The origin of problem 1
We need to build a loss function, its purpose is to let a convolution layer of a filter value maximization, and then we will use stochastic gradient descent to adjust the value of the input image so that the activation value maximization,
Question 1: the period of don't understand, related to the question 2
Question origin 2
Through the stochastic gradient descent to maximize the loss:
Starting from a gray image with noise, rise along the gradient direction to adjust the image, to maximize the loss, the purpose is to let a filter to maximize response and get the input image is selected filter has the maximum response of the image,
Question 2: gradient descent, isn't it, how to use gradient rise? Why to maximize cost rise along the gradient adjustment may give the maximum filter response? Get the input image is selected filter has the maximum response of images and visual filter what is the relationship?
With a focus on the visual principle not understand
Self-study of new shivering, still hope bosses great help, thank you thank you thank you thank you thank you (omit ten thousand words)