About the monotonicity of the predict layer in the coral network

[Tony-958]

Hi, Sorry to bother. I had a problem when understanding the monotonicity of the coral network's predict layer. Here's the only statement I found modifying the bias layer adding to the fc result: self.linear_1_bias = nn.Parameter(torch.zeros(self.num_classes-1).float()) Is it enough to restrict the monotonicity ? Or is there any other statement restricting the monotonicity of the biases？

[rasbt]

I would call it "guaranteeing" monotonicity :). Yeah, it's the bias and the weight sharing together with the loss function. Regarding the changes compared to a regular network, maybe this overview helps? https://github.com/Raschka-research-group/coral-cnn/blob/master/github-images/differences-at-a-glance.pdf

[Tony-958]

Yeah, that helps. Thank you so much.