This two-part article proposes a new approach to understanding neuronal mechanisms, still unexplained despite the immense progress in neuroscience since the 1940s.
The first part ("The boolean brain") first presents a brief history of the steps leading to the Convolutional Networks that now rival the performance of the human visual system. The biological plausibility of these networks is examined, leading to the paradoxical conclusion that McCulloch and Pitts' logic model was a correct approach and that it has been underestimated.
A new model of neural networks, the Elective Neural Networks (ENN), is proposed on this basis, inspired by the Theory of Epigenesis by selective stabilization of synapses (Changeux et al., 1973) [1], and equipped with a logical learning mechanism by synapse elimination. Its capacity to form large-sized networks is examined, taking into account connectivity constraints, and its biological plausibility is defended, including the issue of the binary synapse.
The second part ("The orthogonal brain") proposes a neuronal mechanism with an explanation of the learning curve in a classical conditioning: the proboscis extension reflex in the Apis Mellifera bee. A reinforcement learning mechanism is added to the ENN model, applying to both classical and operant conditioning. A general hypothesis on the implementation of effectors control in a brain is deduced, in which no individual synapse is genetically programmed.
The slide format was chosen for this paper because of its ability to represent complex dynamic phenomena.
Elective Neural Networks. II. The orthogonal brain. On a Heuristic Point of View Concerning the Brain Function.
1. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 1/93
Elective neural Networks.
Elective Neural Networks.
II. The orthogonal brain.
On a Heuristic Point of View Concerning the Brain Function.
Claude ABIN
Translated from French by Virginie ABIN
Except otherwise noted, this work is licensed
under a Creative Commons CC BY 4.0 license.
2. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 2/93
Abstract.
This two-part article proposes a new approach to understanding neuronal mechanisms, still
unexplained despite the immense progress in neuroscience since the 1940s.
The first part ("The boolean brain") first presents a brief history of the steps leading to the
Convolutional Networks that now rival the performance of the human visual system. The biological
plausibility of these networks is examined, leading to the paradoxical conclusion that McCulloch and
Pitts' logic model was a correct approach and that it has been underestimated.
A new model of neural networks, the Elective Neural Networks (ENN), is proposed on this basis,
inspired by the Theory of Epigenesis by selective stabilization of synapses (Changeux et al., 1973) [1],
and equipped with a logical learning mechanism by synapse elimination. Its capacity to form large-
sized networks is examined, taking into account connectivity constraints, and its biological
plausibility is defended, including the issue of the binary synapse.
The second part ("The orthogonal brain") proposes a neuronal mechanism with an explanation of the
learning curve in a classical conditioning: the proboscis extension reflex in the Apis Mellifera bee. A
reinforcement learning mechanism is added to the ENN model, applying to both classical and operant
conditioning. A general hypothesis on the implementation of effectors control in a brain is deduced,
in which no individual synapse is genetically programmed.
The slide format was chosen for this paper because of its ability to represent complex dynamic
phenomena.
Elective neural Networks.
4. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 4/93
Summary of the second part :
1. PER Conditioning : Facts of the case.
2. PER Conditioning : Proposals of the ENN model.
Two types of effectors.
Type II Learning.
Type II Hyperdomain.
Effector arbitration.
Effector control.
PER conditioning.
3. The Orthogonal Brain.
4. Conclusion : Process of brain development.
5. Bibliography.
Elective Neural Networks.
5. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 5/93
The most exciting phrase to hear in science, the one that heralds
the most discoveries, is not "Eureka!", but "That's funny…".
Isaac Asimov
Elective Neural Networks.
6. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 6/93
Elective Neural Networks - PER : Facts of the case.
1. PER Conditioning : Facts of the case.
2. PER Conditioning : Proposals of the ENN model.
Two types of effectors.
Type II Learning.
Type II Hyperdomain.
Effector arbitration.
Effector control.
PER conditioning.
3. The Orthogonal Brain.
4. Conclusion : Process of brain development.
5. Bibliography.
7. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 7/93
1.1. Ivan Pavlov : classical conditioning (1898).
It is considered that the first experiments on learning and memory in animals were carried out by
Ivan Pavlov on dogs in the 1890s. While studying the physiology of digestion by collecting their
saliva in a fistula, he noticed that the salivary glands became active at unexpected times, before the
food was presented.
Through experimentation, Pavlov established the conditions under which this salivation, an
innate reflex, could be triggered by various visual and sound stimuli. His experiments, published in
Russian, were known in the West by isolated publications and earned him the Nobel Prize in 1904.
The whole was only translated into English in 1927 [2].
Pavlov had discovered the concept of the "conditioned reflex": a neutral conditional stimulus
(ringing) can be associated with an unconditional stimulus (presentation of food) that naturally
triggers an innate reflex (salivation). This conditional stimulus then also triggers this reflex
(acquired reflex). This is the "Pavlovian conditioning" or "classical conditioning". The experimental
paradigm imagined by Pavlov in 1898 has since been used in an enormous amount of work.
Elective Neural Networks - PER : Facts of the case.
8. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 8/93
1.2. Edward Thorndike : the "Law of Effect" (1898).
It was also in 1898 that Edward Thorndike published a thesis on animal intelligence [3] [4]. In it,
he demonstrated the abilities of association of cats, dogs and chickens through a system of "Puzzle
Boxes", from which they had to find the mechanism to escape in order to obtain the food presented
outside. The animal moves around in the cage, discovers by chance the action that opens the door
and then remembers it: in later tests, it finds the solution more and more quickly.
This is Thorndike's "law of effect": satisfactory answers are strengthened, unsatisfactory
answers weakened. He later wrote: "Reward, with its double effect as a confirming reaction and as a
satisfier, operates dependably and well nigh universally as a strengthener of a connection" [5].
This theory of Trial and Error Learning, based on the stimulus-response pattern (S-R) of
behaviorist psychology, was called at the time "Connectionist theory of the mind". Thorndike's
conception was that in all animal species, some neurobiological connections tend to elicit a
particular type of response (S-R bonds). "Some of these connections have already been established
in the normal animal at the time of birth; others, which he is potentially capable of forming, are
acquired as a result of post-natal experiences" [6].
In 2002, J.P. Changeux will take up Thorndike's conception to extend it to the neural bases of
what he will call "learning by selection", at the level of cognitive tasks [7].
Elective Neural Networks - PER : Facts of the case.
9. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 9/93
1.3. Burrhus Skinner : operant conditioning (1936).
Thorndike's Law of Effect laid down the founding principle of "operant" (or "instrumental")
conditioning, which was the main subject of Burrhus Skinner's work from 1936 [8] [9]. Its
experimental device, called "Skinner's Box", made it possible to observe the reactions of rats
subjected to various stimuli and whose behavioral responses (moving, pressing on a lever) were
rewarded (food) or punished (electric shocks). Unlike Pavlovian conditioning where we manage to
trigger an innate reflex by a neutral stimulus, operant conditioning has the effect of strengthening
(or weakening) one of the possible behaviors of the animal in response to a particular stimulus.
After confirming Pavlov's discoveries on a large number of vertebrates and invertebrates, the
researchers finally established the latter as standard models of classical conditioning, because
their relatively simple nervous systems make it possible to trace associative phenomena at the
cellular and molecular levels [10].
Insects, in particular, are a fascinating object of study for their sophisticated behaviour [11]. In
1967, Karl Von Frisch discovered the famous "waggle dance", by which the bee communicates to its
fellow bees the direction and distance of a food source [12].
It has been shown that bees see the world in color [13] [14], perceive shapes and patterns [15],
detect their symmetry [16], and have rich and complex olfactory and mechanosensory perceptions
[17].
Elective Neural Networks - PER : Facts of the case.
10. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 10/93
1.4. Experiments on insects: the honeybee (Apis Mellifera).
Amazing visual and cognitive abilities of the Apis Mellifera bee have been studied by Martin
Giurfa's team in Toulouse. Experiments using a Y-maze where the bee is in free flight have shown,
among other things, that the bee is capable of conceptual learning: "identity/difference" [18],
"top/bottom" [19], to master simultaneously two concepts: "top/bottom" and "different" [10], to
generalize, categorize and extract rules [20]. Mushroom bodies seem to play an essential role in
high-level learning [11] [21].
The extension of the proboscis in contact with a sweet solution is an innate reflex (PER,
Proboscis Extension Response) that can be studied on immobilized bees. The first work on the
conditioning of the PER with colours was carried out by Masutaro Kuwabara in 1957 [22]. A
Pavlovian protocol with odours was published by Kimihisa Takeda in 1961 [23], and improved by
Morton Edward Bitterman in 1983 [24].
Olfactory conditioning of the PER has become a valuable tool for the study of learning and
memory, especially since it is possible to trace the stimuli (CS, US, Conditional, Unconditional) in
vivo inside the bee's brain through various electrophysiological and optophysiological recordings
[25] [26].
A new conditioning paradigm was introduced in 2004: mechanosensory conditioning of the PER.
The immobilized bee learns to associate a mechanosensory stimulation on an antenna (light
contact of a toothpick) with the reward of a sweet solution applied on the proboscis [27].
Elective Neural Networks - PER : Facts of the case.
14. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 14/93
1.8. Architecture models of the bee’s brain.
The direct recordings made in vivo inside the bee’s brain, accompanied by behavioural
observations in all these conditioning experiments, make it possible to build models of architecture
and functioning of the insect’s brain.
No experiment, on any brain in any species, allows such a direct and close approach to the
mechanisms of memory and learning. The conditioning phenomena observed in a multitude of
species, including humans, give reason to hope that the progress made in bees will advance the
understanding of neural mechanisms throughout the animal kingdom. It is therefore essential to
elucidate them.
Several authors have constructed proposals for insect brain architectures that could address the
problem of Pavlovian conditioning. : Menzel & Giurfa (2001) [34], Heisenberg (2003) [35], Cassenaer
& Laurent (2007) [36], Galizia (2014) [37], Mizunami et al. (2015) [38]. The structures involved are:
antennal lobes (AL) with projection neurons (PN), and mushroom bodies (MB) with Kenyon cells
(KC). The following pages show three examples.
All these architectures have matrix structures between stimuli (conditional and unconditional)
and muscle commands. The synaptic mechanisms acting at the intersection of each row and each
column are the most mysterious aspect of the problem and can only be analysed as part of an
organised system.
It is here that the conditional learning mode of the Elective Neural Networks model can make a
useful contribution, in particular by also explaining the shapes of the learning curves.
Elective Neural Networks - PER : Facts of the case.
16. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 16/93
1.8.2. Galizia, 2014 : "Olfactory coding in the insect brain: data and conjectures "
[37] (1).
Elective Neural Networks - PER : Facts of the case.
Fig. 5. Galizia, 2014 [37]
(License CC BY-NC-ND 4.0.)
17. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 17/93
1.8.2. Galizia, 2014 : "Olfactory coding in the insect brain: data and conjectures "
[37] (2).
Elective Neural Networks - PER : Facts of the case.
Fig. 6. Galizia, 2014 [37]
(License CC BY-NC-ND 4.0)
18. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 18/93
1.8.3. Mizunami et al. 2015 : "Toward elucidating diversity of neural mechanisms
underlying insect learning " [38].
Elective Neural Networks - PER : Facts of the case.
Fig. 7. Mizunami et al. 2015 [38].
(License CC BY 4.0)
19. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 19/93
1.9. 2011 : the analysis of Evren Pamir (1).
In 2004, Charles Gallistel and his team conducted conditioning experiments on pigeons to
assess the extent to which group average performance measures represented the behavioural
performance of individuals. They found that the classic, gradual, negative acceleration learning
curve was an artifact produced by averaging. Individual behaviour was in fact characterised by an
abrupt change in the level of response [39].
To investigate this question further, Evren Pamir's team published in 2011 the results of
individual olfactory conditioning experiments on 1640 bees in 17 groups, with several conditioning
protocols. The individual behaviour of each bee was recorded [40].
E. Pamir determined that the model that best explains individual behavior is a two-state hidden
Markov model. The team’s conclusions can be summarized as follows:
• There is no progressive variation of the conditioning (of the "strength" of the synapses?),
• On the contrary, at a certain point, an irreversible change occurs in the bee's brain from the
"unconditioned" to the "conditioned" state.
Elective Neural Networks - PER : Facts of the case.
20. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 20/93
Three model-hypotheses are proposed
to account for the learning curve:
A : Simple Learning-Curve (LCM1),
B : Extended Learning Curve (LCM2),
C : Hidden Markov Model (HMM).
1.9. 2011 : the analysis of Evren Pamir (2).
Elective Neural Networks - PER : Facts of the case.
Fig. 8. Pamir et al., 2011 [40]
(License CC BY-NC 4.0)
21. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 21/93
The detailed analysis of the responses of
each bee gives the advantage to the
Hidden Markov Model (HMM).
It therefore appears that there is, at some
point, an irreversible change of internal
state in the bee’s brain: a correct
response results in a very high
probability of correct responses for all
subsequent tests.
1.9. 2011 : the analysis of Evren Pamir (3).
Elective Neural Networks - PER : Facts of the case.
Fig. 8. Pamir et al., 2011 [40]
(License CC BY-NC 4.0)
22. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 22/93
1.10. The problem of the omission procedure.
An experiment with strange results, the "Omission Procedure" is described in a 2004 article by
Martin Giurfa and Dagmar Malun [27] : conditioning is acquired even if the reward is not given.
Giurfa et Malun conclude that "the association established during conditioning is independent of
the response (extension of the proboscis) of the bees and is therefore classical and not operant".
A version of this experiment was reported in a 1983 article by Bitterman, Menzel, Fietz and Shäfer
[24].
Elective Neural Networks - PER : Facts of the case.
Fig. 10. Giurfa, Malun, 2004 [27]
(License CC BY-NC 4.0)
23. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 23/93
1.11. Conclusions and assumptions.
The results of the above-mentioned bee experiments allow the following assumptions to be
made in order to draw up an architectural proposal:
1) The VUMMx1 signal used to "validate" the proboscis extension can be interpreted as a reward
signal, a positive variation of the bee's state of "well-being". Its uniqueness implies that it must
evaluate the last action performed, i.e. the last effector activated in the perceptual configuration of
the moment. If the last action performed does not trigger the reward signal, it is interrupted and
leaves no trace (conditional learning).
2) Successful completion of conditioning requires that the conditional stimulus (CS) be presented
shortly before the unconditional stimulus (US). The occurrence of the CS does not bring any
disturbance in the response to the US.
3) In the case of the PER, the experience of the omission procedure seems to show that something
irreversible happens from the first presentation "US - CS".
4) Evren Pamir's analysis clearly points towards a discrete (and not progressive) modification of
the synapses in this learning process, with a stochastic component.
Elective Neural Networks - PER : Facts of the case.
24. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 24/93
Elective Neural Networks - PER : Proposals of the ENN model.
1. PER Conditioning : Facts of the case.
2. PER Conditioning : Proposals of the ENN model.
Two types of effectors.
Type II Learning.
Type II Hyperdomain.
Effector arbitration.
Effector control.
PER conditioning.
3. The Orthogonal Brain.
4. Conclusion : Process of brain development.
5. Bibliography.
25. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 25/93
"Certes, je ne crois pas que les abeilles se livrent à ces calculs
compliqués, mais je ne crois pas davantage que le hasard ou la seule
force des choses produise ces résultats étonnants".
"Certainly, I do not believe that bees engage in these complicated
calculations, but I do not believe any more that chance or the mere force of
things produces these astonishing results".
Maurice Maeterlinck, La vie des abeilles, 1901.
(Quoted by par Constance Boitard, 2015, p. xiii) [41].
Elective Neural Networks - PER : Proposals of the ENN model.
26. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 26/93
2.1. Conditioning of the PER: proposals of the ENN model.
The first part of this article ("the Boolean brain") dealt only with perception: taking into account
the signals provided by sensors, processing, memorization and recognition of configurations
already seen.
Every living organism possesses, at a very variable level, sensors and effectors ("muscular or
glandular organ which enters into activity in response to a given stimulus").
The first function of the brain is the construction of an internal representation of the
environment, in order to identify, evaluate situations and predict their evolution. The second
function of the brain is to choose and carry out the most favourable actions for the organism. This
process is carried out in two stages: selection of the action and control of the effectors.
This paragraph will therefore deal with the issues of learning conditioned by a reward signal,
effectors and their control, to finally address the mechanisms of conditioning according to the ENN
model.
Three assumptions are added to the ENN model:
1) Two types of effectors: simple, compound.
2) Type II Learning (T2L): Conditional Learning.
3) Type II hyperdomain: restriction of T2L to only one DME at a time.
Elective Neural Networks - PER : Proposals of the ENN model.
27. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 27/93
2.2. Assumption 1 : two types of effectors.
The ENN model assumes the existence of two types of effectors :
The simple effector: one command one action.
The compound effector: a group of antagonistic, mutually exclusive effectors.
Each effector, simple or compound, is managed by a control module.
Number of effectors in bees (TBC) :
8 glands,
200 muscles, some of which are in single control, others in antagonistic control (mutual
exclusion).
Elective Neural Networks - Two types of effectors.
28. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 28/93
2.3. Assumption 2 : Conditional Learning ("of Type II").
Effector control implements a special learning mechanism: Conditional Learning, Type II (T2L).
First phase: same as type I learning.
A new configuration at the input of a DME causes the emergence of the G neuron, which
triggers a process of temporary deactivation of active synapses leading to the inactivation of
the active synapses of a single neuron.
Second phase: conditional completion.
In T1L, the process is irreversible.
In T2L, the process of synapse elimination is only initiated if a particular additional signal
occurs at the time of the extinction of the G. In the absence of this signal, learning is
interrupted and the synapses return to their initial state.
This signal is the "RS" reward signal.
Elective Neural Networks - Type II Learning.
29. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 29/93
Σ E
Σ N
Gf
1. 2. 3. 4. 5. 6.
T1
T2
(Reversible process)
RS
T4
T3
The RS reward signal must occur at the latest T3 after Gf extinction and persist until the end
of the reversibility period of the process: T2 after Gf extinction.
2.4. Conditional Learning Sequence.
Elective Neural Networks - Type II Learning.
Fig. 11.
30. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 30/93
2.5. Why is Type II (Conditional) learning necessary? (1)
Action selection is the crucial problem facing any animal. At any time, its sensors give it
information about its internal state and its environment and it must use this data to choose the most
appropriate action.
