How Modular should Neural Module Networks Be for Systematic Generalization?

Authors: Vanessa D'Amario, Tomotake Sasaki, Xavier Boix

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In a series of experiments on three VQA datasets (VQA-MNIST, SQOOP, and CLEVRCo Gen T), our results reveal that tuning the degree of modularity, especially at the image encoder stage, reaches substantially higher systematic generalization.
Researcher Affiliation Collaboration Vanessa D Amario 1, 3 Tomotake Sasaki 2, 3 Xavier Boix 1, 3 1 Massachusetts Institute of Technology, USA 2 Fujitsu Limited, Japan 3 Center for Brains, Minds and Machines, USA
Pseudocode No The paper describes module definitions with mathematical formulas but does not include structured pseudocode or algorithm blocks.
Open Source Code Yes Code and data can be found in the following github repository: https://github.com/vanessadamario/understanding_reasoning.git.
Open Datasets Yes Code and data can be found in the following github repository: https://github.com/vanessadamario/understanding_reasoning.git.
Dataset Splits Yes We perform a grid-search over the NMNs hyper-parameters, and evaluate the accuracy using an in-distribution validation split at the last iteration of the training.
Hardware Specification No The paper does not provide specific details about the hardware used to run the experiments, such as GPU or CPU models.
Software Dependencies No The paper does not provide specific version numbers for software dependencies or libraries used in the experiments.
Experiment Setup Yes For each experiment, we tuned the hyper-parameters by performing a random-search over 100 sets of hyper-parameters. We used the Adam optimizer [23] with learning rates equal to {10e-3, 10e-4, 10e-5, 10e-6}, and dropout rates equal to {0, 0.2, 0.4, 0.6, 0.8}.