GTC Silicon Valley-2019: Molecular Generative VAEs: Parallelization, Optimization, and Latent Space Analysis on the DGX-1
Note: This video may require joining the NVIDIA Developer Program or login
GTC Silicon Valley-2019 ID:S9417:Molecular Generative VAEs: Parallelization, Optimization, and Latent Space Analysis on the DGX-1
Ellen Du(The Dow Chemical Company),Joey Storer(The Dow Chemical Company)
Generative Variational Autoencoders (VAE) in molecular discovery and new materials design have recently gained considerable attention in academia as well as industry (Gomez-Bombarelli, 2017). In this talk, we will present results from a combined Dow Chemical and NVIDIA development effort to implement a VAE for chemical discovery. We'll discuss challenges associated with applying deep learning to chemistry and highlight recently developed methods. Highlights from our presentation will include a discussion of methods to analyze and sample from an organized latent representation in a conditioned variational autoencoder, tips for training a complex architecture, distributed multi-node training using Horovod, and results showing the generation of molecular structure with associated property prediction.