The way of making this choice (within a space of actions) has been the subject of many works in
the field of neuroscience and in robotics, where the goal is to optimize the behavior of autonomous
agents (Girard, 2010) [42].
Reinforcement learning (Sutton, Barto, 1998) [43], derived from Machine Learning, "is the
problem faced by an agent who must learn his behaviour through trial-and-error interactions with a
dynamic environment" (Girard, 2003, p. 33) [44].
The agent must have a system of evaluation that constantly informs him about the "quality of his
behaviour". The evaluation is delivered in the form of a signal, a neural activation that must be used
by choosing the behaviour or action from among all the possibilities offered by the effectors.
Elective Neural Networks - Type II Learning.
31. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 31/93
2.5. Why is Type II (Conditional) learning necessary? (2)
Type I learning in the ENN model is unsupervised learning that simply memorizes coincidences
and redundancies in perceptions.
Type II learning, from perceptual configurations thus organized, records the result of each action
selected in each perceptual configuration (which represents both the external environment and the
internal state of the animal).
In the ENN model for the bee, the action space is the set of actionable effectors. The time horizon
of the evaluation system is less than 3 seconds. This is therefore the simplest and most naive
version of reinforcement learning. Once the action has been decided and the effector activated, the
result of the action is immediately evaluated:
1) The action is favorable: it is stored (T2L) and will be repeated in the same context,
2) The action is not favorable: it is forgotten (T2L interrupted).
Elective Neural Networks - Type II Learning.
32. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 32/93
2.6. Assumption 3 : Type II Hyperdomain.
The uniqueness of the evaluation device (reward circuit), which may or may not validate the last
action performed, means that Type II learning must only be applied to the last activated effector.
There is therefore a mechanism for selecting the effector to be taught.
Three proposals for this selection scheme:
1. One of the preferred paths of evolution being to adapt and reuse existing devices, it is
logical, by analogy to the theory of neuronal recycling of S. Dehaene [45] [46], to consider a
type II hyperdomain simply derived from type I, in which only learning will be restricted to
one domain at a time, without inhibiting other actions.
2. Conditioning experiments tend to think that the activation of an effector acts on the device
by moving the selection on this effector (or this group of effectors) so as to allow a new
perceptive configuration to activate it (by a type II learning).
3. Whatever the type of effector, the control of the effectors is carried out by the activation of a
neuron belonging to a Domain of Mutual Exclusion. Selecting an effector therefore
corresponds to selecting a DME.
Representation: each DME working with T2L will have a specific symbol :
= Symbol of the RS signal enabling T2L.RS
Elective Neural Networks - Type II Hyperdomain.
33. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 33/93
In a Type II hyperdomain, all domains remain active, but only one can benefit from
learning at a time.
N
G
N
N
N
N
H
N1G1G0
N
G
N
N
N
N
H
N1G1G0
Type I Hyperdomain ≠ Type II Hyperdomain
2.7. Type I and II Hyperdomains, differences :
Elective Neural Networks - Type II Hyperdomain.
Fig. 12.
34. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 34/93
MS1-EF1 MS2-EF1 MSn-EF1
vEF1
MS1-EF2 MS2-EF2 MSn-EF2
vEF2
vEF
MS1 MS2 MSn MSnMS2MS1
Level 0
Level 1
Level 2
2.8.a. Arbitration (selection) of effectors:
For type II hyperdomains, only level 0 is different from
type I: T2L is restricted to only one effector at a time
(see next page).
Elective Neural Networks - Effector arbitration.
(Validation of
Effector 1)
Fig. 13.
36. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 36/93
Example n°1 :
A simple effector,
Only one sensory modality.
Each new stimulus (perceptual configuration) triggers an APT2,
activates a new N neuron and causes activation of the effector.
The effect on the organism of this activation is immediately
evaluated by the reward system:
If the RS signal is not generated, the T2L is interrupted in its
reversible period; the N neuron is deactivated and returns to its
initial state.
If the RS signal is generated, the T2L is terminated correctly. The
N neuron will now be activated at each new reception of the same
perceptual configuration, as well as the same effector.
The number of perceptual configurations that can activate this
effector is limited by the number of N neurons in this DME.
N
Gf
N
N
N
N
EF1
RS
2.9.a. Effector Control.
Elective Neural Networks - Effector control.
(Symbol of T2L)
Fig. 15.
37. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 37/93
Each new stimulus triggers a T2L, activates a new N neuron and
results in the activation of one of the two effectors, with a
probability that depends on their respective connectivities.
The effect on the organism of this activation is immediately
evaluated by the reward system:
If the RS signal is not generated, the T2L is interrupted in its
reversible period; the N neuron is deactivated and returns to its
initial state.
If the RS signal is generated, the T2L is terminated correctly. The
N neuron will now be activated at each new reception of the same
perceptual configuration, as well as the same effector.
N
Gf
N
N
N
N
EF1a EF1b
RS
Example n°2 :
A compound effector (two antagonistic effectors),
Only one sensory modality.
2.9.b. Effector control.
Elective Neural Networks - Effector control.
(Symbol of T2L)
Fig. 16.
38. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 38/93
Situation :
One of the two effectors presents a habitual and systematic
response to a particular stimulus (US) presented in one of the two
sensory modalities (MS1).
For the conditioning test, another stimulus (CS) is presented in
the second sensory modality (MS2), followed closely by the US in
the first modality :
EF1a EF1b
?
US CS
MS1-EF1
EF1
MS2-EF1
Example n°3 :
A compound effector (two antagonistic effectors),
Two sensory modalities.
2.9.c. Effector control.
Elective Neural Networks - Effector control.
[28]
Fig. 17.
Fig. 18. (Excerpt from Fig. 1.)
39. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 39/93
2.10. Architecture of effector control: constraints and assumptions.
What follows is a reverse engineering work. The aim is to design a scheme assembling domains
of mutual exclusion according to the ENN model while managing to respect the constraints listed in
the conclusion of § 1.
Regarding the effector control, here are the assumptions related to the ENN model:
Each sensory modality (gustatory, olfactory, haptic, visual, etc...) has a hyperdomain in which
each particular sensation is decoded and activates a neuron.
Each effector (muscle, gland) can be activated by a large amount of signals from the
hyperdomains of sensory modalities. This corresponds to a large logic gate "OR".
The activation of an effector is the result of the propagation of a signal that passes through a
path that has been previously "enabled". All the possible paths pre-exist, only some of them are
"validated" and effective. It is this validation process that will be described using the diagrams on
the following pages.
Any alternative hypothesis would lead to the acceptance that synapses can be created at a specific
place, at a specific time, by a still unknown (and improbable) process.
Elective Neural Networks - Effector control.
40. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 40/93
2.11. Architecture of effector control: choices and justifications.
Why are two intermediate levels necessary?
The uniqueness of the reward signal necessitates another intermediate stage to limit the
validation of the signal path to a single sensory modality to a single effector. This stage consists
of a Type II hyperdomain, where only learning is limited to a DME. The reward signal then enables
learning in the only active DME (that of the sensory modality where the event takes place) and for
the only effector validated by effector arbitration (positioned by the previous activation of the IS).
The fact that several sensory modalities are connected to each effector module makes it
possible to carry out contradictory actions, especially on compound effectors. To avoid these
contradictions and manage access conflicts, the proposed schema includes an intermediate
stage where the DMEs of all the sensory modalities are grouped in a type I hyperdomain, which
limits access to the effector module to only one modality at a time.
By its purely logic (Boolean) character, this architecture is compatible with Evren Pamir's
conclusions. Uncertainty in the resolution of metastability in a DME brings the stochastic aspect he
observed.
In the bee, the proboscis effector control module is a DME in which each neuron is connected to
the muscles (or groups of muscles) that ensure the protraction (extension) or retraction of the
proboscis. The maxillae of the bee, unlike other insects, do not have retraction muscles; the
proboscis depends for its retraction on the retraction muscles of the labium (Snodgrass, 1956) [47].
Elective Neural Networks - Effector control.
41. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 41/93
EF1a EF1b
US CS
EF1
MS1' MS2'
MS1-EF1 MS2-EF1
EF1a : Proboscis extension.
EF1b : Proboscis retraction (tbc).
(Extension) (Retraction)
Type II Hyperdomain : only the effector selected
by the arbitration can undertake a learning
process and "add" a new perceptual pattern.
Type I Hyperdomain : Only one perceptual pattern
can activate the effector at a time.
Two sensory modalities : MS1, MS2.
(The other modalities MS3, MS4,... are not shown).
Elective Neural Networks - Effector control.
Fig. 19.
42. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 42/93
EF1a EF1b
US CS
EF1
MS1' MS2'
Type II Hyperdomain : validation of the effector selected to receive
a learning. The domains of this stage are subject, for all new
learning, to validation by the EF1 effector.MS1-EF1 MS2-EF1
Type I Hyperdomain : arbitration of access to the effector module.
This intermediate stage is necessary to avoid conflicts of access to
the effector module from stimuli coming from different sensory
modalities (arbitration is not represented).
The EF1 Domain is the Control Domain of the EF1 effector.
The EF1 effector is an effector composed of two antagonistic
effectors EF1a and EF1b.
