Improving black-box optimization in VAE latent space using decoder uncertainty

Authors: Pascal Notin, José Miguel Hernández-Lobato, Yarin Gal

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We illustrate these advantages across several experimental settings in digit generation, arithmetic expression approximation and molecule generation for drug design.
Researcher Affiliation Academia Pascal Notin Department of Computer Science University of Oxford Oxford, UK pascal.notin@cs.ox.ac.uk José Miguel Hernández-Lobato Department of Engineering University of Cambridge Cambridge, UK jmh233@cam.ac.uk Yarin Gal Department of Computer Science University of Oxford Oxford, UK yarin@cs.ox.ac.uk
Pseudocode Yes Algorithm 1: Importance sampling estimator of MI Algorithm 2: Bayesian Optimization with uncertainty censoring Algorithm 3: Uncertainty-constrained gradient ascent
Open Source Code Yes We are open sourcing the code in the repository at the following address: https://github.com/pascalnotin/uncertainty_guided_optimization.
Open Datasets Yes We train a Character VAE (CVAE) [4] on 80, 000 expressions generated by the formal grammar, then perform optimization in the latent space. ... For both architectures, we trained our models on a set of 250k drug-like molecules from the ZINC dataset [29].
Dataset Splits Yes Detailed information for datasets and model hyperparameter values used across experiments are detailed in Appendix B-E (one section dedicated to each experimental setting) and G (reproducibility). ... We consider 4 distinct sets of points in latent: embeddings into latent space of a random sample from the train and test sets...
Hardware Specification Yes Compute Resources: All experiments were run on NVIDIA V100 GPUs.
Software Dependencies Yes All models were implemented with Python 3.5.2 and PyTorch (version 1.7.1) [35].
Experiment Setup Yes Detailed information for datasets and model hyperparameter values used across experiments are detailed in Appendix B-E (one section dedicated to each experimental setting) and G (reproducibility).