The following introduction about the creation of artificial intelligence abstract art is intended for less technical readers and tries to present the topic as simply as possible.
What is a generative model?
The core for the generation of abstract art by artificial intelligence is a so called generative model. A generative model is able to generate new data based on a large amount of data it was trained with.
In the case of modern abstract or concrete art, this means that the model has been trained with a data set of a large number of real paintings and is able to generate new paintings after the training phase. The model has been designed to learn the general principles and develop the ability to generate new works of art that appear real. These data is generally mentioned as training data.
The model tries to determine so called features from the individual pixels of the training data. This is a complex task in the context of image generation for modern artworks, as the pixels vary widely in their arrangement and value (e.g. colour or intensity). So the generative process is all about to capture the characteristics of the real artworks.
Another characteristic of a generative model is that it is not deterministic but rather probabilistic. In concrete terms, this means that the model contains a random component. A deterministic model would always contain defined and reproducible states and with the same input the same output would always follow.
The spreading of generative models
In addition to generative models, discriminative models are very common in the field of machine learning, since a large number of problems in industry can be solved using these models. Furthermore, generative models are more difficult to evaluate in comparison, since the quality of the results is often subjective.
An example of a discriminative model is the classification of images. The base here is also a large amount of data, however these are labelled and the class is known. In the field of art, an example would be the classification of paintings according to their artist. The procedure here would be that the discriminative model is trained from labelled paintings and after the training phase it is able to recognize by whom (e.g. Van Gogh or Leonardo Da Vinci) a new painting has been created – that means we discriminate between artworks by the creator.
The progress of generative models has greatly increased in recent years. On the one hand, this increase is due to the development of new processes and methods from the field of deep learning, such as so called GANs (generative adversarial networks) or autoencoders. On the other hand, the quality of these models has received an enormous boost due to the ever increasing amount of data, since the quantity and quality of training data is the biggest contribution to a successful outcome. Further exemplary applications are the generation of music, videos, whole articles or books.
The challenges for generative models
Works of art often consist of complex structures and contexts with different dependencies. The model must therefore be able to reflect this complexity. Furthermore, these diversities must be presented in the training data with a certain distribution. Basically there are two main challenges for the creation of a successful generative model:
1. The amount of data must be sufficiently large
2. The algorithm behind the model must be able to recognize and learn these complex interrelationships. (Note: a concrete amount of data cannot be done in advance and is based on experience)
Today, probably the best known, most promising and most widespread generative model is the Generative Adversarial Network (short: GAN – Goodfellow et al., 2014).
Concept Generative Adversarial Network
In its simplest form, a GAN consists of two models, or named Artificial Neural Networks (ANNs). These networks are highly complex and have millions of parameters and structures. The special thing about the architecture is that these two models work in competition with each other. One model is acting as a generator, which produces new data, while the second model is acting as a discriminator and has the task of determining authenticity.
The discriminator is fed from two sources: 1. A large number of real images from the field of abstract/modern art. 2. The newly generated paintings created by the generator.
The aim of the discriminator is to identify that the images of the generator are not real. Whereas the goal of the generator is to create images that appear as much real as possible that the discriminator does not recognize them as fake.
This process of working against each other takes place in the training phase and incrementally an optimization of the two models is performed using a special and complex procedure. For the generator this optimization means that it is becoming more and more capable of producing abstract paintings, and for the discriminator it means that it is improving continuously to detect what forces the generator to improve its paintings.
Due to this principle and the highly complex structure, the generator is able to create new works of art.
See following some of our results