The US stimulus comes from the MS1 sensory modality, the CS
stimulus comes from MS2 (the other modalities MS3, MS4,...
are not represented).
Elective Neural Networks - Effector control.
(Extension)
Fig. 20.
43. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 43/93
2.12. PER conditioning sequence.
The following pages present the PER conditioning sequence according to the ENN model :
1 : Presentation of Conditional Stimulus: neutral odour.
2a, 2b, 2c, 2d : Presentation of Unconditional Stimulus: sucrose.
3a à 3d : Following presentations: odour alone.
3e, 3f : Odour alone, proboscis extended, successful conditioning.
3g, 3h, Odour alone, non-extended proboscis, unsuccessful conditioning.
Elective Neural Networks - PER Conditioning.
44. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 44/93
EF1a EF1b
US CS
EF1
MS1' MS2'
MS1-EF1 MS2-EF1
(Sucrose)
1. Presentation of Conditional Stimulus: neutral odour.
The CS stimulus, connected to the input of all MS2-EFn domains, is
presented. Since none of the effectors are enabled, nothing happens.
(Extension)
(Odour)
Elective Neural Networks - PER Conditioning.
EF1a : Proboscis extension.
EF1b : Proboscis retraction (tbc).
(Sucrose) (Odour)
US CS
Fig. 21.
45. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 45/93
EF1a EF1b
EF1
MS1' MS2'
MS1-EF1 MS2-EF1
The US stimulus, connected to the input of all MS1-EFn domains, is
presented. This stimulus triggers the extension of the proboscis
through the following steps:
Activation of the relevant N neuron in the MS1-EF1 domain.
2a. Presentation of unconditional stimulus: sucrose.
Elective Neural Networks - PER Conditioning.
(Sucrose) (Odour)
US CS
(Extension)
Fig. 22.
46. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 46/93
EF1a EF1b
EF1
MS1' MS2'
MS1-EF1 MS2-EF1
2b. Presentation of unconditional stimulus: sucrose.
The US stimulus, connected to the input of all MS1-EFn domains, is
presented. This stimulus triggers the extension of the proboscis
through the following steps:
Activation of the relevant N neuron in the MS1-EF1 domain.
This activation enables the EF1 effector in the effector
arbitration hyperdomain, which triggers a Type II learning in the
MS2-EF1 domain.
Elective Neural Networks - PER Conditioning.
(Sucrose) (Odour)
US CS
(Extension)
Fig. 23.
47. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 47/93
EF1a EF1b
EF1
MS1' MS2'
MS1-EF1 MS2-EF1
2c. Presentation of unconditional stimulus: sucrose.
Elective Neural Networks - PER Conditioning.
(Sucrose) (Odour)
US CS The US stimulus, connected to the input of all MS1-EFn domains, is
presented. This stimulus triggers the extension of the proboscis
through the following steps:
Activation of the relevant N neuron in the MS1-EF1 domain.
This activation enables the EF1 effector in the effector
arbitration hyperdomain, which triggers a Type II learning in the
MS2-EF1 domain.
Activation of the relevant neuron in the intermediate domain MS1'.
This activation has the collateral effect of inhibiting all intermediate
domains of the same Type I Hyperdomain, including MS2'.
(Extension)
Fig. 24.
48. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 48/93
EF1a EF1b
EF1
MS1' MS2'
MS1-EF1 MS2-EF1
2d. Presentation of unconditional stimulus: sucrose.
Elective Neural Networks - PER Conditioning.
(Sucrose) (Odour)
US CS The US stimulus, connected to the input of all MS1-EFn domains, is
presented. This stimulus triggers the extension of the proboscis
through the following steps:
Activation of the relevant N neuron in the MS1-EF1 domain.
This activation enables the EF1 effector in the effector
arbitration hyperdomain, which triggers a Type II learning in the
MS2-EF1 domain.
Activation of the relevant neuron in the intermediate domain MS1'.
This activation has the collateral effect of inhibiting all intermediate
domains of the same Type I Hyperdomain, including MS2'.
Activation of the neuron N concerned in the control domain of the
EF1 effector, which activates the EF1a effector.
The bee expands its proboscis, which reaches the drop of
sucrose solution. This favourable sensation activates the neuron
VUMmx1 and validates the current T2L: that of MS2-EF1.
From that moment on, an N neuron in the MS2-EF1 domain will be
able to recognize the neutral odor that has just been presented (as a
favorable stimulus).(Extension)
Fig. 25.
49. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 49/93
EF1a EF1b
EF1
MS1' MS2'
MS1-EF1 MS2-EF1
3a. Following presentations : odour alone.
Elective Neural Networks - PER Conditioning.
(Sucrose) (Odour)
US CS
(Extension)
Fig. 26.
50. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 50/93
EF1a EF1b
EF1
MS1' MS2'
MS1-EF1 MS2-EF1
An N neuron in the MS2-EF1 domain recognizes the odour,
which is presented again.
3b. Following presentations : odour alone.
Elective Neural Networks - PER Conditioning.
(Sucrose) (Odour)
US CS
(Extension)
Fig. 27.
51. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 51/93
EF1a EF1b
EF1
MS1' MS2'
MS1-EF1 MS2-EF1
This neuron triggers a T2L in the intermediate domain MS2':
a neuron N emerges randomly in this domain (*).
3c. Following presentations : odour alone.
(*) This explains the variability of neuronal connectivity between two
individuals for the same function.
Elective Neural Networks - PER Conditioning.
(Sucrose) (Odour)
US CS
An N neuron in the MS2-EF1 domain recognizes the odour,
which is presented again.
(Extension)
Fig. 28.
52. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 52/93
EF1a EF1b
EF1
MS1' MS2'
MS1-EF1 MS2-EF1
This neuron N of MS2' triggers an T2L in the control domain of the
EF1 effector: a neuron N randomly emerges in the EF1 domain,
causing the activation of EF1a or EF1b.
3d. Following presentations : odour alone.
Elective Neural Networks - PER Conditioning.
(Sucrose) (Odour)
US CS
An N neuron in the MS2-EF1 domain recognizes the odour,
which is presented again.
This neuron triggers a T2L in the intermediate domain MS2': a
neuron N emerges randomly in this domain (*).
(Extension)
Fig. 29.
53. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 53/93
EF1a EF1b
EF1
MS1' MS2'
MS1-EF1 MS2-EF1
First case: EF1a is activated, the proboscis is extended.
3e. Following presentations : odour alone.
Elective Neural Networks - PER Conditioning.
(Sucrose) (Odour)
US CS
An N neuron in the MS2-EF1 domain recognizes the odour,
which is presented again.
This neuron triggers a T2L in the intermediate domain MS2': a
neuron N emerges randomly in this domain (*).
This neuron N of MS2' triggers an T2L in the control domain of the
EF1 effector: a neuron N randomly emerges in the EF1 domain,
causing the activation of EF1a or EF1b.
(Extension)
Fig. 30.
54. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 54/93
EF1a EF1b
EF1
MS1' MS2'
MS1-EF1 MS2-EF1
EF1a is activated, the proboscis is extended:
The RS reward signal is distributed.
Domains EF1 and MS2' finish their T2L correctly.
A complete link is established between the odour (CS) and EF1a
(proboscis extension): the conditioning of the PER is acquired.
3f. Following presentations : odour alone.
Elective Neural Networks - PER Conditioning.
(Sucrose) (Odour)
US CS
(Extension)
Fig. 31.
55. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 55/93
EF1a EF1b
EF1
MS1' MS2'
MS1-EF1 MS2-EF1
Second case: EF1b is activated, the proboscis is not extended.
3g. Following presentations : odour alone.
Elective Neural Networks - PER Conditioning.
(Sucrose) (Odour)
US CS
(Extension)
Fig. 32.
56. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 56/93
EF1a EF1b
EF1
MS1' MS2'
MS1-EF1 MS2-EF1
EF1b is activated, the proboscis is not extended:
The RS reward signal is not distributed,
Domains EF1 and MS2' abort their T2L.
No new learning is done, but the MS2-EF1 domain is still able to
recognize the odour.
If the odour presentation is renewed, a new T2L will be carried out in
MS2' and EF1, each time with a probability of activating EF1a and
obtaining the proboscis extension.
3h. Following presentations : odour alone.
Elective Neural Networks - PER Conditioning.
(Sucrose) (Odour)
US CS
(Extension)
Fig. 33.
57. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 57/93
At each presentation of the Conditional Stimulus, any N
neuron has the same probability to emerge. The probability
of proboscis extension depends on the connectivity
between the EF1 domain and the EF1a and EF1b effectors.
In the circuit presented here, the neurons from A1 to A5
trigger the extension, the others do not.
Let's look at what happens with an effector control circuit
where the probability of obtaining the reward (proboscis
extension) is :
k = #A / #B
(ratio of the number of neurons in list A to the number of
neurons in list B).
Proboscis
Extension
A5
A4
A3
A2
A1
B5
B4
B3
B2
B1
EF1a EF1b
EF1 Domain:
2.13. Effects of EF1 effector control domain connectivity :
Elective Neural Networks - PER Conditioning.
Fig. 34.
58. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 58/93
In each test, the conditional stimulus (CS) is presented first, followed by the unconditional
stimulus (US). We reason according to the bees that have not yet succeeded, to have an equation
of the form "1-X":
At each test, the probability of success is P1 = k, the probability of failure is P = 1 - k :
Every bee that succeeds once never fails again :
• Test 1 : (1 – k) fail,
• Test 2 : (1 - k)² fail again,
• Test 3 : (1 - k)3 fail again, etc…
• …
• Test n : (1 - k)n fail again…
The success curve of a bee population therefore follows the equation : P = 1 - (1 - k)n
2.14. Cumulative probability equation of the PER :
Elective Neural Networks - PER Conditioning.
59. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 59/93
An example, with k = 0,5 :
P = 1 - (1- k)n (k = 0,5)
(The curve obtained must be adjusted for the fact that 10 to 20% of the bees are non-learner bees).
Proboscis
Extension
A5
A4
A3
A2
A1
B5
B4
B3
B2
B1
EF1a EF1b
0
10
20
30
40
50
60
70
80
90
100
0 1 2 3 4 5 6
n
P (%)
% PER
Elective Neural Networks - PER Conditioning.
Fig. 35.
Fig. 36.
60. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 60/93
P = 1 - (1- k)n (k = 0,1)
B6
B7
B8
B9
A1
B5
B4
B3
B2
B1
EF1a EF1b
Sting
Extension ?
Another example, with k = 0,1 :
0
10
20
30
40
50
60
70
80
90
100
0 1 2 3 4 5 6
n
P (%)
% SER
Elective Neural Networks - PER Conditioning.
Fig. 37.
Fig. 38.
62. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 62/93
2.15. Solving the problem of the omission procedure:
incremental enrichment of the reward circuit.
The answer of the ENN model to the problem raised by the omission procedure is the incremental
enrichment of the reward circuit: the notion of immediate benefit following the activation of an
effector is not fixed (limited to the innate), but evolves with experience :
The direct association of the SC with the reward as soon as the proboscis comes into contact
with the sucrose solution explains the result in the case of the omission procedure for the PER.
The existence of aversive learning shows that there is also a mechanism for evaluating and
remembering negative events [53], which is not discussed here.
MS1
MS2
MS3
GPP
RS
Innate (Nature)
Acquired
(Nurture)
The reward circuit is gradually
enriched from an innate core.
GPP = Global Perceptual Pattern, elaborated in Mushroom Bodies.
Elective Neural Networks - PER Conditioning.
Fig. 40.
63. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 63/93
2.16. Non-standard analysis of Pavlov's experiment :
This proposal for an evolution of the reward circuit makes it possible to "revisit" Pavlov's dog
experience (before rigorous protocols were established):
The sight of the meat after the bell is not the first unconditional stimulus :
1. The first unconditional stimulus is the contact of the meat on the dog’s tongue.
2. First conditioning: the sight of the meat before contact with the tongue.
The single view of the meat then activates the reward circuit and becomes an US.
3. Second conditioning: the bell before the sight of the meat makes the dog salivate.
4. Third conditioning: transporting the dog to the experiment room, before the bell that precedes
the sight of the meat... makes the dog salivate.
5. Fourth conditioning: the arrival of the laboratory employee in a white blouse which precedes
the transport to the experiment room, which precedes the sight of the meat, etc.
Elective Neural Networks - PER Conditioning.
64. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 64/93
2.17. "Predictions" of the ENN model for the PER :
1. In the absence of learning, the probability that an unknown stimulus triggers a response
is non-zero (it is equal to the inverse of the number of activable effectors).
2. The "failed" tests of the conditioning actually correspond to the control of another
effector in the same module (for the PER: proboscis retraction).
Elective Neural Networks - PER Conditioning.
65. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 65/93
1. PER Conditioning : Facts of the case.
2. PER Conditioning : Proposals of the ENN model.
Two types of effectors.
Type II Learning.
Type II Hyperdomain.
Effector arbitration.
Effector control.
PER conditioning.
3. The Orthogonal Brain.
4. Conclusion : Process of brain development.
5. Bibliography.
Elective Neural Networks - The Orthogonal Brain.
66. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 66/93
3.1. Logical consequence of effector control according to the ENN model.
The ENN model is compatible with the idea of gradual implementation of effector activation under
the influence of the reward circuit. This circuit evolves in parallel with each experience of the animal,
starting from a core of innate actions.
Can we go further in trying to restrict this core?
The rationale presented in the following pages attempts to link two seemingly unrelated facts:
1) The genetic code doesn't contain enough information to encode the synapses.
2) In a conditioning, the CS can belong to all sensory modalities, the US to all types of effectors.
Elective Neural Networks - The Orthogonal Brain.
67. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 67/93
3.2. Too little information in the genetic code to encode the synapses:
Jean-Pierre Changeux
Collège de France - Communications cellulaires
Inaugural lecture delivered on January 16, 1976:
" …Si dans ses moindres détails l'organisation fonctionnelle du système nerveux est soumise à un
déterminisme génétique aussi strict, existe-t-il suffisamment de gènes dans le stock chromosomique
pour en rendre compte ? Chacune des 100 000 milliards de synapses du cerveau humain est-elle
codée par un gène différent ? Evidemment, la réponse est non ".
" …If in its smallest details the functional organization of the nervous system is subjected to such
strict genetic determinism, are there enough genes in the chromosomal stock to account for it? Is
each of the 100,000 billion synapses in the human brain encoded by a different gene? Of course, the
answer is no ".
"…Comment est-il possible d'engendrer la formidable complexité du système nerveux à partir d'un
nombre aussi restreint de gènes ? La réponse est à chercher dans la mécanique du développement,
l'Entwicklungs Mechanik de Wilhelm Roux qui, à partir de l'œuf, construit l'adulte ".
" …How is it possible to generate the tremendous complexity of the nervous system from such a
small number of genes? The answer is to be researched in the mechanics of development, the
Entwicklungs Mechanik of Wilhelm Roux, which from the egg builds the adult ".
(Changeux, 1976, § 61) [49]
Elective Neural Networks - The Orthogonal Brain.
68. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 68/93
3.3. CS : all sensory modalities, US : all types of effectors:
Jean Delacour, "Apprentissage et mémoire : une approche neurobiologique" (1987) :
("Learning and memory: a neurobiological approach")
" 6. Différentes formes de conditionnement classique.
Une grande variété de stimulus et de réponses se prêtent à l'établissement d'un conditionnement. Le
SC peut appartenir à toutes les modalités sensorielles.
… Les réponses conditionnables peuvent être somatomotrices mais aussi viscérales (par exemple
modifications gastriques ou du péristaltisme intestinal, cf. Bykov, 1956), endocriniennes (Graham et
Desjardins, 1980) ou immunologiques (Freed, 1984, Ader et Cohen, 1985). Des phénomènes
neurobiologiques peuvent aussi être conditionnés: rythmes EEG et activité unitaire (cf. chap. IV et V),
concentration ou métabolisme de certains neuromédiateurs (Ennaceur et coll. 1985). "
" 6. Different forms of classical conditioning.
A wide variety of stimuli and responses are suitable for conditioning. The CS can belong to all
sensory modalities.
… Conditionable responses may be somatomotory but also visceral (e.g., gastric or intestinal
peristalsis changes, cf. Bykov 1956), endocrine (Graham and Desjardins 1980) or immunological
(Freed 1984, Ader and Cohen 1985). Neurobiological phenomena can also be conditioned: EEG
rhythms and unitary activity (cf. Chap. IV and V), concentration or metabolism of certain
neuromediators (Ennaceur et al. 1985). "
(Delacour, 1987, Ch.II, §II.2.A.6, p.22) [50]
Elective Neural Networks - The Orthogonal Brain.
69. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 69/93
3.4. Proposal of the ENN model: towards the orthogonal brain.
The scheme of the effector control and the conditioning mechanism make it possible to formulate
the following proposal:
The initial brain wiring potentially enables any perceptive pattern to activate any effector.
To describe this proposal in detail, can be found in the following pages:
1) A proposal for a general scheme of the effector system in bees,
2) A complementary hypothesis for the connection of the olfactory system to the effector system
(problem of 50000 Kenyon cells),
3) The formulation of the "Orthogonal Brain" hypothesis.
Elective Neural Networks - The Orthogonal Brain.
70. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 70/93
Arbitration of effectors
MS1
MS2
CPG
EF1 EF2a EF2b EF3 EFna EFnb
EF1 EF2 EF3 EFn
Effectors
Sensors
(external and
internal), and
Global
Perceptual
Pattern GPP
(from
Mushroom
Bodies)
Single Compound
3.5a. General diagram of the effector system in bees (ENN model).
Elective Neural Networks - The Orthogonal Brain.
Fig. 41.
71. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 71/93
MS1
MS2
CPG
Effectors EF2a EF2b
EF2
Compound
Type II Hyperdomain: only the effector
selected by the arbitration can undertake a
learning and "add" a new perceptive
configuration.
Type I Hyperdomain: only one perceptive
configuration can activate the effector at a
time.
Elective Neural Networks - The Orthogonal Brain.
Sensors
(external and
internal), and
Global
Perceptual
Pattern GPP
(from
Mushroom
Bodies)
3.5b. General diagram of the effector system in bees (ENN model).
Fig. 42.
72. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 72/93
3.6. General diagram of the effector system : " Connectivity" issue.
The coding of odours in the brain is a fascinating question that is the subject of wonderful
research in both insects [51] [36] [52] [53] and mammals [54]. In insects, the Kenyon cells, located in
the Mushroom Bodies, are very numerous, are each activated infrequently and respond specifically
to particular odours. It is frequently accepted that they correspond to the concept of "grandmother
cell" (one cell, one odour), although there is no general consensus on this point [55]. It is also in the
Mushroom Bodies that associations between different sensory modalities are made [21].
In the ENN model, odour coding is radically simple. However, a question remains: how to
"connect" sensory modalities signals to effector controls ?
In the bee's brain, each of the two Mushroom Bodies contain about 170,000 Kenyon cells
responsible for decoding odours:
According to the ENN model, these 170,000 cells could be organized into 3400 DME (*) each with
50 cells. These 3400 DMEs would be clustered into a single hyperdomain, so that only one odour
could be decoded at a time (even when it is a mixture of odours, which can give rise to a particular
olfactory perception).
The diagram on the next page shows how the connection to the effector control could be made.
(*) Domain of Mutual Exclusion.
Elective Neural Networks - The Orthogonal Brain.
73. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 73/93
3.7. ENN hypothesis of connection of the olfactory system to the effectors.
EF2a EF2b
EF2
Antennal Lobe Mushroom Bodies
170000 Kenyon cells
(3400 DME)
Projection
Neurons
(800)
Type I Hyperdomain
Type II Hyperdomain
Elective Neural Networks - The Orthogonal Brain.
Fig. 43.
74. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 74/93
3.8. The Orthogonal Brain hypothesis.
The instruction set of a computer is said to be "orthogonal" when all instructions can
be applied to all types of data.
The architecture proposed by the ENN model allows each perceptive configuration,
whatever the sensory modality, to activate any effector.
By analogy with the definition of computer science, we can speak here of orthogonal
architecture.
Elective Neural Networks - The Orthogonal Brain.
75. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 75/93
3.9. The Orthogonal Brain hypothesis: consequences (1).
What can this diagram explain (not only for the bee)?
1. The great diversity of stimuli and conditionable responses: each sensory modality has access to
each effector (the idea that the brain learns by creating individual synapses "on demand" is definitely
eliminated).
2. Responses to ambiguous stimuli: Mushroom bodies, which produce complex signals combining
information from all sensory modalities, also have access to effectors. The ability to generate
complex logic functions, including the "Exclusive Or", was briefly described in the first part of this
presentation ("The Boolean Brain").
3. The paradox of the non-linearity between the number of genes and that of neurons: J.P. Changeux
has pointed out this paradox of evolution: the ratio of the numbers of genes of the mouse and the
man does not reflect at all the gap of size and performance of their brains [7].
Despite their apparent complexity, all the patterns in the ENN model are very easy to specify: a few
types of neurons, a few types of interactions, maximum connectivity between certain types of
neurons. Brain functions (perception, action) are installed by epigenesis. This reduces the number of
"genetic code lines" required.
Elective Neural Networks - The Orthogonal Brain.
76. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 76/93
3.9. The Orthogonal Brain hypothesis: consequences (2).
What is the link between the two facts at the origin of this reflection? (§ 3.1)
The orthogonal architecture (each sensory pattern can activate each effector) allows to select by
epigenesis the most appropriate actions in each context. This mechanism is perfectly efficient and
requires little genetic code.
Elective Neural Networks - The Orthogonal Brain.
77. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 77/93
1. PER Conditioning : Facts of the case.
2. PER Conditioning : Proposals of the ENN model.
Two types of effectors.
Type II Learning.
Type II Hyperdomain.
Effector arbitration.
Effector control.
PER conditioning.
3. The Orthogonal Brain.
4. Conclusion : Process of brain development.
5. Bibliography.
Elective Neural Networks - Conclusion : Process of Brain Development.
78. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 78/93
4.1. Reflections: Process of brain development.
The ENN model offers a common mechanism for classical and operant conditioning.
Classical conditioning appears there as a variant of operant conditioning, in which the "space
of choice" of the action of the CS is restricted and oriented at the time of the presentation of the
US.
Let us imagine a "virgin" brain, in which no action would be programmed by the genes: the
effectors would be set up only by a series of operant conditionings. The theoretical obstacle is the
excessive width of the action choice space.
This choice space is naturally reduced during certain periods in the brain: the critical periods
in the development of the nervous system.
We can therefore imagine the implementation of all the effectors according to a programmed
sequence of critical periods: biochemical equilibrium (glands), vital functions (internal muscles,
glands), modes of locomotion (external muscles). Each stage requires the prior implementation of the
processing of sensory information.
Elective Neural Networks - Conclusion : Process of Brain Development.
79. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 79/93
4.2. Brain development and Critical Periods (1).
In cerebral development, a critical period (or sensitive period) is "a period during which a cortical
region changes morphologically and physiologically according to the information it receives" [56], "a
window, typically in early development, during which a system is open to structuring or restructuring
on the basis of input from the environment" [57].
The best known example of a critical period is the imprinting phenomenon discovered by Konrad
Lorenz with geese in the 1930s. Immediately after birth, the young bird learns to follow his mother,
who protects him, guides him, and gives him the best protection and the greatest chance of survival.
But if the mother is not present, the chick can become permanently attached to all kinds of objects
that meet certain conditions (sounds, movements), living beings or mechanical objects [58] [59]. The
phenomenon always occurs at the same time for each species of bird (hatching or leaving the nest)
and lasts only a few hours.
In the 1960s, Torsten Wiesel and David Hubel explored the visual cortex of cats using electrodes
to observe the activation of individual neurons. Their discoveries on the functional organization of
the visual cortex earned them the Nobel Prize in 1981 [60] [61]. In particular, they showed that the
primary visual cortex (V1) was organized in columns of ocular dominance (right eye, left eye). Layer
IV of V1 receives information from both eyes: the same number of cells are assigned to each eye. If
an eye is forced to remain closed (monocular depivation) for a particular period after birth, the
primary visual cortex cannot be organized fairly, and the closed eye is definitely under-represented in
layer IV, which causes a permanent functional blindness of this eye, which is nevertheless intact. For
the cat, this critical period is approximately between six days and three months [62].
Elective Neural Networks - Conclusion : Process of Brain Development.
80. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 80/93
4.2. Brain development and Critical Periods (2).
In mice, the critical period for the visual system opens at P20, 20th day after birth (P20, Postnatal
day) and closes at P40 [65]. In monkeys, the maturation of the visual system involves several critical
periods: basic spectral sensitivity of cones and rods (3-6 months), monocular spatial vision (25
months), binocular vision (>25 months) [63]. In humans, a deficit of visual stimulation in one of the
two eyes at the beginning of life can induce amblyopia: difference in visual acuity between the eyes,
which cannot be explained by an organic lesion. The critical period for the visual system in humans
is thought to end at about six or seven years of age. After that age, treatments for amblyopia are less
effective [64].
The development of the complete cortex is a succession of critical periods: first the primary
sensory areas (smell, hearing, vision), then those of locomotion and finally the cognitive areas [65].
The question of the existence of a critical period for language acquisition now seems to be replaced
by the search for the biological mechanisms of this acquisition [57].
Studies on critical periods aim to understand precisely its mechanisms. In particular, research is
looking for a way to reopen these critical periods in adults by identifying the "brakes" on plasticity, in
order to restore deficient functions, for example to cure amblyopia [66]. The setting up of binocular
vision in mice is the most widely used model [56].
Elective Neural Networks - Conclusion : Process of Brain Development.
81. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 81/93
4.2. Brain development and Critical Periods (3).
The teams of Alain Prochiantz, T. Hensch and A. Simeone demonstrated in 2012 that the critical
period for the mouse visual cortex (opening between P20 and P40) was controlled by the
internalization of the Otx2 protein by parvalbumin neurons (PV cells) [67] [68]. PV cells are large
inhibitory neurons called “basket cells”, present in the cortex and hippocampus, and whose
maturation plays an essential role in modulating the excitation/inhibition ratio. "With maturation,
these neurons surround themselves with fixed peri-neuronal networks (proteins + sugars) (PNN,
Peri-Neuronal Nets), and their axons with myelin sheaths, which greatly reduce their plasticity" [69].
For J. Werker et T. Hensch, "The cellular substrates of CP plasticity include four sets of players:
CP triggers, mediators, brakes, and reopeners. The PV cell serves as a pivotal plasticity switch, and
its excitatory-inhibitory balance determines the timing of the CP… A consolidated state (CP closure)
is finally maintained by molecular brakes, both functional (such as Lynx1, NgR1) and physical (such
as PNN, myelin)". Ways to restore plasticity can be imagined [57].
"Windows of plasticity, therefore, arise between the maturation of an optimal E-I balance triggering
the machinery of synaptic pruning and a later-emerging consolidating set of brake-like factors, which
persistently limit rewiring throughout adult life" [57]. We also find here the importance of synaptic
pruning in the setting up of cortical circuits.
Elective Neural Networks - Conclusion : Process of Brain Development.
82. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 82/93
4.3. Brain Development Process (1).
From all the previous hypotheses, it is now possible to formulate a process of development
of the bee's brain defining two parts :
1) The part of genes, defined by J.P. Changeux (JP Changeux, 2008),
2) The part of epigenesis, according to the ENN model.
