thesis project on biological artificial intelligence

1. Università degli studi di Trieste
Dipartimento di Ingegneria e Architettura
Laurea Triennale in Ingegneria Elettronica e Informatica
Summary of: “Ortus: an Emotion-Driven
Approach to Artificial Biological
Intelligence”
Supervisor:
Prof. Eric Medvet
Student:
Martina Silvestrini
matr. IN0500478
Anno Accademico 2020/2021
Trieste

2. Contents
Introduction 2
General Characteristics 2
System Design 2
Experiment 4
Future Developments 5
Conclusion 5
References 7
1

3. Introduction
This Thesis consist of a summary of the experimental project conducted by Andrew W.E. Mc-
Donald, Sean Grimes and David Breen at the Drexel University, Philadelphia and published in
August 2020 with the title of: “Ortus: an emotion-driven approach to artificial biological intel-
ligence”.
The article aimed to investigate the increasing possibilities and potentials of biological artificial
intelligence and its applicability. The study has been conducted through the implementation of
an artificial organism called Ortus developed with the purpose of testing this new biologically-
base learning approach .
General Characteristics
Ortus is a virtual system that exhibit biological intelligence, composed of a neural network
made of biologically-inspired non-spiking neurons and a not-well characterized body, living in a
biologically-alike environment with air to breath and water. It interacts with its environment
though sensory neurons and act in response to it through motor neurons (output layer of the
network). It has an emotional-driven intelligence thanks to which it is able to learn and therefore
refine its network during all its life.
The model organism upon which Ortus’ non-spiking and specialized neurons are based on is
the nematode C.Elegans, chosen as a sample for its neural network simplicity (it only has 302
neuron) and efficiency.
System Design
Ortus is virtual organism whose parts (ODE) were drafted based on their biological correspond-
ing. Each element of the design is implemented in the neural network using either different types
of neurons, learning mechanisms or rules of construction.
1. ODE1 Respiratory Circuit: it is the fundamental element of any living organism, necessary
to balance O2 and CO2 and therefore permit life sustain. Control of O2 and CO2 level was
assigned to two sensory neurons sO2 and sCO2 that work complementarily: Rising of CO2
level increase the activation of sCO2 neuron; this cause the excitation of lungs to obtain
more oxygen from the external environment and expel the excessive CO2.
Vice versa rising of O2 level cause the increasing activation of sO2 sensory neuron, with
the direct consequence of inhibiting the motor neuron responsible for lungs activation that
implies a decrease in oxygen level and the establishment of equilibrium.
It is well known that the maintenance of O2 / CO2 balance is influenced not only by
internal factors, but also by external one, responsible for emotion arise. For instance, in
a dangerous situation, the increase of fear may lead to the lost of O2 /CO2 balance and
consequently the organism might respond with hyperventilation.
2. ODE2 Psychological Sphere: the emotional response of the organism to external inputs
influences the functioning of the entire system. In Ortus’ implementation two emotions
(fear and pleasure) where featured through emotional interneurons, respectively eFear and
ePleasure. Through them associative learning was built, aiming to re-conduct new stimuli
to well-known ones thanks to these emotions.
2

4. Associative learning in animal behavior is defined as any learning process in which a new
response becomes associated with a particular stimulus. It is a learning mechanism often
associated with conditioning, where two types of stimulus can be detected: an uncondi-
tional stimulus that invariably elicit a certain response (absence of O2 reinforces the eFear
interneuron that communicates to the motor neuron to evoke the inspiration ) and the
conditional stimulus, which ability to elicit a certain response is conditional to a particular
set of experiences (experiencing a new situation when absence of O2 is detected activates
a sense of fear connected with it).
It can also be pointed out that in mammal, the relation inhale pleasure / exhale fear
exist on some level because it is connected with the concepts of survival and longevity
and it is consequently reasonable to suggests that associative learning is funneled through
innate behavior circuit to assign positive or negative emotions to neural sensory stimuli.
3. ODE3 Differentiated Interneurons: 4 classes of interneurons were implemented in Ortus’ :
ˆ SEIs (Sensory Extension Interneurons): they take input from sensory neurons (sO2,
sCO2, sH2O)
ˆ SCIs (Sensory Consolidatory Interneurons): they take input from SEIs, each having
an incoming synaptic weight equal to ( 1
nr. SEIs connected ). It is important that this
interneuron is able to combine multiple sensory inputs to form associations to trigger
certain emotions.
ˆ EEIs (Emotion Extension Interneurons): it is an auxiliary interneuron specific for each
emotion. Each receives input through chemical synapses from SCIs. The idea behind
the chemical synapses from EEI to SCI is to allow an emotional state to trigger or
cause Ortus to remember a stimulus previously evoked. The remembering is defined as
measurably increased activation in the SCI that the EEI onto; the connection between
EEIs and the emotion interneuron happens through a GAP junction.
ˆ Emotion interneurons (eFear, ePleasure)
Figure 1: The diagram shows Ortus’ emotional
integration. The circles are the sensory neu-
rons, while the ovals are the different types of
interneurons. Dark green and red arrows rep-
resent the excitatory chemical synapses, while
light green and pink ones the weak chemical
synapses. The black junctions symbolize bidi-
rectional gap junctions and the dark red binder
between eP0 and eF0 describe an inhibitory re-
lationship that embodies the emotion hierarchy
where Fear overcomes Pleasure.
As for humans and other animals, a hierarchy of emotions is present: this means that
those emotions indispensable for survivor will prevail on the others. In this specific case,
since only two emotions have been defined, fear overcomes pleasure. The implication is
that those interneurons connected with eFear are stronger and therefore, anything learned
through fear is more likely going to remain in memory.
3

