Variational Inference and Model Selection with Generalized Evidence Bounds

Authors: Liqun Chen, Chenyang Tao, Ruiyi Zhang, Ricardo Henao, Lawrence Carin Duke

ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical evidence is provided to validate our claims. To compare the performance of our new bound and its predecessors, we empirically evaluate the sharpness of these bounds on a toy distribution, and benchmark them on a series of VI tasks.
Researcher Affiliation Academia Affiliation: Electrical & Computer Engineering, Duke University, Durham, NC 27708, USA. Correspondence to: Chenyang Tao <chenyang.tao@duke.edu>, Liqun Chen <liqun.chen@duke.edu>.
Pseudocode No The paper does not contain any explicitly labeled 'Pseudocode' or 'Algorithm' blocks or figures.
Open Source Code No Details of the experimental setup are in the SM, and source code is available (upon publication) from https: //www.github.com/Liqun Chen0606/glbo.
Open Datasets Yes on the MNIST dataset. We further evaluate GLBO on the more complex Celeb A face dataset (Liu et al., 2015). We use ten datasets from the UCI Machine Learning Repository (Lichman, 2013)...
Dataset Splits No We use a random 90%/10% split for training and testing, and use test root mean squared error (RMSE) and log-likelihood (LL) for evaluation. This only mentions train/test, not validation. Other sections don't specify splits either.
Hardware Specification No The paper does not provide specific details regarding the CPU, GPU, or other hardware used for running the experiments.
Software Dependencies No The paper mentions several models and implementations but does not specify the version numbers of any software dependencies or libraries used for the experiments.
Experiment Setup Yes The encoders and decoders are implemented with L {1, 2} neural network layers and leveraging K {5, 50} posterior samples. We choose the VR-Max estimator for RVB and set GLBO to CLBO(x; T, K) with T = 200. For CLBO and R enyi we fixed T = 2.