INTERSPECIFICS

CODEX VIRTUALIS_
GENESIS
 CODEX VIRTUALIS_GENESIS

INDEX

[01] /INTRODUCTION

[02] /CODEX
VIRTUALIS_
GENESIS

[03] /TECHNOLOGY ROADMAP

3.1 Network Training
3.2 Image sampling
3.3 Latent space exploration
3.4 Transfer Learning
3.5 Clustering / Self-labelling
3.6 Class A Cycles
3.7 Dimensionality-
Reduction Visualization
3.8 Node Selection
3.9 Class-B Cycles
3.10 Taxonomic navigation
3.11 3D Modelling

[04] /CREDITS

[05] /INTERSPECIFICS

[1]
I N T E R S P E C I FI C S

/INTRODUCTION

Codex Virtualis is an AI-art-science research framework oriented

towards the image synthesis and evolution of an open-ended
taxonomic collection of new-to-nature speculative life forms. An
aesthetic journey through an ecosystem of neural networks and
algorithms that reflects on the role form and association account
for changes in the natural world.

The project, designed to be delivered in four phases, begins with

Codex Virtualis_Genesis: the origin, initial classification, and
multidimensional representation of these entities. Continues
with Codex Virtualis _ Habitat, establishing a relationship with
a meta-environment by exposure to simulated inputs. Later, in
Codex Virtualis _ Emergence, intra-activity is ensued between
living biological samples to try and transfer features in real-time. To
finally test the resilience of one of these organisms and see it living
virtually in Codex Virtualis _Life. A long-scope research strategy
and distinctive feature of previous projects.

[2]
CODEX VIRTUALIS_GENESIS

[3]
I N T E R S P E C I FI C S

/I N T R O D U C T I O N

Codex Virtualis is rooted in a field of the second level and the human-machine
analogies that symbolically compares intentionality enactment is the third level
genetic expression with probabilistic data of symbiotic relationships embedded in this
distribution, transfer learning with genetic project.
recombination, and horizontal gene
transfer with style transfer to materialize a Therefore, Codex Virtualis can be
generative morphogenesis source in a virtual understood as a readymade of technology
environment. The two main inputs fed into and theoretical frameworks that seek
and orchestrated by our artificial generative to sharpen our perception of the
system are microscopy and cellular automata creative function in machine terms, and
images. A dynamic feature transfer between questions conventional definitions of life,
biotics Naro and Agua datasets, and Bert experimenting with algorithmic behaviors
Chan’s 2019 automaton project Lenia. In that progressively become novel lifeforms in
this sense, within the core of this project lies themselves.
the exploration of a life and life-simulation
morphological interplay. A space from which The project joins a cultural and artistic
a meta-deep biosphere emerges within the contemporary praxis inquiring around
technosphere. the concept of life through codices that
recount ancestral cosmogonies, mechanical
The symbiotic narrative of evolution is representations, computational simulations,
for Codex Virtualis a conceptual ground and mechanisms of life. Building an
to speculate on different levels of fusion epistemic bridge between microbiological
and metamorphosis. In the first level, a knowledge and the artificial intelligence
selection of resilient model organisms: connectivism approach, Codex Virtualis
Chroococcidiopsis, Deinococcus imagines ways of living together, of deep
radiodurans, Bacillus subtilis, and Xanthoria interspecific relations that may enhance our
Elegans, constitute the initial line of kinship possibilities of survival.
from which a series of morphological
symbioses will take place. The cooperative
nature of an algorithmic ecosystem in

