Activation Maximization

Finding the input that maximizes the activation of a Unit, mostly used to visualize the feature detectors in a CNN.

For a single neuron we can write an optimization problem like this $im{g}^{\ast }=\arg {\max }_{img}{h}_{n,x,y,z}(img)$ where

• $h$ is the Activation Function
• $x,y$ encode the position of the Unit
• $n$ is the layer
• $z$ is the channel index

The mean activation for an entire channel can be achieved by averaging over all neruons in that channel $im{g}^{\ast }=\arg {\max }_{img}{\sum }_{x,y}{h}_{n,x,y,z}(img)$

Using the minimum instead of the maximum is not working. It will correspond to …

You can start with an image from

• the training data
• random noise
  • need to do Regularisation like
    • jittering
    • rotation
      • linear combination of all channels leads to entangled features (makes interpretability harder)
    • scaling
    • frequency penalization to reduce variance of neighboring pixels
    • GANs
    • denoising Autoencoder