Note: The two intialization methods will result in the same images in conv1 (because the first layer by itself is linear). Try the other layers to see what happens. Also, Vgg16 and ResNet18 are currently not supported.
Explanations of the figures:
Zero-Initialized: Gradient ascent is performed without providing any prior information. The starting tensor is made of zeros.
Top-patch Initialized: The image patch that results in the strongest response of the unit (i.e., artificial neuron) is used as the starting tensor for gradient ascent.
Top Patch (original): The top image patch (within the 50,000 ImageNet test set) was identified by my previous research. See the Main page to learn more.