[4]
 CODEX VIRTUALIS_GENESIS

Model Organisms_

[5]
I N T E R S P E C I FI C S

[6]
 CODEX VIRTUALIS_GENESIS

/CODEX
VIRTUALIS_
GENESIS
In a broader metaphorical sense, the AI Often in artificial genetic modeling,
generative models used in the workflow of symbiotic evolution is described by genetic
our multi-layered computational ecosystem fusion operators within a genotype
become analogous to the role of evolution representation level. However, in our GAN
within biology. In this process, a change in architectures, coevolution is implemented
the gene pool of a population (genotype via a game-theoretical phenotype level
expression) can produce a continuous approach, in which information exchange,
stream of novel organisms (phenotype deception and imitation constitute a
selection). Similarly, by using in silico primordial set of inner interactions that
evolutionary architectures, we explore lead to a divergent, continuously evolving,
different algorithmic recipes under which collection of visual forms incited by the
the biological concepts of variation, heredity, generative and discriminative network
fusion, and cooperation can aesthetically interplay. In addition, we embed in the last
manifest themselves and lead to de novo stage of our system a perceptive layer, which
hybrid morphological profiles. has the role of directing the exploration
within the latent space of our network by
One of the fundamental processes some visual attribute metrics extracted by
adopted within our system is the mixed computer vision algorithms. We interpret
training of our generative adversarial this exploration as navigating through an
networks (StyleGAN2). In this task, two abstract fitness landscape where the variation
qualitatively different datasets merge by along a morphology-related parameter results
adopting domain adaptation machine in specific morphotypes.
learning strategies, leading to the emergence
of new visual samples. These bimodal Finally, using 16s and 18s ribosomal RNA
architectures allow us to explicitly represent sequences from model organisms and
an information transfer between domains, language model GPT-NEO, we generated
analogous to the horizontal gene transfer DNA sequences to create hybrids between
mechanisms in biology. The latter being an the biotic and the generative DNA segments.
increasingly acknowledged relevant source
of evolutionary novelty due to genome
acquisition, mutualistic interactions, and
synergy between organisms.

[7]
I N T E R S P E C I FI C S

[8]
 CODEX VIRTUALIS_GENESIS

/TECHNOLOGY
ROADMAP

[9]
I N T E R S P E C I FI C S

1. NETWORK TRAINING 2A. IMAGE SAMPLING 2B. IMAGE SAMPLING SET 3

sci.db ig.db naro.db

seed

images

images
normalized l
el de
od mo
.m et.
et nn
nn
tf.records

x(seeds)
x(seed)

img_seed

el
od
t.m Image set
e
nn

6. CLASS-A CYCLES 7. DIMENSIONALITY REDUCTION 8. NODES SELECTION 9

Labels Ys Zs Ys
Ys Xs

Convex Hull
Path
Visualize Z Vectors
Dimensionality
reduction

Handpick triads

T z1 z2 z3

Loops_A 2D / 3D Positions
LA(seeds)

[10]
 CODEX VIRTUALIS_GENESIS

3. LATENT SPACE EXPLORATION 4. TRANSFER LEARNING 5. CLUSTERING/SELF-LABEL

Path/ image set

polygon I(X)

+
l
de
mo
et. tf.records
nn
XS

Clustering
>
Xs

Labels Y(I)
Latent Walk
Video

9. CLASS-B CYCLES 10. TAXONOMIC NAVIGATION 11. 3D MODELLING (SOFTSHELL)

Latent Walk
Video

Xs Ys Zs Is

Feature Families

Pixel
Mapping
Interpolation

Heightmap

Loops_B Taxon Navigator

LB(Ts)
3D Body

[11]
I N T E R S P E C I FI C S

[12]
 CODEX VIRTUALIS_GENESIS

3.1 NETWORK TRAINING

sci.db ig.db naro.db

We iteratively calculate the
connection weights between neurons
by continuous exposure to real-life
images to find the optimal values images
that produce non-existent though
statistically accurate samples. Next, a
base StyleGan2 model is trained on
the following microorganism data images
sets: normalized

AGUA: 1,300 crowdsourced

image collection. tf.records
NARO: 1,200 images from
the NARO genebank public
collection.

LENIA: 1,500 image

collection from continuous
cellular automata Lenia.

MODEL ORGANISMS:
extreme condition resilient l
de
organism collection inspired
.mo
by astrobiology research. et
nn

[13]
I N T E R S P E C I FI C S

[14]
 CODEX VIRTUALIS_GENESIS

3.2 IMAGE SAMPLING We select points from the

n-dimensional space of the
network to infer and generate a
corresponding image. This action
is repeated to group samples with
similar spatial distribution features.

seed

el
de
l mod
mo .
. et
et nn
nn
x(seeds)
x(seed)

img_seed

Image set

[15]
I N T E R S P E C I FI C S

[16]
 CODEX VIRTUALIS_GENESIS

3.3 LATENT SPACE

EXPLORATION

Path/
Latent—or hidden—space refers to polygon
the abstract dimension containing
the attribute values of an image. It
cannot be interpreted directly but
encodes meaningful representational
data. In order to identify relationships
between image attributes and the
latent space region that generates
them, a compression and clustering el
phase is set to identify structural mod
.
et
similarities and trajectories that nn
produce interesting visual patterns.
Xs

