EpsilonRule(; ϵ=1f-6)

Constructor for LRP-$ϵ$ rule. Commonly used on middle layers.

Arguments:

• ϵ: Optional stabilization parameter, defaults to 1f-6.

GammaRule(; γ=0.25)

Constructor for LRP-$γ$ rule. Commonly used on lower layers.

Arguments:

• γ: Optional multiplier for added positive weights, defaults to 0.25.

Gradient(model)

Analyze model by calculating the gradient of a neuron activation with respect to the input.

IndexNS(index)

Neuron selector that picks the output neuron at the given index.

InputTimesGradient(model)

Analyze model by calculating the gradient of a neuron activation with respect to the input. This gradient is then multiplied element-wise with the input.

LRP(c::Chain, r::AbstractLRPRule)
LRP(c::Chain, rs::AbstractVector{<:AbstractLRPRule})

Analyze model by applying Layer-Wise Relevance Propagation.

Keyword arguments

• skip_checks::Bool: Skip checks whether model is compatible with LRP and contains output softmax. Default is false.
• verbose::Bool: Select whether the model checks should print a summary on failure. Default is true.

References

[1] G. Montavon et al., Layer-Wise Relevance Propagation: An Overview [2] W. Samek et al., Explaining Deep Neural Networks and Beyond: A Review of Methods and Applications

MaxActivationNS()

Neuron selector that picks the output neuron with the highest activation.

ZBoxRule()

Constructor for LRP-$z^{\mathcal{B}}$-rule. Commonly used on the first layer for pixel input.

ZeroRule()

Constructor for LRP-0 rule. Commonly used on upper layers.

analyze(input, method)
analyze(input, method, neuron_selection)

Return raw classifier output and explanation. If neuron_selection is specified, the explanation will be calculated for that neuron. Otherwise, the output neuron with the highest activation is automatically chosen.

check_model(method::Symbol, model; verbose=true)

Check whether the given method can be used on the model. Currently, model checks are only implemented for LRP, using the symbol :LRP.

Example

julia> check_model(:LRP, model)

check_ouput_softmax(model)

Check whether model has softmax activation on output. Return the model if it doesn't, throw error otherwise.

drop_singleton_dims(a)

Drop dimensions of size 1 from array.

flatten_model(c)

Flatten a Flux chain containing Flux chains.

heatmap(expl::Explanation; kwargs...)
heatmap(attr::AbstractArray; kwargs...)

heatmap(input, analyzer::AbstractXAIMethod)
heatmap(input, analyzer::AbstractXAIMethod, neuron_selection::Int)

Visualize explanation. Assumes Flux's WHCN convention (width, height, color channels, batch size).

Keyword arguments

• cs::ColorScheme: ColorScheme that is applied. When calling heatmap with an Explanation or analyzer, the method default is selected. When calling heatmap with an array, the default is ColorSchemes.bwr.
• reduce::Symbol: How the color channels are reduced to a single number to apply a colorscheme. The following methods can be selected, which are then applied over the color channels for each "pixel" in the attribution:
  • :sum: sum up color channels
  • :norm: compute 2-norm over the color channels
  • :maxabs: compute maximum(abs, x) over the color channels in
  When calling heatmap with an Explanation or analyzer, the method default is selected. When calling heatmap with an array, the default is :sum.
• normalize::Symbol: How the color channel reduced heatmap is normalized before the colorscheme is applied. Can be either :extrema or :centered. When calling heatmap with an Explanation or analyzer, the method default is selected. When calling heatmap with an array, the default for use with the bwr colorscheme is :centered.
• permute::Bool: Whether to flip W&H input channels. Default is true.

Note: these keyword arguments can't be used when calling heatmap with an analyzer.

modify_denominator(rule, d)

Function that modifies zₖ on the forward pass, e.g. for numerical stability.

modify_layer(rule, layer)

Function that modifies a layer before applying relevance propagation. Returns a new, modified layer.

modify_params(rule, W, b)

Function that modifies weights and biases before applying relevance propagation. Returns modified weights and biases as a tuple (ρW, ρb).

safedivide(a, b; eps = 1f-6)

Elementwise division of two matrices avoiding near zero terms in the denominator by replacing them with ± eps.

set_params(layer, W, b)

Duplicate layer using weights W, b.

stabilize_denom(d; eps = 1f-6)

Replace zero terms of a matrix d with eps.

strip_softmax(model)

Remove softmax activation on model output if it exists.