RelevancePropagation.jl

Documentation	Build Status

Julia implementation of Layerwise Relevance Propagation (LRP) and Concept Relevance Propagation (CRP) for use with Flux.jl models.

This package is part of the Julia-XAI ecosystem and compatible with ExplainableAI.jl.

This package supports Julia ≥1.6. To install it, open the Julia REPL and run

julia> ]add RelevancePropagation

Let's use LRP to explain why an image of a castle gets classified as such using a pre-trained VGG16 model from Metalhead.jl:

using RelevancePropagation
using VisionHeatmaps             # visualization of explanations as heatmaps
using Flux
using Metalhead                  # pre-trained vision models

# Load & prepare model
model = VGG(16, pretrain=true).layers
model = strip_softmax(model)
model = canonize(model)

# Load input
input = ...                      # input in WHCN format

# Run XAI method
composite = EpsilonPlusFlat()
analyzer = LRP(model, composite)
expl = analyze(input, analyzer)  # or: expl = analyzer(input)
heatmap(expl)                    # show heatmap using VisionHeatmaps.jl

We can also get an explanation for the activation of the output neuron corresponding to the "street sign" class by specifying the corresponding output neuron position 920:

analyze(input, analyzer, 920) 

Heatmaps for all implemented analyzers are shown in the following table. Red color indicate regions of positive relevance towards the selected class, whereas regions in blue are of negative relevance.

Analyzer	Heatmap for class "castle"	Heatmap for class "street sign"
LRP with EpsilonPlus composite
LRP with EpsilonPlusFlat composite
LRP with EpsilonAlpha2Beta1 composite
LRP with EpsilonAlpha2Beta1Flat composite
LRP with EpsilonGammaBox composite
LRP with ZeroRule (discouraged)

Adrian Hill acknowledges support by the Federal Ministry of Education and Research (BMBF) for the Berlin Institute for the Foundations of Learning and Data (BIFOLD) (01IS18037A).