Elective Neural Networks - Conclusion : Process of Brain Development.
83. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 83/93
4.3. Brain Development Process (2).
1) The part of genes: setting up the organization (JP Changeux, 2008) [64] :
" A genetic envelope determines the main features of the anatomical organization of the brain that are
species-specific and altered by gene mutation, but retained after suppression of any release of
neurotransmitters; these are :
1. Morphogenesis of the neural tube and brain,
2. Division, migration and differentiation of nerve (and glial) cells,
3. Setting up maximum connectivity,
4. The spontaneous activation of neurons in the network,
5. Regulation of synapse assembly and evolution through the circulatory activity. "
(adapted from Jean-Pierre Changeux, 2008) [70].
Elective Neural Networks - Conclusion : Process of Brain Development.
84. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 84/93
4.3. Brain Development Process (3).
2) The part of epigenesis: connectivity adjustments (according to the ENN model).
6. Perception : implementation of the logic equations representative of the perceptual patterns
detected by the sensors, thanks to the Type 1 Learning (T1L) of the ENN model, which proceeds only
by synapse elimination starting from maximum connectivity.
7. Action : Progressive implementation of actions, in a sequence of critical periods, with the
selection of the most suitable effector for each perceptual pattern, through a trial-and-error process
under the control of the reward circuit. The conditional Type II learning mechanism (T2L) of the ENN
model, which also proceeds by synapse elimination, performs and stores this selection. The reward
circuit evolves in parallel thanks to the same T2L mechanism.
This reasoning brings out a new perspective :
No individual synapse is genetically programmed.
Elective Neural Networks - Conclusion : Process of Brain Development.
85. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 85/93
4.4. Conclusion and reasonable outlook.
Although not explicitly stated (to our knowledge), the idea of a Critical Period Genetic Sequencer
(CPGS) seems to be present in some papers.
Thus, the first cry of the newborn might not be innate: before birth, the CPGS selects the
respiratory effectors (two antagonistic groups of respiratory muscles), then an operant conditioning,
guided by the blood oxygen level sensors, is carried out to set up the inspiration-expiration
sequence. In this way, the complete development of all the nervous systems could be completely
reconsidered.
This conception seems strange. It is natural, in humans, to reject strange ideas in order to keep
familiar ones. Here are two examples of familiar ideas:
1. We learn and form our memories by "creating new synapses".
2. Some behaviours are "encoded in the genes".
If we think about the necessary mechanisms behind these two ideas, we realize that they are not
supported (for the time being) by any precise and solid hypothesis. The ENN model suddenly loses
some of its strangeness… Will the hypotheses that constitute it be able to intrigue curious minds and
thus open a new path to understanding how the brain works?
_____________________________________
Elective Neural Networks - Conclusion : Process of Brain Development.
86. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 86/93
1. PER Conditioning : Facts of the case.
2. PER Conditioning : Proposals of the ENN model.
Two types of effectors.
Type II Learning.
Type II Hyperdomain.
Effector arbitration.
Effector control.
PER conditioning.
3. The Orthogonal Brain.
4. Conclusion : Process of brain development.
5. Bibliography.
Elective Neural Networks - Bibliography.
87. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 87/93
Elective Neural Networks - Bibliography.
Bibliography.
1. Changeux, J. P., Courreges, P. & Danchin, A. (1973). A theory of the epigenesis of neuronal networks by
selective stabilization of synapses. Proc. Natl. Acad. Sci. USA 70, 2974-2978.
2. Pavlov, I. P. Conditioned reflexes: An investigation of the physiological activity of the cerebral cortex. Ann.
Neurosci. 17,(1927).
3. Thorndike, E. L. (1898). Animal intelligence: An experimental study of the associative processes in animals. The
Psychological Review: Monograph Supplements, 2(4), i–109. https://doi.org/10.1037/h0092987
4. Junca, P. Bases comportementales et génétiques des apprentissages aversif et appétitif chez l'abeille, Apis
mellifera. Sciences cognitives. Université Paris Sud - Paris XI, 2015. Français. <NNT : 2015PA112240>. <tel-
01293788>
5. Thorndike, E.L. The Fundamentals of Learning. New York : Teachers College Press, 1932.
6. Goodenough, F. L. (1950). Edward Lee Thorndike: 1874-1949. The American Journal of Psychology, 63(2),
291–301.
7. Changeux, JP. (2002), L'homme de vérité, Paris, Odile Jacob.
8. Skinner BF. 1936. The reinforcing effect of a differentiating stimulus. J Gen Psychol 14(2): 263-278.
9. Skinner, B. F. (1936). Conditioning and extinction and their relation to drive. The Journal of General Psychology,
14(2), 296-317.
10. Avarguès-Weber, A., & Giurfa, M. (2013). Conceptual learning by miniature brains. Proceedings. Biological
sciences, 280(1772), 20131907. https://doi.org/10.1098/rspb.2013.1907
11. Giurfa, M. (2013). Cognition with few neurons: higher-order learning in insects. Trends Neurosci. 36, 285–294.
doi: 10.1016/j.tins.2012.12.011
88. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 88/93
Elective Neural Networks - Bibliography.
Bibliography.
12. von Frisch, K. (1967). The Dance Language and Orientation of Bees (Cambridge, Harvard University Press).
13. Menzel, R., and Backhaus, W. (1991). Colour Vision in Insects. In Vision and Visual Dysfunction The Perception
of Colour, P. Gouras, ed. (London, MacMillan Press), pp. 262-288.
14. Giurfa, M., Nunez, J., Chittka, L., & Menzel, R. (1995). Colour preferences of flower-naive honeybees. Journal
of Comparative Physiology A, 177(3), 247-259.
15. Giurfa, M. and Menzel, R. (1997). Insect visual perception: complex abilities of simple nervous systems. Curr.
Opin. Neurobiol. 7, 505-513.
16. Giurfa, Martin & Eichmann, Birgit & Menzel, Randolf. (1996). Symmetry perception in insects. Nature. 382. 458-
461. 10.1038/382458a0.
17. Dacher, M., Lagarrigue, A., and Gauthier, M. (2005). Antennal tactile learning in the honeybee: Effect of nicotinic
antagonists on memory dynamics. Neuroscience 130, 37-50.
18. Giurfa, M. et al. (2001) The concepts of ‘sameness’ and ‘difference’ in an insect. Nature 410, 930–933
19. Avargues-Weber, A, Dyer, AG, Giurfa, M. 2010. Conceptualization of above and below relationships by an
insect. Proc. R. Soc. B 278, 898–905. (doi:10.1098/rspb.2010.1891)
20. Benard, J. (2007). Apprentissage visuel chez l'abeille Apis mellifera: de la généralisation à l'extraction des
règles. Thèse de doctorat, Université Toulouse III – Paul Sabatier , Centre de Recherches sur la Cognition
Animale (CRCA).
21. Devaud, J. et al. (2015) Neural substrate for higher-order learn- ing in an insect: Mushroom bodies are
necessary for configural discriminations. Proc. Natl.Acad.Sci.USA 112, E5854–E5862
22. Kuwabara, M. (1957). Bildung des bedingten Reflexes von Pavlovs Typus bei der Honigbiene, Apis mellifica
JFacHokkaido UnivSerVI Zool 13, 458-464.
89. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 89/93
Elective Neural Networks - Bibliography.
Bibliography.
23. Takeda K. (1961). Classical conditioned response in the honey bee. J Insect Physiol 6: 168–179.
24. Bitterman, M. E., Menzel, R., Fietz, A, & Schafer, S. (1983). Classical conditioning of proboscis extension in
honeybees (Apis mellifera). Journal of Comparative Psychology, 97, 107-119
25. Giurfa M. 2003. Cognitive neuroethology: Dissecting nonelemental learning in a honeybee brain. Curr Opin
Neurobiol 13: 726–735.
26. Giurfa M, Sandoz, J.C. (2012). Invertebrate learning and memory : fifty years of olfactory conditioning of the
proboscis extension response in honeybees. Learn. Mem. 19, 54–66. (doi:10.1101/lm. 024711.111)
27. Giurfa M, Malun, D. (2004). Associative mechanosensory conditioning of the proboscis extension reflex in
honeybees. Learn Mem 11:294–302. doi: 10.1101/lm.63604
28. Matsumoto, Y., Menzel, R., Sandoz, J.-C. and Giurfa, M. (2012). Revisiting olfactory classical conditioning of
the proboscis extension response in honey bees: a step toward standardized procedures. J. Neurosci. Methods
211, 159-167. doi: 10.1016/j.jneumeth.2012.08.018.
29. Vergoz V, Roussel E, Sandoz JC, Giurfa M. 2007. Aversive learning in honeybees revealed by the olfactory
conditioning of the sting extension reflex. PLoS One 2: e288. doi: 10.1371/journal.pone. 0000288.
30. Giurfa M, Fabre E, Flaven-Pouchon J, Groll H, Oberwallner B, Vergoz V, Roussel E, Sandoz JC. 2009. Olfactory
conditioning of the sting extension reflex in honeybees: Memory dependence on trial number, interstimulus
interval, intertrial interval, and protein synthesis. Learn Mem 16: 761–765.
