What is Latentism?

Artificial Intelligence Art as Exploration of the Latent Space

 

            In recent years, Artificial Intelligence (AI) has made tremendous progress with regards to working with images. This progress has been made possible by hardware innovations in GPUs as well as new techniques related to deep learning. Artists and Painters on the other side still use their imagination to put on canvas the message they want to convey. We propose to use AI as a tool which can supplement the artist’s imagination in creating new striking images. This tool is difficult to implement and presents several barriers to entry but by its nature it follows in an explicit way what the artist does in an implicit way.

 

            No artist lives in isolation, and each artist finds himself the outcome of a cultural landshaft that makes up his perception and his sensitivity. All artist are also influenced by what artists of the same generation are producing, on top of being aware of what had been done before them by historical masters. Artists are also the product of their social environment and of the political status of the area they live in. Finally they all strive to develop their own twists and their own uniqueness with their approach to art. In this way, the artists must look into the past and be aware of the present. Once this step is done, for example while studying the history of art in Academies, then they can use their imagination to create new art, but the root of their imagination will be in all that substrate that is now impregnating them. What if there was a way to make this process more straightforward and analytical. AI is here to do just that.

 

            Recently, new structures such has convolutional neural networks (CNN) have allowed helping solving problems in fields as diverse as medicine, image recognition and self-driving cars. Schmidhuber, one of the fathers of modern AI has described Science as the history of compression progress. Indeed, when the human mind can generalize phenomena observed in the wild through a formula then it can compress an enormous amount of information in a simple algorithm taking just a few lines of code to run. Kepler, Newton and Einstein have all managed to generalize observed phenomena and derive a formula which can summarize this phenomena in a compact form. Similarly neural networks transform initial data given to them and map this data into a different dimension called the “latent space” in order to make things looking similar in the real world being as close as possible in that new dimension, in order to make their separation and classification easier.

           

            When applied to visual arts, which in the end are just images, the process is the same. Images are taken as input and some neural network maps them to a latent space where there is more order. Other networks can now take points in that latent space and turn them into pictures. So points close by in the latent space will tend to produce similar images. If there was an algorithm capable of first compressing and then generating images then by just feeding as many images as possible to it, one could produce new art just by exploring that latent space. We now call it “Latentism”.

 

            Such algorithms actually do exists and come in different shapes and forms: GANs, VAE and a few others allow to generate new images by following a given pipeline. For example in the case of GANs, there are two competing networks. One tries to produce images as best as it can to fit into some category, and the other does as best as it can to distinguish between those fakes and the real true images belonging to that category. By competing against each other the first network eventually gets very good at generating images. Following this, one can feed images of paintings to the network and let it map those images to the latent space. This is similar to what the artist does when he studies the ancient masters. Then by exploring the latent space, the artist can generate new fresh pictures and this is similar to what the artist does when using his imagination to produce new art. So we can draw a very close parallel between what artist do and what can be done with the help of AI.

 

            This method can raise a few criticisms. The first one is that the AI lacks imagination and cannot produce anything new or draw object it has never seen, so that art is trapped in plainly repeating what has been done before. We reject this criticism for the simple reason that this AI is trained via a random process. At the core of the training strategy used to teach them how to encode this information in the latent space is randomness. This source of randomness introduces unexpected variations into the images and what was supposed to be a book can become a guitar or a hat can become a shark on a lady’s head. Just as artists use their imagination to introduce subtle variations in their thought, the algorithms are using random numbers to veer the image generation process towards an unpredictable and surprising outcome.

 

            Another criticism is that what is generated by the AI is not art. We also reject this criticism in the sense that the actual art there is the process used to generate new image. The result of the process is the image, but the art is the process. And if this is art, then the role of the artist is to sift through possible outcomes while traveling the latent space and select those pictures which respond to his wishes. The artist might be looking for the aesthetic value in a picture or for an image which will surprise the viewers of his work. With his objective in hand, he can start his journey.

 

            Latentism is a form of art which use AI to help the artists’s imagination in producing new images. It is still the artist responsibility to translate those images into an oil painting, a drawing or an etching. It must be noted that there are several barriers to entry which make this art difficult. First to gain an understanding of the theoretical aspects of neural networks takes some long and quantitative education which the most majority of artists to not have to this day. Then to implement those algorithms, it requires to have experience in coding and programming which once again is not very likely given the characteristics of current art education. Finally, in order to train those AI algorithm, one needs extremely powerful computing machines which are not even available for sale to the large public. Training those algorithms on a simple computer would take months. So it really takes the convergence of the cooperation of highly trained mathematicians and inspired artists to put all this together.

 

            To conclude, while all those ideas seem extremely new because they can now be implemented in the real world, they are in fact quite similar to Plato’s notion of the Theory of Forms. It postulates that there are perfect concepts which exist in a different dimension of their own and that object and concepts in our world are mere reflections of those perfect shapes. What Latentism does is that in a Platonician sense it goes out to explore that world of perfect concepts, of perfect images and of perfect paintings and allows to bring back glimpses of it via generated images. What seems to be like a new idea has actually already been postulated 2500 years ago.