Latent Walk
Video

[17]
I N T E R S P E C I FI C S

[E]

[24]
 CODEX VIRTUALIS_GENESIS

3.4 TRANSFER LEARNING

To transfer visual elements from

one image collection to another,
we implement a transfer learning

+
technique that replaces all images at
an advanced phase in training with
a new set of images. This technique
pushes the model to transfer features
from dataset A to dataset B, creating tf.records
variation in visual components while XS
maintaining attributes belonging to

>
both datasets, a process analogous to
genetic recombination in biology.

The following base/target

combinations are used to create
mutated families of organisms
descended from the original
organisms in dataset A + traits in
dataset B:

[A] AGUA + NARO

[B] AGUA + BS ( Bacillus subtilis)

[C] AGUA + CH ( Chroococcidiopsis)

[D] AGUA + DR ( Deinococcus

radiodurans)

[E] AGUA + XE ( Xanthoria elegans)

[F] AGUA + LENIA (algorithmic

organisms datasets)

[25]
I N T E R S P E C I FI C S

[26]
 CODEX VIRTUALIS_GENESIS

3.5 CLUSTERING /
SELF-LABELLING
image set
I(X)
We perform image sampling to
generate the corresponding set
of images for this training stage.
As mentioned above, every set of
hybrid organisms create new visual-
attribute to latent-space-coordinate
relationships.Groups of similar
images are identified using K-means
and Agglomerative Clustering so as
to find out how these relationships
Clustering

may work. This clustering step assigns

each hybrid organism family a group
id or label, and organisms in the same
group share visual characteristics of
shape, color, and texture. The latter
makes it easier to identify latent space
vector points that could conform to
a neighborhood, in other words, a
region in space where a configuration
of similar visual features is
maintained, and to which to generate
cycles that exhibit uniform behavior.
Labels Y(I)

[27]
I N T E R S P E C I FI C S

[28]
 CODEX VIRTUALIS_GENESIS

3.6 CLASS-A CYCLES

Labels Ys
A cycle refers to a video where a
newly generated organism shows an
organized, uniform, and continuous
behavior. These sequences are
produced by traversing the trajectory Convex Hull
of a given set of points in the latent
Path
space, interpolating the intermediate
ones, and inferring a frame for each
of them. Class-A cycles are generated
automatically by selecting an ordered
cluster subset—which represents a
family—, sampling a point within
each individual cluster’s convex hull
and tracing the trajectory between
them.

Loops_A
LA(seeds)

[29]
I N T E R S P E C I FI C S

[30]
 CODEX VIRTUALIS_GENESIS

3.7 DIMENSIONALITY-
REDUCTION VISUALIZATION
Ys Xs

Moving through points in latent

space allows the appearance of the
resulting images to be modified.
However, it is difficult to visualize
and interpret their location and
displacement direction in this abstract Dimensionality
geometric dimension. To facilitate
reduction
this exploration and identify regions
of interest, we project the vectors of
the sampled organisms into a lower-
dimensional resolution. The resulting
2D and 3D visualizations show the
corresponding points for every image
and allow a selection of the target
transition’s direction to form another
class of cycles.

2D / 3D Positions

[31]
I N T E R S P E C I FI C S

[32]
 CODEX VIRTUALIS_GENESIS

3.8 NODE SELECTION

Zs Ys
After visualizing points in the latent
space, we manually select groups
of points to visit through in an
interpolation path to generate class-B
cycles. The purpose of this manual
Visualize Z Vectors
choice is to find the most suitable
individuals that will constitute
our genetic pool to form a new
generation.

Handpick triads

T z1 z2 z3

[33]
I N T E R S P E C I FI C S

3.9 CLASS-B CYCLES

Class-B cycles contain transitions

between latent space points manually
selected through the low-dimensional
visualizations.Selecting for organized,
uniform and continuous behavior,
the manual selection criterion acts as
a fitness function, allowing to merge
features of organisms and targeted
mutations.

