Multi-task learning has been shown to improve the predicting performance of machine learning models despite data scarcity, but imbalanced training datasets often degrade its efficacy through negative transfer. Here, the authors introduce adaptive checkpointing with specialization, a training scheme that mitigates detrimental inter-task interference, and demonstrate its practical utility by predicting sustainable aviation fuel properties.
- Basem A. Eraqi
- Dmitrii Khizbullin
- S. Mani Sarathy
