Work                  Events                 About

RIVEN RATANAVANH



Making Visual Art with GANs with Derrick Schultz

Week 1 Notes

Part of learning how to work with Machine Learning models is learning a way to interact with computers - much different from interacting with it through Processing etc.

Every model is slightly different. What is important is matching what you need to do with what the model does. Also: lower your expectations. You have to work with the parameters, and essentially, limitations of each model.

How you approach the dataset will change how your results.

Process

In this class, we’ll work with Narrow A.I.
As a way of working, we have to break down what we want to do in order to work backwards.

The ML Process involves:
1. Dataset
2. Training
3. Model

1. If we want to create a flower, we need a generalized dataset of flowers. Working with datasets will be the hardest part of this class, becuase unique output will need unique input.

2. Training is the architecture of the model(?). Training something like StyleGAN can take up to 2 weeks.

3. The model is essentially a package of code that comes out of the training. The input and the output will be of specific type.


GANs

Generative means that we are generating something new from our models, as opposed to classifying something.

If you spend 5 weeks with a model, you will likely have spent more time than the researcher has spent with it.



Technique != Art

The machine is not art.
What can you, as an artist, make with a machine?

Do you want to work with hardware? (e.g. make an image and build it).
Make video? Make performances?



LINKS
Alexander Mordvintsev
Sofia Crespo
Mario Klingemann
Mario Pickard - trained a 2D Model and turned it into 3D stained glass
Memo Akten - live waves
John Gall - scuptures
Robbie Barat
Scott Eaton
Anna Ridler
Everest Pipkin... critical AI
From Fran - AI as a tool in the toolchain, or as inspirational tool
Robbie Barat

Note: artists have tended to lean into the surreal, impressionistic, or psychedelic to work with the affordances of current GANs.










Horse to Zebra.
https://www.youtube.com/watch?v=9reHvktowLY