Maximize activation

Visualize the high-level features of deep Networks ：

1. Activate maximize ：
   1. Find the input mode that maximizes the activation value of a given hidden layer unit
   2. Template convolution ： If a convolution template is more similar to the module in the image , The greater the response .
   3. If an image input maximizes the output excitation value of this neuron , It can be considered that this neuron is to extract features similar to this input , So look for the one that can maximize the input of this neuron x Then it can visually and meaningfully express the features learned by this neuron     
   4. That is, once the network training is completed , Parameters w Is to determine the , Then we can find the corresponding x 了 ：

   5. It can be seen from the formula that ： This is a nonconvex optimization problem , There are many local minima . Find a local minimum through gradient descent ：

      1. Get two or more local minima . In either case , The features extracted by the neural node can be described by finding one or more minimum values . If there are multiple minimum values , Then you can find the one that maximizes the activation value or average all , Or show everything .

      2. Starting from the initialization of different random values, we iterate to get the same minimum value at last

2. sampling
3. Linear combination method