Design and independent training of composable and reusable neural modules

David Castillo-Bolado, Cayetano Guerra-Artal, Mario Hernández-Tejera

Abstract: Highlights•Neural networks acquire behaviour when assigned a role during training.•Neural data types favour compositionality, generalization and reuse of modules.•Guidelines allow building a gradient-approximator module for training and selection.•Modular training facilitates incremental development of complex neural networks.•Modular-compositional methodology improves network accuracy.