Principle Feature Visualisation in Convolutional Neural Networks

Abhinav Jaivishnu Choudhary, Animesh Maheshwari

Keywords: CNN, Explainable AI, Visualisation

Abstract:

Explainable AI is crucial in determining the performance of machine learning systems and debugging the code wherever necessary. This motivated us to pursue the project for establishing the reproducibility of the work on "Principal Feature Visualisation
in Convolutional Neural Networks", a paper in ECCV 2020.

Scope of Reproducibility:
We validated the reproducibility of the work on "Principal Feature Visualisation
in Convolutional Neural Networks" by Marianne Bakken1. We experimented with this technique on various images as well as various CNN pre-trained models.

Methodology:
We used the author's code and re-implemented it on various images to validated the 4 major claims of the paper. We changed the pre-trained models in the code and noted the best for the tested image classes amongst them.

Results:
We used the authors code and implemented it on the images used by the authors as well as other image classes. We compared the results with the Grad-Cam approach and validated all the claims made by authors in the paper.

What was easy:
Implementing the code was easy and required very less computational power. The comments with the code by the authors made it easy to comprehend the code and conduct various experiment with it.

What was difficult:
The description provided in the paper was difficult to comprehend in the terms of reproducing the code.

Communication with original authors:
We contacted an author of the paper. We wanted to understand the reasons about the colours appearing on the result images and in understanding the reason behind the results not appearing appropriately with the inception-v3 model.

Paper Url: https://openreview.net/forum?id=Uz_uJd7WCc6&noteId=vYr_GrQVhe8&referrer=%5BML%20Reproducibility%20Challenge%202020%5D(%2Fgroup%3Fid%3DML_Reproducibility_Challenge%2F2020)