Project

Doppelgänger

Last year my husband and I began planning and construction of what we refer to as the Doppelgänger Project. As the name implies, our objective is to find a biologically unrelated double of my artistic identity.

We are in the process of creating and customizing a GAN (Generative Adversarial Network) that will be trained primarily, if not solely, using images of my artwork. The goal is to eventually have a neural network that is able to produce images that seamlessly identify and reflect my artistic tendencies.

Training Set A

The main challenge that we currently face is that neural networks usually require massive training sets in order to produce an adequate result and I have been producing paintings that are coherent with my current style over only the past three years or so.

Training Set A is a compilation of my paintings, which I have been adding to by making small paintings daily, that will help increase the range and quality of examples that the GAN can learn from. My goal is to have a collection of approximately 300 original paintings to work with. I currently have close to 250 original paintings in total.

The new paintings that I have been making for Training Set A are spontaneous and impromptu paintings intended to communicate and codify my raw, creative vision while maintaining maximum flexibility and authenticity. Each painting takes 1 to 2 hours to create and is documented and indexed for future access.

Training Set B

The second training set is a combination of my original paintings and 'First Generation Mutations' made using Deep Dream Generator. The purpose of this training set is to provide quantity of data so that we can get higher quality initial results from the GAN.

With this option, Training Set B could contain up to 90,000 images that are directly based on my artwork. However, there is a cap to how many images I can produce daily. The image count for Training Set B has reached 1,000 and we will continue to add more images until we have reached a solid starting point for the GAN to work off of.

Future Iterations:

After the initial build and development process, we hope to shift to training the Doppelgänger solely off of Training Set A, which consists of my original paintings along with the addition of what I will be currently and naturally producing going forward.

The goal is that once we have finished jumpstarting the system, we will be able to map and chart its long-term development and essential characteristics in comparison to my own.

Development of the GAN

Alex is using Python as the primary language for the creation of the GAN and has specifically been using the Keras machine learning framework.

In addition to referencing online code libraries, he has also looked to Udemy courses to help bring him up to speed on deep learning and has been reading Deep Learning with Python by François Chollet.

One of the crucial aspects of this project is to both rely upon and add to general understanding of neural networks. This project is in no way meant to be exclusive and will be entirely transparent from start to finish.

The intention is to have the final system be entirely open source so that other artists have the opportunity to experiment with Project Doppelgänger in their own practice.

If you would like to take a look at the code behind Project Doppelgänger, simply click here.

Where we are at...

Alex has finished building a Convolutional Neural Network that is able to distinguish my mutated artworks from my original paintings with ~90% accuracy, given a data set of approximately 1,000 images.

This Convolutional Neural Network uses images from Training Set B and more or less functions the same as the discriminator for the GAN will, given some slight modifications. This is an essential barrier to developing a successful GAN and will ensure that the GAN's discriminator is able to successfully identify the characteristics of my artwork.

Alex is now in the process of fine-tuning and adjusting the initial GAN structure so that it is able to generate new, visually compelling images based off of Training Set B. Once again, the difficulty is in crafting a GAN that is able to function using the limited number of images we have. This sort of process requires a touch of alchemy in order to succeed.

Next Steps:

The next benchmarks in the process will be to keep iterating with the GAN's structure until we achieve something that resembles my paintings.

We hope to have a functioning and proficient GAN in the next few months and will continue to finesse it over the course of the next several years.

Intended Outcomes:

What does the future project look like?

Would like to have completed the initial build and have a system that is able to produce images that, while perhaps not as high in quality, do represent similar imagery and replicate my artistic style. Keep adding to the quantity of data

In the next six months:

Would like to have completed the initial build and have a system that is able to produce images that, while perhaps not as high in quality, do represent similar imagery and replicate my artistic style. Keep adding to the quantity of data

In the next year:

Would like to have completed the initial build and have a system that is able to produce images that, while perhaps not as high in quality, do represent similar imagery and replicate my artistic style. Keep adding to the quantity of data

In the next year:

In the next five years:

In the Next Five Years:

We will continue to develop and expand on what the GAN is able to do and find ways to integrate more training data so as to give it equal footing with me.

This will include trying to find ways to leverage data so that it is maybe even one step ahead of me - allowing it to keep track of daily changes and advances in science, popular culture, art, etc.

In the Next Year:

We will have continued to develop and track the GAN's progress and will have continued to increased its online social presence.

During this period of time we also hope to have an exhibition featuring its work alongside my own, detailing social responses to its work and development.

In the Next Six Months:

We would like to have completed the initial build and have a system that is able to produce images that, while perhaps not as high in quality, do demonstrate similar imagery and replicate my artistic style. Over this period of time we will also keep adding to the quantity of data