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

Area of Science:

• Computational chemistry
• Machine learning
• Artificial intelligence

Background:

• Transfer learning (TL) leverages existing knowledge to improve performance on related tasks.
• Data scarcity is a major challenge in chemical property prediction.
• Latent variables facilitate information exchange between tasks in TL models.

Purpose of the Study:

• Investigate the impact of task correlation on TL model performance in chemical property prediction.
• Determine limitations on the number of simultaneously trained tasks before performance degradation.
• Assess the effect of TL on ancillary model properties.

Main Methods:

• Utilized an autoencoder latent space as a latent variable for TL models.
• Employed the QM9 dataset supplemented with semiempirical quantum chemistry calculations.
• Incorporated a linear predictor model to organize the latent space.

Main Results:

• Property prediction improved counterintuitively with a simpler linear predictor model.
• TL showed dramatic improvement in data-scarce prediction tasks.
• Little adverse impact of TL was observed in data-rich prediction tasks.

Conclusions:

• TL is a highly advantageous supplement for chemical property prediction models.
• The demonstrated TL approach offers significant benefits with no implementation downside.
• This method effectively addresses data scarcity limitations in predictive modeling.