5. 4. ODE4 Relations between Elements: Ortus’ basic elements are neurons and muscles;
these elements are connected with 4 possible types of relations:
ˆ [±]A causes [±]B that represent the chemical synapses
ˆ A correlated B : representing the bidirectional gap junction
ˆ A opposes B : where A and B inhibit each other
ˆ A domains B : where A inhibits B
These relations have few attributes such as mutability index (MI) and polarity all set on
standard and static values in this first version of Ortus.
The building process of the network had to simulate the way brain grows itself based on
genetic instructions. Therefore, to accomplish the task, a specific Ortus Development Rules
language was developed by the authors .
5. ODE5 Hebbian learning with Stentian extension: this learning approach has been proven
to be highly representative of the biological learning mechanism .
According to it a synapse strengthens, or weakens whenever occurs the synchronous, or
(respectively) asynchronous, activity of each pre-synaptic neuron connected with it, mean-
ing (in our case) that the combination of multiple, synchronous sensory neurons should
reinforce the chemical synapses connected to the correspondent emotional interneuron.
In Ortus’ implementation this has been done through a mutability index (MI) that defines
a synapse’s potential to be modified.
The Experiment
The purpose of the experiment was testing Ortus’ emotional-driven intelligence. To do so a pe-
riodical immersion into water was simulated using a series of sensory neurons:sO2 , sCO2 sH2O
that acted as the fist layer of Ortus’ neural network. The hidden layer of interneurons SCIs,
SEIs, EEIs, eFear, ePeasure was also implemented, as well as the output layer made of motor
neurons responsible for lungs stimulation.
Training and Conditioning: Ortus was exposed to water via sH2O stimulation 4 times, with
100 time-steps pause to allow neutral injuction to decay. In each exposition fear was avouched
preventing Ortus to inhale O2 and exhale CO2. This was repeated 4 times.
Figure 2: Conditioning phase: CO2 exhalation
is prevented. each time H2O is introduced under
this conditions Ortus’ fear response grows.
Emotional-Driven Learning: after 200 time-steps pause the immersion in water was repeated,
but this time no constrains were applied to the respiratory circuit. The response is shown in
picture:
4

6. Figure 3: Experiment result: after the 4 rounds
of conditioning, exposing Ortus to H2O is suffi-
cient to induce a state of fear.
Explanation: during the learning phase Ortus associated sCO2 activation with fear and
sH2O with sCO2 because any time it was exposed to water sCO2 activation was inducted. The
two sensory stimuli were synchronous enhancing the synaptic connection with eFear, according
to Hebbian learning. For that matter, when Ortus was put into water once again, without
respiratory restrictions, eFear was activated automatically and rapidly, as it was expected after
the conditioning process.
On contrary, without any conditioning sH2O activation led to a light increase of fear due to the
slow realization of the decease of O2 level.
Figure 4: In the absence of conditioning: the
growing of an emotional state of fear is very sub-
tle and slow
Future Developments
Ortus’ first attempt is quite simplistic but there is room for improvement:
1. The static weight (MI) of the synapses can be varied during the network life, to give more
stability to the system.
2. A new approach over neural connections must be think of to add more complexity to the
system without compromising efficiency: for example giving neurons multiple functionali-
ties could reduce network complexity.
3. It would be nice to give Ortus more nuances of the learning spectrum, making it able to
distinguish little sub changes.
4. The learning method could be improved to do not allow the network to unlearn certain
things.
Conclusion
Ortus shows the amazing potential of the biologically-inspired artificial intelligence to replicate
on some level mechanisms and fundamental principles observed in organic systems (e.g., cyclic
5

7. respiratory circuit) and organic nervous systems (e.g., emotionally-driven associative learning).
Nevertheless, the article could have been more precise defining Ortus’ nature and components,
for example, authors could have explain the muscles better, since it is not clear how (and even
if) have they been implemented.
6

8. References
Andrea H. McEwan, Catharine H. Rankin. “Mechanosensory Learning and Memory in Caenorhab-
ditis elegans”. In: Handbook of Behavioral Neuroscience, Volume 22, Issue null, Pages 91-111
22 (1905), pp. 99–111.
Andrew W.E McDonald Sean Grimes, David Breen. “Ortus: an Emotional-Driven approach to
(Artificial) Biological Intelligence”. In: (2020).
Mackintosh, Nicholas John. Animal learning. url: https://www.britannica.com/science/
animal-learning.
Sjöström, Jesper and Wulfram Gerstner. Spike-timing dependent plasticity. url: http://www.
scholarpedia.org/article/Spike-timingdependentplasticity/STDPandHebbianlearn
ingrules.
7