main text. |
| Hardware Specification | No | The paper states, 'We list the types of nodes and runtimes for the algorithms presented,' but no specific hardware models (e.g., GPU/CPU models, cloud provider instances) are detailed in the provided text. |
| Software Dependencies | No | The paper mentions 'JAX' and 'Weights and Biases' but does not specify version numbers for these or any other software dependencies needed to replicate the experiments. |
| Experiment Setup | No | The paper states that 'Extensive written configurations are included in the supplement' and 'Code released contains hyperparameters for reproduction of results,' but concrete hyperparameter values (e.g., learning rate, batch size) are not provided in the main text itself. |