home projects sketches open source talks

Deep Dreaming

Deep Dreaming is a technique originally out of google to see how a vision neural network interprets images. The technique works by training an image classifier, then trying to maximize the activation of certain neurons and displaying that output onto an image. Most of the images you see of deep dreaming on the internet are pictures of dogs hallucinated into an image. I wanted to take this technique and play around with it.

Different Datasets

I trained my own classifiers on dozens of different classes of objects: Buddhas,Jesus, lizards,frogs, apple, cheesburges,flowers, Hillary Clinton, Donald Trump, brocolli, cats,and more. I then hallucinated those objects together with different images to try and create weird juxtapositions like putting hillary and trump together or cheesburgers and apples together.

Modifying the objective function

In machine learning,most models "learn" by comparing the ground truth with the model's generated results in a special objective function. I experimented with modifying the objective function to see if I could train on multiple types of images and generate output that was the combination of the different types.

Source code can be found here I generated thousands of images combing through to find the most interesting images. Below is a very small selection of some of the interesting ones.

The Mona Lisa with Buddhas

Praying Mantises in Tron

Buddha Forest

Recursive buddha - buddhas being hallucinated onto Buddha

Cats being hallucinated into the United States map

Donut Forest

Donald Trump and skyscrapers hallucinated onto the United States

Cheeseburgerz in the Great Wall

Skulls on a Plane