sebastian-lapuschkin
commented
3 years ago first, preset A applies epsilon-LRP to dense layers, alphabeta to conv layers. the use of epsilon (I assume in your case) near the output allows to transport negative model outputs. once some neuron has relevance, you do not need to care about its sign, to phrase it bluntly, for further propagation.
second, alphabeta-LRP transports any propagated relevance incoming from upstream neurons, wrt alpha- and beta- weighted pre-activations of the layer. that means, the rule can propagate negative relevance just fine, it just weights the preactivations between layer inputs and outputs differently for the purpose of routing relevance downstream.
also consider, that positive preactivations (leading to potentially firing relu-activated neurons) may provide evidence speaking against your class of interest. these neurons will then usually receive negative relevance.
i hope that helps.
for a more specific answer, I would require info about the model, some example images, heatmaps and model outputs, etc, ie, details.
best
adrhill
Hello,
I am using the LRP preset A for CNNs and I was wondering why I obtain heatmaps with negative (blue) relevance when the applied rules are alpha-beta (with alpha =1, beta = 0) and epsilon rule? This is how I initiate my analyzer: analyzer = innvestigate.create_analyzer('lrp.sequential_preset_a', model=model, allow_lambda_layers=True)
Thank you in advance!