Neural networks

The term artificial intelligence is very broad and is applied to human-mimicking behavior performed by machines. Think household robot, the automatic lawn mower, the self driving car, the industrial robot or the clever chess engine. Some of these systems are created by forming a set of logical rules that determine the outcome and may perform extremely well. But such systems require that the architect creating them can forecast all possible scenarios that the system may be exposed to. If the system reaches a new situation that was not thought of the architect must add new logical rules so that the system can handle them properly. Such systems become increasingly cumbersome as their tasks become more complex (complexity here refers to “being able to make more general choices” rather than solving a specific task).

I am interested in building systems that find the rules of the game themselves and where a small set of initial rules may develop into complicated and (hopefully) sophisticated behavior. Such systems require a different approach than the ones depicted in the previous paragraph. Neural network bear such capacity, but have the drawback that they sometimes behave a bit too unpredictably. In many situations this is a problem, but for a creative process where one often seeks new perspectives it is sometimes advantageous.

Art using neural networks

I started combining neural networks with art in an attempt to become a better artist. I was already working with artificial intelligence and had picked up painting as a hobby. Initially I used the networks in an attempt to classify and find patterns and common themes in artworks. Through this process I found that the neural nets can learn to paint themselves, and the results were fantastic! I tried to unleash increasingly complex networks to create more complicated artworks but began to feel passive in this process. The methods I was using then were Generative Adversarial Networks where two separate neural networks compete against each other and where an increasingly elaborate battle evolves and an increasingly “intelligent” behavior emerges, on the verge of a computational singularity. The challenge here, as the architect, is to find clever ways of avoiding the singularity in order to rise the intelligence and at the same time keep control of the process for meaningful content.

I have since then started to develop methods for staying in control and collaborate more directly with these types of neural networks. There are several interesting directions that I try to pursue. First, I try to develop interesting architectures that increase the creativity of the AI itself and which also allow for more human interaction. The second is more of an hands-on approach where I take turns with the AI and paint. The last approach is to develop specific hardware where the intent is for the human artist to stay in a creative workflow and be able to use the AI as more of a tool, just as another type of brush. I wish I had more time to develop this, because in writing moment I have my hands full with the previous two approaches. Which still are extremely interesting!

A print process

I’ve printed a few of the artworks together with Sandeng Printspace in Gothenburg. The prints are made in limited editions on lovely fine art archival grade paper and are signed and numbered by me. If you are interested in such a print or have seen a particular image that you like, just let me know.