Context Aware Second-Order Interpretation

Code for reproducing experiments in "Understanding Impacts of High-Order Loss Approximations and Features in Deep Learning Interpretation".

Prerequisites

• Python, NumPy, Pytorch, Argparse, Matplotlib
• A recent NVIDIA GPU

Basic Usage

To evaluate the interpretation with default parameters on the given toy image, run python main.py. To access all the parameters use python main.py --help.

Examples

Impact of Group Features

To generate the following examples use python main.py --lambda1 LAMBDA

Low values of λ₁ lead to sparse saliency maps. Noise is removed and salient features are highlighted.

Impact of Hessian

To generate the CAFO example, use python .\main.py --image_path=IMAGE_NAME --lambda1=0 --magnitude
For the CASO example, use python .\main.py --image_path==IMAGE_NAME --lambda1=0 --magnitude --second-order

For low confidence in the predicted class, CASO and CAFO can be very different (turtle example). For high confidence, CASO and CAFO are similar (duck example).