An Intriguing Failing of Convolutional Neural Networks and the CoordConv Solution

By Rosanne Liu et al

Published on Dec. 3, 2018

Read the original document by opening this link in a new tab.

Table of Contents

1. Abstract
2. Introduction
3. Not-so-Clevr dataset
4. The CoordConv layer
5. Supervised Coordinate tasks
6. Conclusion

Summary

The paper discusses the failure of convolutional neural networks in solving the coordinate transform problem and introduces the CoordConv solution. It shows how Convolutional networks fail in tasks involving spatial representations and explains how CoordConv allows networks to learn translation invariance or dependence. The CoordConv layer is described as an extension to standard convolution, enabling networks to access input coordinates. The paper presents results from various experiments, demonstrating the effectiveness of CoordConv in improving performance on tasks like image generation and object detection. The research explores the fundamental challenge of learning coordinate transforms and offers CoordConv as a solution that can be easily integrated into convolutional networks.