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

paper
Author

Austin Tripp, Erik Daxberger, José Miguel Hernández-Lobato

Official version Version I consider authoritative

Summary

ML algorithm to optimize structured inputs (e.g. molecules) via weighted retraining of a generative model.

My contribution

I conceived the project and performed the experiments / did writing jointly with Erik.

Thoughts

(as of 2023-11-24)

This is the first paper that was really “mine”. While there is nothing explicitly wrong in the paper, in hindsight the approach and methodology is more naive and less original than I previously thought. The algorithm proposed in the paper has two parts.

  • The “weighted retraining”: I think this is sensible and would expect it to work well on other problems. However, I do think it is basically an instance of an “Estimation of distribution algorithm” and am unsure whether it would work really better in practice than other EDAs.

  • The “latent space optimization” (LSO): I tried to extend the LSO to other problems for ~1 year but had poor results. I think the fundamental difficulty is that it is at least as hard to create an informative probabilistic model in a continuous latent space as it is in a discrete space like molecules. In fact, it might even be harder: e.g. it is probably easier to specify a prior for smoothness with respect to molecular structure directly in molecule space than in some abstract latent space. I think the reason this method got strong results in our paper is because the objective we optimized, logP, is essentially just a proxy for molecular size and is easy to optimize by finding large out-of-distribution molecules.

    Overall I would not recommend using LSO: instead I would create a model directly in whatever discrete input space is being used (e.g. a GP with the Tanimoto kernel for molecules) and optimize acquisition functions using genetic algorithms. I know there has been subsequent work on LSO which is more positive, but I think these works just lack strong BO baselines which don’t use a latent space.