Layerwise-Relevance-Propagation

Implementation of Layerwise Relevance Propagation for heatmapping "deep" layers, using Tensorflow and Keras.

Results

MNIST

VGG

Instructions

MNIST

• Run train.py to train model.
• Weights will be saved in logs/.
• Run lrp.py for Layerwise Relevance Propagation.

NOTE: If using Tensorflow version < 1.5.0, you need to change tf.nn.softmax_cross_entropy_with_logits_v2 to tf.nn.softmax_cross_entropy_with_logits.

VGG

• Feed a list of images to run Layerwise Relevance Propagation on all images.
• All results will be saved in results/.
• Run lrp.py <image_1> <image_2> ... <image_n>.

Reference

• On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation
• Explaining NonLinear Classification Decisions with Deep Taylor Decomposition
• Understanding Neural Networks with Layerwise Relevance Propagation and Deep Taylor Series
• A Quick Introduction to Deep Taylor Decomposition
• Tutorial: Implementing Layer-Wise Relevance Propagation