Towards interpreting computer vision based on transformation invariant optimization

TL;DR

This paper introduces a transformation-invariant visualization method for deep neural networks in computer vision, enhancing interpretability by incorporating rotation and scaling invariance into generated images that activate specific network predictions.

Contribution

It proposes a novel back-propagation based visualization technique that integrates transformation invariance, improving the clarity and insightfulness of neural network interpretation.

Findings

Transformation invariance improves visualization quality.

Generated images better activate target classes.

Method provides new insights into neural network behavior.

Abstract

Interpreting how does deep neural networks (DNNs) make predictions is a vital field in artificial intelligence, which hinders wide applications of DNNs. Visualization of learned representations helps we humans understand the vision of DNNs. In this work, visualized images that can activate the neural network to the target classes are generated by back-propagation method. Here, rotation and scaling operations are applied to introduce the transformation invariance in the image generating process, which we find a significant improvement on visualization effect. Finally, we show some cases that such method can help us to gain insight into neural networks.