Variational Inference with Locally Enhanced Bounds for Hierarchical Models
Authors: Tomas Geffner, Justin Domke
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We present an extensive empirical evaluation of our approach using two tightening methods: importance weighting (Burda et al., 2016) and uncorrected Hamiltonian annealing (Geffner & Domke, 2021b; Zhang et al., 2021). The former is based on importance sampling, while the latter uses Hamiltonian Monte Carlo (Neal et al., 2011; Betancourt, 2017) transition kernels to build an enhanced variational distribution. We observe empirically that the proposed approach yields better results than plain variational inference and a traditional application of tightening methods. |
| Researcher Affiliation | Academia | 1College of Information and Computer Sciences, University of Massachusetts Amherst, MA, USA. |
| Pseudocode | No | The paper does not contain pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | No | The paper does not provide any links to or statements about the availability of source code. |
| Open Datasets | Yes | We used data from Movie Lens100K (Harper & Konstan, 2015). |
| Dataset Splits | No | The paper describes characteristics of the datasets used (number of groups and observations per group) but does not provide explicit training/validation/test splits for reproducibility, such as percentages or specific sample counts for each split. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for experiments. |
| Software Dependencies | No | The paper mentions 'Adam (Kingma & Ba, 2014)' as an optimizer but does not specify version numbers for any software dependencies. |
| Experiment Setup | Yes | We optimize using Adam (Kingma & Ba, 2014) with a stepsize η = 0.001. ...and train for 50k steps. ...we use subsampling with M = 10 to estimate gradients at each step using the reparameterization trick... We initialize all methods to maximizers of the ELBO... |