Skip to content

Latest commit

 

History

History
15 lines (9 loc) · 989 Bytes

README.md

File metadata and controls

15 lines (9 loc) · 989 Bytes

Explaining Neural Networks by Decoding Layer Activations

Code for Paper: Explaining Neural Networks by Decoding Layer Activations Accepted at Intelligent Data Analysis (IDA), 2021 (https://ida2021.org/) https://arxiv.org/abs/2005.13630

It uses Python and Pytorch (The original code in the paper used TensorFlow).

It creates plots showing original and reconstructed samples (Explanations) by ClaDec. Things that are blurry or poorly reconstructed are not very relevant for classification. Sharp and well-matching parts between original and explanation are more important to the classifier. It can be observed, for example, that color (or grey scale information) or patterns (of clothes) are not relevant for the classifier, while shape is.

In the paper, you can see more samples, in particular a comparison with a reference decoder.

Original and Explanations for last (logit) Layer

Original and Explanations for last conv Layer