A Simple Neural Network Module for Relational Reasoning

Implemented in MXNet as part of NIPS Implementation Challenge

Link to paper : Arxiv | pdf

About this implementation

This implementation tries to replicate the results in the paper for the sort-of-clevr dataset. The code for generating the dataset has been borrowed from here.

Apache MXNet is used as the framework of choice, specifically MXNet's symbolic interface.

Results

Baseline Model(CNN + MLP) Accuracy at 30th Epoch

Overall	0.639880
Non-Relational	0.599121
Relational	0.683105

Model with Relation Network (CNN + RN) at 30th Epoch

Overall	0.899305
Non-Relational	0.978027
Relational	0.821289