And thanks for this wonderfull tool. I am using it to better understand the different ways to interpret a CNN :)
So I was wondering, for Class Activation Maps, could we seed a random noise and let it converge until it maximizes the activation for one specific class ? Using an input image (like in visualize_cam) helps him to converge then ?
I really see this problem as a generative model, so we could also use a regularizer on the loss (from a trained GAN for example, or other generative model).
Hi everyone,
And thanks for this wonderfull tool. I am using it to better understand the different ways to interpret a CNN :)
So I was wondering, for Class Activation Maps, could we seed a random noise and let it converge until it maximizes the activation for one specific class ? Using an input image (like in
visualize_cam
) helps him to converge then ? I really see this problem as a generative model, so we could also use a regularizer on the loss (from a trained GAN for example, or other generative model).Thanks,