Latent Space Bayesian Optimization With Latent Data Augmentation for Enhanced Exploration.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Chemistry

Background:

• Latent space Bayesian optimization (LSBO) integrates generative models like variational autoencoders (VAE) with Bayesian optimization (BO) for de novo object generation.
• Existing LSBO methods struggle with exploration due to a mismatch between VAE and BO objectives, leading to inefficiencies.
• This mismatch creates latent inconsistency, hindering the effectiveness of LSBO.

Purpose of the Study:

• To introduce and address the problem of latent inconsistency in LSBO.
• To enhance the exploration capabilities and sample efficiency of LSBO.
• To develop a novel LSBO framework that improves de novo generation.

Main Methods:

• Proposed the concept of latent consistency/inconsistency to diagnose LSBO challenges.
• Developed a latent consistent aware-acquisition function (LCA-AF) to leverage consistent latent points.
• Introduced LCA-VAE, a VAE variant employing latent space data augmentation and inconsistency penalization to foster latent consistency.
• Combined LCA-VAE and LCA-AF to create the LCA-LSBO framework.

Main Results:

• The proposed LCA-LSBO framework demonstrates high sample efficiency.
• Effective exploration capabilities were achieved by addressing latent inconsistency.
• The novel incorporation of latent space data augmentation within LCA-VAE significantly improved LSBO performance.

Conclusions:

• Addressing latent consistency is crucial for improving LSBO efficiency and exploration.
• LCA-LSBO, combining LCA-VAE and LCA-AF, offers a robust solution for de novo generation tasks.
• The study highlights the effectiveness of latent space data augmentation in enhancing generative model-based optimization.