Interpolation

Loops_B
LB(Ts)

[34]
 CODEX VIRTUALIS_GENESIS

3.10 TAXONOMIC
NAVIGATION

The agglomerative clustering in Xs Ys Zs Is

phase 4 allows us to build a hierarchy
of similarity between groups or
neighborhoods. By adding this
hierarchical organization to a 3D
visualization—where each node is an
organism connected to others in the
same group by proximity—we obtain
a form of taxonomic division of the Feature Families
population for each family.

Taxon Navigator

[35]
I N T E R S P E C I FI C S

[36]
CODEX VIRTUALIS_GENESIS

[37]
I N T E R S P E C I FI C S

[38]
 CODEX VIRTUALIS_GENESIS

3.11 3D MODELLING
Latent Walk
Video

We model a three-dimensional outer

membrane body by drawing points
in a 3D environment using videos
as procedural textures.To this end,
we map the x/y position, draw two
equidistant points to the base plane
in opposite directions, and use the
video as a heightmap of positive and
negative values. Having an opposite
plane gives the generated body a Pixel
spherical shape. Mapping

This process solves the need to

abandon the two-dimensional space
of the StyleGAN2 results and—
in future phases—will allow the
organism to receive environmental Heightmap
inputs and respond via reshaping its
softshell.

3D Bodie

[39]
I N T E R S P E C I FI C S

[44]
CODEX VIRTUALIS_GENESIS

[45]
I N T E R S P E C I FI C S

/CREDITS
INTERSPECIFICS / Codex Virtualis_ Genesis was
created during the SETI X AI @ Ars
Leslie García Electronica Futurelab 2021 artist
Paloma López
residency.
Emmanuel Anguiano
Felipe Rebolledo
Alfredo Lozano The SETI x AI residency is part of
Maro Pebo the AI Lab (European ARTificial
Carles Tardío
Nickole Klinckwort Intelligence Lab). It offers
international artists working in
AI a chance to win a residency at a
[PARTNERS] scientific partner institution and the
Futurelab of Ars Electronica.Based
in Linz, Ars Electronica is a globally
Ars Electronica
unique platform for art, technology,
Futurelab / and society that has been analyzing
and commenting on the digital
European Artificial
Intelligence Lab / revolution since 1979. The focus is
always on current developments and
SETI Institute /
possible future scenarios and how
Co-founded by the they will change people’s lives.
Creative European
Program of the
European Union / The SETI Institute Artist In
Residence (Air) Program connects
contemporary artists with SETI
Institute researchers and facilitates
an exchange of ideas to catalyze new
perspectives, insights, and modes
[CONTACT] of comprehension.Founded in
MAIL / 1984, the SETI Institute is a non-
[email protected] profit, multi-disciplinary research
and education organization whose
interspecifics.cc mission is to explore, understand, and
int-lab.cc/codex
explain the origin and nature of life
in the universe and the evolution of
intelligence.

[46]
CODEX VIRTUALIS_GENESIS

[47]
I N T E R S P E C I FI C S

INTERSPECIFICS

Independent artistic research bureau Our work has been supported by

founded in Mexico City in 2013. We DAAD, International Cities for
have focused our research on the use Advanced Sound, Laboratorio Arte
of sound and A.I., to explore patterns Alameda, Telefonica Foundation,
emerging from biosignals and the Bancomer BBVA Foundation,
morphology of different living The National Fund for Culture
organisms as a potential form of non- and the Arts in México. Bauhaus-
human communication. With this Universität Weimar and Universitat
aim, we have developed a collection of der Kunste Berlin in Germany.
experimental research and education Museum of Modern art in Medellin
tools we call Ontological Machines. in Colombia. National Council for
Our work is deeply shaped by the Culture and the Arts and Museum
Latin American context where of Contemporary art in Chile and
precarity enables creative action shown at FACT Liverpool, European
and ancient technologies meet Congress for Artificial Intelligence in
cutting-edge forms of production. York, Spektrum, Acud Macht Neu,
Our current lines of research are CTM Festival, and CLB in Berlin,
shifting towards exploring the hard ICAS Festival in Dresden, Casa del
problem of consciousness and the Lago, Centro Cultural de España,
close relationship between mind and INDEX at the University Museum
matter, where magic appears to be of Contemporary Art, Tamayo
fundamental. Sound remains our Museum.
interface to the universe.

[48]
CODEX VIRTUALIS_
GENESIS
INTERSPECIFICS

Mexico City 2021.

INT-LAB.CC/CODEX