31. Carcaud J, Roussel E, Giurfa M, Sandoz JC. 2009. Odour aversion after olfactory conditioning of the sting
extension reflex in honeybees. J Exp Biol 212: 620–626. doi: 10.1242/jeb.026641.
32. Pouzat, C. (2006). Olfaction chez les insectes. Cours biologie CNRS.
http://xtof.perso.math.cnrs.fr/pdf/Pouzat_111206.pdf
90. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 90/93
Elective Neural Networks - Bibliography.
Bibliography.
33. Hammer M. 1993. An identified neuron mediates the unconditioned stimulus in associative olfactory learning in
honeybees. Nature 366:59–63.
34. Menzel, R. & Giurfa, M. Cognitive architecture of a mini-brain : the honeybee. Trends Cognit. Sci. 5, 62,
(2001). DOI : https://doi.org/10.1016/S1364-6613(00)01601-6.
35. Heisenberg, M. Mushroom body memoir: from maps to models. Nat Rev Neurosci 2003 ; 4:266-75.
36. Cassenaer S, Laurent G. 2007. Hebbian STDP in mushroom bodies facilitates the synchronous flow of
olfactory information in locusts. Nature 448:709–13
37. Galizia, C.G. (2014). Olfactory coding in the insect brain: data and conjectures. Eur J Neurosci 39:1784–
1795. doi: 10.1111/ejn.12558
38. Mizunami, M., Hamanaka, Y., and Nishino, H. (2015). Toward elucidating diversity of neural mechanisms
underlying insect learning. Zool. Lett. 1, 8. doi: 10.1186/s40851-014-0008-6
39. Gallistel CR, Fairhurst S, Balsam P. 2004. The learning curve: Implications of a quantitative analysis. Proc Natl
Acad Sci 101:13124–13131
40. Pamir, E., Chakroborty, N. K., Stollhoff, N., Gehring, K. B., Antemann, V., Morgenstern, L., et al. (2011).
Average group behavior does not represent individual behavior in classical conditioning of the honeybee. Learn.
Mem. 18, 733–741 10.1101/lm.2232711
41. Boitard, C. Identification des réseaux neurobiologiques gouvernant les apprentissages ambigus chez l'abeille
Apis mellifera. Biologie animale. Université Paul Sabatier - Toulouse III, 2015. Français. 〈NNT : 2015TOU30125
〉.〈tel-01378527〉
42. Girard, B. (2010). Modélisation neuromimétique : Sélection de l'action, navigation et exécution motrice. Paris,
France. HDR. Université Pierre et Marie Curie (UPMC). http://www.isir.upmc.fr/files/GirardHdR.pdf
43. Sutton & Barto (1998). Reinforcement Learning : An Introduction. MIT press.
91. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 91/93
Elective Neural Networks - Bibliography.
Bibliography.
44. Girard, B. (2003) Intégration de la navigation et de la sélection de l'action dans une architecture de contrôle
inspirée des ganglions de la base. Neurosciences [q-bio.NC]. Université Pierre et Marie Curie - Paris VI, (2003).
Français. 〈tel-00007683〉
45. Dehaene, S. (2005) Evolution of human cortical circuits for reading and arithmetic: The ‘‘neuronal recycling’’
hypothesis, in From Monkey Brain to Human Brain, Dehaene, S. Duhamel, J.R. Hauser, M. and Rizzolati, G. eds.
Cambridge, MA: MIT Press, 133–157.
46. Dehaene, S. Les neurones de la lecture. Odile Jacob, 2007, 478 p.
47. Snodgrass, R.E. (1956). The Anatomy of the Honeybee. (Ithaca, NY: Cornell University Press).
48. Tedjakumala, S.R., Giurfa, M. (2013). Rules and mechanisms of punishment learning in honey bees: the
aversive conditioning of the sting extension response. J Exp Biol 216 (16): 2985–2997. DOI: 10.1242/jeb.086629
49. Jean-Pierre Changeux, Communications cellulaires [1976], Paris, Collège de France, coll. « Leçons inaugurales
du Collège de France », n° 73, septembre 2013. http://books.openedition.org/cdf/1313
50. Delacour, J. (1987) "Apprentissage et mémoire. Une approche neurobiologique", Masson, Paris.
51. Laurent G. Olfactory network dynamics and the coding of multidimensional signals. Nat Rev Neurosci. 2002;
3:884–95. [PubMed: 12415296] Malsburg Ch Von der (1973) Self-organization of orientation sensitive cells in the
striate cortex. Kybernetik 14:85-100
52. Aso, Y., Hattori, D., Yu, Y., Johnston, R.M., Iyer, N.A., Ngo, T.T., Dionne, H., Abbott, L.F., Axel, R., Tanimoto, H.,
and Rubin, G.M. (2014). The neuronal architecture of the mushroom body provides a logic for associative
learning. eLife 3, e04577.
53. Aso, Y., Hattori, D., Yu, Y., Johnston, R.M., Iyer, N.A., Ngo, T.T., Dionne, H., Abbott, L.F., Axel, R., Tanimoto, H.,
and Rubin, G.M. (2014). The neuronal architecture of the mushroom body provides a logic for associative
learning. eLife 3, e04577.
92. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 92/93
Elective Neural Networks - Bibliography.
Bibliography.
54. Uchida N, Poo C, Haddad R. 2014. Coding and transformations in the olfactory system. Annual Review of
Neuroscience 37:363–385. doi: 10.1146/annurev-neuro-071013-013941.
55. Jortner, R., Farivar, S. S. & Laurent, G. A simple connectivity scheme for sparse coding in an olfactory system.
J. Neurosci. 27, 1659–1669 (2007).
56. Prochiantz, A. Processus morphogénétiques Cours et travaux du Collège de France. Annuaire 113e année,
Collège de France, Paris, avril 2014, p. 331-353. ISBN 978-2-7226-0329-5.
http://journals.openedition.org/annuaire-cdf/2432 ; DOI : https://doi.org/10.4000/annuaire-cdf.2432
57. Werker JF, Hensch TK. 2015 Critical periods in speech perception: new directions. Amu. Rev. Psychol
66:173–96
58. K. Lorenz, Studies in Animal and Human Behaviour (Harvard University Press, Cambridge, Massachusetts,
1970), Vol. 1.
59. K. Lorenz, Studies in Animal and Human Behaviour (Harvard University Press, Cambridge, Massachusetts,
1971), Vol. 2.
60. Hubel, D.H., Wiesel, T.N.. Receptive fields of single neurons in the cat’s striate cortex. Journal of Physiology,
148:574–591, 1959.
61. Wiesel TN. Postnatal development of the visual cortex and the influence of environment. Nature 299:583-591
(1982).
62. Wiesel TN, Hubel DH. Comparison of the effects of unilateral and bilateral eye closure on cortical unit responses
in kittens. J Neurophysiol 28:10209-1040 (1965).
63. Harwerth RS, Smith EL 3rd, Duncan GC, Crawford ML, von Noorden GK. Multiple sensitive periods in the
development of the primate visual system. Science. 1986;232(4747):235‐238. doi:10.1126/science.3952507
93. 13/10/2020 Elective neural Networks 2. The Orthogonal Brain 93/93
Elective Neural Networks - Bibliography.
Bibliography.
64. Scheiman MM, Hertle RW, Beck RW, Edwards AR, Birch E, Cotter SA, Crouch ER, Cruz OA, Davitt BV,
Donahue S, Holmes JM, Lyon DW, Repka MX, Sala NA, Silbert DI, Suh DW, Tamkins SM, Pediatric Eye Disease
Investigator Group (2005) Randomized trial oftreatment of amblyopia in children aged 7 to 17 years. Arch
Ophthalmol 123:437–447.
65. Testa, D. Contrôler la plasticité du cortex cérébral adulte à travers l'action non-autonome de l'homéoprotéine
Otx2. Neurosciences [q-bio.NC]. Sorbonne Université, 2018. Français. ⟨NNT : 2018SORUS234⟩. ⟨tel-02479650⟩
66. Bavelier D, Levi DM, Li RW, Dan Y, Hensch TK. Removing brakes on adult brain plasticity: from molecular to
behavioral interventions. J Neurosci. 2010;30(45):14964‐14971. doi:10.1523/JNEUROSCI.4812-10.2010
67. M. Beurdeley (co-first), J. Spatazza (co-first), H.H.C. Lee (co-first), S. Sugiyama, C. Bernard, A.A. Di Nardo, T.K.
Hensch* et A. Prochiantz*, « Otx2 binding to perineuronal nets persistently regulates plasticity in the mature visual
cortex », J. Neurosci, 32, 2012, 9429-9437.
68. Apulei J, Kim N, Testa D, et al. Non-cell Autonomous OTX2 Homeoprotein Regulates Visual Cortex Plasticity
Through Gadd45b/g. Cereb Cortex. 2019;29(6):2384‐2395. doi:10.1093/cercor/bhy108
69. Dehaene, S., (2015) Education, plasticité cérébrale et recyclage neuronal. Cours du Collège de France :
Fondements cognitifs des apprentissages scolaires. https://www.college-de-france.fr/media/stanislas-
dehaene/UPL597651672277817909_Cours2015_1.pdf
70. Changeux, J.-P., 2008, Du vrai, du beau, du bien. Une nouvelle approche neuronale, Odile Jacob.
___________________________________