Painting Deepdreams

March 15, 2017

my first use-case for an AI art tool

I started getting serious about learning ML sometime after starting to paint.

Around the same time, I took an art class about "Painting the Personal Narrative," and we needed to make a past / present series of still life paintings.

I started with a "present" painting, which I called "startup dinner" (a joking reference to my somewhat unhealthy eating habits left over from pre-Pinterest life).

startup_dinner

Emulating Surrealism?

When we were asked to make a painting of the "past," the prompt was to make it "nightmarish." I had another photo chosen already, which i jokingly titled "startup breakfast". The only problem was, I didn't know how to make a dream effect.

Enter Deepdreams

Thankfully, code for making deepdreams had just been announced by Google Research.

deepdream examples

Running my first Deepdream

Pareidolia is a type of apophenia, which is a more generalized term for seeing patterns in random data. Some common examples are seeing a likeness of Jesus in the clouds or an image of a man on the surface of the moon. (source)

Painting it

What's going on here?

"This looks like a statement about how animals relate to the food we eat" - someone in my art class

The deepdream method evolved out of attempts to understand what Neural Networks "see".

One way to do this is by understanding which area of the image caused the prediction:

(distill.pub)

Instead, the a-ha moment for the deepdream authors was that we could get the neural network to "hallucinate" and how us "what picture would make this look more like X"?

"I paint not the things I see but the feelings they arouse in me" - Franz Kline

My Work: exploring high and low layers

How does it work?