Compositionality for Deep Learning

Partners involved: Hammer*, Mukherjee, Stegle

ML biological process chains include various learning paradigms: unsupervised processes such as nonlinear dimension reduction, and supervised processes such as classification by deep learning . Currently, the methods of the chain are typically trained independently, and the integration of additional knowledge is ad hoc, resulting in a time-consuming interactive design process. The goal of WP4 is to realize the compositionality of ML components in order to achieve

i) autonomous training of the different model components including hyper-parameter optimization and

ii) extensibility of the models to important aspects such as interpretability or data protection.

The initial focus is on three typical goals scRNA-seq data analysis: identification of cell types, function prediction, and identification of causal relationships.