GitHub - jurgyy/CLRP: Contrastive Layer-wise Relevance Propagation or CLRP is a modification of standard Layer-wise Relevance Propagation (LRP) with the goal of making the output (more) class discriminative. This notebook will use the LRP library Innvestigate to attempt to implement CLRP

Implementing Contrastive Layer-wise Relevance Propagation with Innvestigate

Contrastive Layer-wise Relevance Propagation[1] or CLRP is a modification of standard Layer-wise Relevance Propagation[2] (LRP) with the goal of making the output (more) class discriminative. This notebook will use the LRP library Innvestigate[3] to attempt to implement CLRP.

A general overview of CLRP is as follows:

Given an output neuron which represent concept we try to construct a dual virtual concept $\overline O$ which represents the opposite concept of .
This concept $\overline O$ can be represented in two different ways:
A. CLRP1: The concept is represented by the selected classes with weights $\overline W = \{W^1, W^2, ..., W^{L-1}, W^L_{-j}\}$ . Here $W_{-j}$ means the weights connected to the output layer excluding the neuron.
B. CLRP2: The concept is represented by the selected classes with weights $\overline W = \{W^1, W^2, ..., W^{L-1}, W^L_{-j}\}$ . Which means all the weights are the same, except for the last layer where the weights to neuron $j$ are negated.
(?) The score $S_{y_j}$ of target class is uniformly redistributted to other classes.
$R_{\text{LRP}} = f_{\text{LRP}}(X, W, S_{y_j})$
Given the same input example LRP generates an explanation $R_{\text{dual}} = f_{\text{LRP}}(X, \overline W, S_{y_j})$ for the dual concept.
Then CLRP is defined as follows: $R_{\text{CLRP}} = \max(0, R - R_{\text{dual}})$

Here are some results from the CLRP paper which shows a very clear class discrimitative property. These results are from using VGG16 pre-trained on imagenet and applying the $z^\beta$ -rule in the first convolution layer and for all the other convulutional layers the -rule. For more details read the paper.