Debiasing Evidence Approximations: On Importance-weighted Autoencoders and Jackknife Variational Inference

Authors: Sebastian Nowozin

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

Reproducibility Variable Result LLM Response
Research Type Experimental We now empirically validate our key claims regarding the JVI method: 1. JVI produces better estimates of the marginal likelihood by reducing bias, even for small K; and 2. Higher-order bias reduction is more effective than lower-order bias reduction; To this end we will use variational autoencoders trained on MNIST.
Researcher Affiliation Industry Sebastian Nowozin Machine Intelligence and Perception Microsoft Research, Cambridge, UK Sebastian.Nowozin@microsoft.com
Pseudocode Yes Algorithm 1 Computing ˆLJ,m K , the jackknife variational inference estimator
Open Source Code Yes 1The implementation is available at https://github.com/Microsoft/ jackknife-variational-inference
Open Datasets Yes Our setup is purposely identical to the setup of Tomczak & Welling (2016), where we use the dynamically binarized MNIST data set of Salakhutdinov & Murray (2008).
Dataset Splits Yes We train on the first 50,000 training images, using 10,000 images for validation.
Hardware Specification Yes All our models are implemented using Chainer (Tokui et al., 2015) and run on a NVidia Titan X.
Software Dependencies No The paper mentions 'Chainer (Tokui et al., 2015)' but does not specify a version number for this software dependency.
Experiment Setup Yes Hyperparameters are the batch size in {1024, 4096} and the SGD step size in {0.1, 0.05, 0.01, 0.005, 0.001}.