Federated multi-task learning in Chemoinformatics

In pharmaceutical research and bioinformatics, learning problems with low sample size are common (protein-protein and drug-target interaction prediction, prediction of phenotypes, etc.). In these domains, the sample complexity reduction achieved by jointly training a large number of related tasks (called multitask learning) can be essential. The need for efficient multi-task learning methods is further motivated by the emerging problem of federated learning, an effort to learn these tasks jointly without sharing the underlying data sets. Besides its clear practical value, multitask learning is well motivated from a theoretical perspective: it is one of the principles that biological learning agents exploit to achieve extremely low sample complexity learning. The research aims at Developing efficient multitask learning methods with favorable properties for federated learning applications, and methods to assess their performance. The main directions of study are new neural-network based architectures with flexible parameter sharing schemes (instead of the classical trunk-head partitioning in federated learning), task subset selection, and grouping approaches. Furthermore, assessing the performance based on predictive performance, predictive uncertainty, computational complexity, and communication requirement in case of a future federated learning implementation will be an important topic in this research.