side-area-logo

Sempervivum Rickshaw and other stories

Cultivation, Evolution, Transition, Variation,

A Sempervivum Rickshaw with Brennan Goddard

 

Limited edition of 20 copies

signed by the artists/authors

Printed by Vassel Graphique 2021

Cultivation,

Native to the mountain ranges of Europe, Morocco and Iran, Sempervivum evolved from ancestral plants five to nine million years ago. Today 4,000 hybrids have been identified, each evolving different adaptations to survive the extreme mountainous climates.

The beauty and fractal composition of Sempervivum grows from a logarithmic spiral. Their succulent leaves form rosettes which vary between fine to broadly lobed with supple nodes, vicious jagged edges, covered either entirely with filaments and other varieties have eccentric tufts and swirls often sometimes covering the entire plant.

Summer 2019 I planted on my roof 2000 plants grown in the nursery of Franck Poly, producer of Sempervivum. Having observed them change through the seasons, I grew a desire to document their magnificent multiple survival techniques, textures and colours. Perhaps it is not only their differences but also their soft mammal-like forms, my excitement of seeing the “daughter” rosettes growing awkwardly from their “mother” or their impossibly complicated geometric patterns that challenged me to paint 101 Sempervivum portraits.

Painted at the foot of the Jura mountains, the following drawings depict a mere handful of sempervivum natural hybrids and cultivars. 

 

Evolution

At the start of the neural network training, the results are random noise, with no meaning or evident similarity to the original pictures. But, as in natural selection, despite the initial poor quality, some of the images are slightly more similar to the originals than the others, for some settings of the artificial neurons. In the next generation, the neuron settings are adjusted to increase this similarity, in a very gradual way. This process is repeated, and the images produced become a tiny, tiny bit closer to the originals. After tens of thousands of generations, the output images become much more similar to the originals. This numerical evolution occurs over hours or days, instead of aeons in the natural world. In the Evolution series, each set of images is generated from the same set of seed latent numbers Z, and the collection shows how each digital Sempervivum evolves from noise.

 

Transition,

In the Generative Adversarial Networks used for Sempervivum Rickshaw, two separate neural networks compete. One, the encoder, takes a digitised image of millions of pixel values as its input, and compresses the information contained in the image through many convolutional layers of neurons into a set of latent numbers. This is the Z space, with 512 different values which describe the features of the image. The decoder network does the reverse: it takes the 512 latent numbers as its input, and from these produces the millions of pixels corresponding to an image. Choosing 512 different numbers in the Z space gives different images outputs – much as another DNA sequence in the natural world produces a different living organism. Changing the Z numbers modifies the digital code and morphs the artificial Sempervivum. In the Transition series, a line is drawn in 512-dimensional space between two Z numbers encoding different Sempervivum paintings. Points sampled from this line produce new hybrids as the forms transition from one into the other.

 

Variation

With the AI Generative Adversarial Network, the latent Z number or digital code can be varied to change the output image. There are an uncountably large number of image possibilities. Even if each of the 512 Z numbers were just 0 or 1, there would still be 10 to the power of 154 different combinations – one followed by 154 zeros. There are so many possible combinations that the code which produces the images closest to the original painting is impossible to know. The best that can be done is to search the Z space by trying a random set of numbers, checking the results against the original, then taking the best number as the basis for the new search, the ‘parent’, and repeating for thousands of generations. This kind of genetic algorithm search is based exactly on evolution through survival of the fittest. Using it, a Z number can be retrieved to generate an image which corresponds closely to each original. In Variation, images of the original paintings are surrounded by close relatives, where the Z code has been changed a small random amount.