START GUIDE

How to make neural network dream

This article shows how to create fantastic art using artificial neural networks.

P.S. I made this banner

The convolution neural network may contain several stacked layers, images fed as an input to neural network travel through subsequent layers, and the final decision made by the output layer. But, there exist several questions, such as

• How layers communicate with one another?
• What does each layer see?
• What kind of information passes from one layer to another?

Visualizing the output of the layer of interest by enhancing the input image helps to understand what is happening at each neural network layer. A trained convolution neural network progressively extracts features of the input image. The first few layers may look for edges, the middle layers may look for shapes, and the final layer may use all the information from previous layers to make decisions.

Progression of an image through the neural network. (By Author)

But what if we feed the neural network iteratively on its output?

The result is an endless stream of images with a dream-like effect. Alexander Mordvintsev came up with the concept of the deep…