Sample-Efficient Optimization in the Latent Space of Deep Generative - - PowerPoint PPT Presentation

Sample-Efficient Optimization in the Latent Space of Deep Generative Models via Weighted Retraining

Erik Daxberger*, Austin Tripp* & Jos´ e Miguel Hern´ andez-Lobato

RealML @ ICML2020

2020-07-18

Problem

Optimization of expensive, black box functions on structured input spaces. Examples: Drug Design Materials Discovery Neural Architecture Search

Erik Daxberger*, Austin Tripp* & Jos´ e Miguel Hern´ andez-Lobato Keyword: Weighted Retraining 2/4

Latent Space Optimization

Optimize in the latent space Z of a deep generative model (instead of data space X)

X X Z Z

Normal

X X Z Z

With weighted retraining

Erik Daxberger*, Austin Tripp* & Jos´ e Miguel Hern´ andez-Lobato Keyword: Weighted Retraining 3/4

Results

Chemical Design Task

100 200 300 400 500 Number of objective function evaluations 5 10 15 20 weight; retrain weight; no retrain no weight; retrain no weight; no retrain

• riginal

works with a variety of models easy to implement huge increase in performance and sample efficiency More results in the paper!

Erik Daxberger*, Austin Tripp* & Jos´ e Miguel Hern´ andez-Lobato Keyword: Weighted Retraining 4/4