In the previous chapters, we identified the optimal weights that result in classifying an image into the right class. The output class of an image can be changed by varying the following:

• The weights connecting the input to the output layer, while the input pixels remain constant
• The input pixel values, while the weights remain constant

In this chapter, we will employ these two techniques to generate images.

In the case studies of an adversarial attack, the neural style transfer and DeepDream will leverage the technique of changing the input pixel values. In the techniques involving a Generative Adversarial Network (GAN), we will leverage the technique of changing certain weights that connect input pixel values to the output.

The first three case studies in this chapter will leverage the technique of changing the input pixel values, while the rest leverage a change in weights that connect the input to the output.