Variational Inference for Discriminative Learning with Generative Modeling of Feature Incompletion

Authors: Kohei Miyaguchi, Takayuki Katsuki, Akira Koseki, Toshiya Iwamori

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We have also empirically confirmed the effectiveness of the proposed method through numerical experiments, employing the variational autoencoders (VAEs) (Kingma & Welling, 2013) as the base generative model.
Researcher Affiliation Industry Kohei Miyaguchi, Takayuki Katsuki, Akira Koseki & Toshiya Iwamori IBM Research Tokyo miyaguchi@ibm.com, {kats,akoseki,iwamori}@jp.ibm.com
Pseudocode Yes Algorithm 1 Variational Inference for DIG (v DIG)
Open Source Code No The paper does not provide any explicit statements about making the source code available, nor does it include a link to a code repository.
Open Datasets Yes All the datasets used in the experiments are taken from UCI Machine Learning Repository. AQ-CO, AQ-NMHC and AQ-NOx are taken from the Air Quality dataset (De Vito et al., 2008)... Year Pred is a part of the Year Prediciton MSD dataset (Bertin-Mahieux, 2011)... Boston and Diabetes are respectively taken from the dataset of the same names.
Dataset Splits Yes We randomly hold 20% of the training split out for validation. Algorithms receive as the input only the remaining 80%.
Hardware Specification Yes CPU RAM GPU Py Torch Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz 64GB NVIDIA TITAN X 1.9.0
Software Dependencies Yes Py Torch 1.9.0; Iterative Imputer from scikit-learn (ver. 0.24.2); Ada Belief optimizer (Zhuang et al., 2020)
Experiment Setup Yes The size of minibatch is always taken to be 512. In particular, for sampling-based methods (i.e., VAE, VAE*, CVAE, DVAE, DVAE*), the stochastic gradient is computed with the reparametrization trick. We iterate the loop 2000 times for each configuration. We choose α = 2 for the parameter of the R enyi divergence.