Colab Notebooks: Google Deep Dream

What’s this about?

DeepDream visualizes pattern recognition, interpretation and iterative generation by Neural Networks. By increasing this creative interpretation you can produce dream-alike imagery.

Neural Networks act like our brain in the case of Pareidolia: it looks for familiar patterns, which derive from datasets they were trained on.

Sources — on the left: NASA, Public Domain, 1976 // on the right: Mona Lisa, transformed with Deep Dream, CC-by nixtown, 2016

The example above (a screen from my presentation on the AI Meetup Frankfurt, November 2019) demonstrates how our brain recognizes a face in the rock formations of the Cydonia region on Mars. A user Nixtown transformed Da Vinci’s Mona Lisa by continuous DeepDream iterations — and AI recognizes weird patterns.

Often our brain recognizes patterns or objects which aren’t there. But if our human perception does it, why AI shouldn’t?

Links:

• Original Blog post by Alexander Mordvintsev
• GitHub
• Colab Notebook

Things to try out:

Try to generate different patterns and to resize images (octaves). Use more iterations. And don’t be afraid of insanity — the results could be unsettling.

Fun fact:

Initially, DeepDream used to recognize in every pattern mostly dog faces. According to FastCompany, the Network was trained on…

a smaller subset of the ImageNet database released in 2012 for use in a contest… a subset which contained “fine-grained classification of 120 dog sub-classes (FastCompany).

Read more about my experiments with DeepDream.

Index of Series "Google Colab Notebook".

Full essay "12 Colab Notebooks that matter"