How Modular should Neural Module Networks Be for Systematic Generalization?

Neural Module Networks (NMNs) aim at Visual Question Answering (VQA) via composition of modules that tackle a sub-task. NMNs are a promising strategy to achieve systematic generalization, ie. overcoming biasing factors in the training distribution. However, the aspects of NMNs that facilitate systematic generalization are not fully understood. In this paper, we demonstrate that the degree of modularity of the NMN have large influence on systematic generalization. In a series of experiments on three VQA datasets (VQA-MNIST, SQOOP, and CLEVR-CoGenT), our results reveal that tuning the degree of modularity, especially at the image encoder stage, reaches substantially higher systematic generalization. These findings lead to new NMN architectures that outperform previous ones in terms of systematic generalization. IMAGE ENCODER(S) to obtain visual features INTERMEDIATE MODULE(S) to carry out sub-tasks CLASSIFIER(S) to provide an answer group 1 group 1 [arg] group 1 group | group | group sub-task sub-task sub-task sub-task | sub-task | sub-task all | group | group all group 1 [arg] group 1 all | all | all all all [arg] all group 1 group | all | all all [arg] all '2' left_of green '2' all [green] all all ['2'] all [left_of] color category Question: "Is the green object left of '2'?" Program layout: group | group | group sub-task | sub-task | sub-task all | group | group all | all | all group | all | all left_of green '2'