Simpler is Better: How Linear Prediction Tasks Improve Transfer Learning in Chemical Autoencoders

Transfer learning is a subfield of machine learning that leverages proficiency in one or more prediction tasks to improve proficiency in a related task. For chemical applications, transfer learning models represent a promising approach for addressing intrinsic data scarcity by utilizing potentially abundant data across adjacent domains. For chemical applications, it is still largely unknown how correlation between the prediction tasks affects performance, the limitations on the number of tasks that can be simultaneously trained in these models before incurring performance degradation, and if transfer learning positively or negatively affects ancillary model properties. In this talk we investigate these questions using an autoencoder latent space as a latent variable for transfer learning models trained on the QM9 dataset that have been supplemented with quantum chemistry calculations. We explore how property prediction can be improved by utilizing a simpler linear predictor model, forces the latent space to reorganize linearly with respect to each property. The linear organization of the latent space has further applications to novel structure generation by increasing the quality of generated species and facilitating targeted structure searching.