Where is mymind ?How embodiment redefines the limits ofthe model and the role of the observer

Nicolas P. RougierINRIA Bordeaux Sud-Ouest

Institute of Neurodegenerative Diseases

13ème Forum des Sciences Cognitives22 Mars 2014 - Réfectoire des Cordeliers, Paris

What is cognition ?No unique definition, many forms in many species

The case of Caenorhabditis elegansA few neurons for a complex behavior

Sensory motor behaviorQ. Wen, M.D. Po, E. Hulme, S. Chen, X. Liu, S. Wai Kwok, M. Gershow, A. M. Leifer, V. Butler, C. Fang-Yen, T.Kawano, W.R. Schafer, G. Whitesides, M. Wyart, D.B. Chklovskii, M. Zhen, A.D.T. Samueln, ProprioceptiveCoupling within Motor Neurons Drives C. elegans Forward Locomotion. Neuron, 2012.

Plasticity, learning andmemoryH. Sasakura and I. Mori, Behavioral plasticity, learning, andmemory in C. elegans, Current Opinion inNeurobiology, 2013.

Decision makingT.A. Jarrell, Y. Wang, A.E. Bloniarz, C.A. Brittin, M. Xu, J.N. Thomson, D.G. Albertson, D.H. Hall, S.W. Emmons,The Connectome of a Decision-Making Neural Network, Science, 2012.

302 neurons, is that enough ?→ For a worm life, definitely

→ For cognition, unlikely (but unproven)

What really matters for cognition ?Is there a critical mass ?

Some figures

• C.Elegans→ 302 neurons

• Mouse→ 71,000,000 neurons

• Rat→ 200,000,000 neurons

• Macaque→ 6,376,000,000 neurons

• Human→ 86,000,000,000 neurons

Some wrong asumptions

• Body size matters

• Brain size matters

• Proportions and relative size matter

The human brain in numbers [...], S. Herculano-Houzel,Frontier in Human Neuroscience, 2009

What really matters for cognition ?The amazing case of the octopus

Specific structures ?

• Neocortex

• Basal Ganglia

• Amygdala

• Frontal cortex

Specific architecture ?

• Connectivity

• Density

• Modularity

• Self-organisation

Specific abilities ?

• To recognize self/others

• To imitate

• To use tools

• To communicate

Where do we start ?Theoretical frameworks

Biological framework

→ Anatomical facts

→ Physiological recordings

→ Experimental data

Cognitive framework

→ Subsumption architecture

→ Embodied cognition

→ A�ordances, emotions, etc.

Computational framework

→ Computational paradigm

→ Plasticity & learning

→ Evaluation of models

Philosophical framework

→ Strong AI / Weak AI

→ Emergence

→ Theories of mind

Where do we start ?How to model cognition ?

What is/are the right biological level(s) of description ?

• Molecule ?→ neurotransmitters

• Organelle ?→ axons, dendrites, synapses

• Cell ?→ neurons, glial cells

• Tissue ?→ brain lobes & structures

• Organ ?→ brain

The curse of the homunculusInfinite regression

Cartesian theater (Descartes, 1641)Cartesian dualism separates the Mindfrom the Body, the earlier controlling thelatter.

Central Executive (Baddeley & Hirsh, 1974)The central executive is responsible for thecontrol and regulation of cognitiveprocesses.

Implicit supervisor (Kohonen, 1982)Kohonenmaps need a supervisor todesignate the winner.

Computational FrameworkDistributed, Asynchronous, Numerical & Adaptive

A unit is a set of arbitrary values that can vary along time under the influence of otherunits and learning.

• Distributed→ No supervisor nor executive

• Asynchronous→ No central clock

• Numerical→ No symbols

• Adaptive→ To learn something

Group Gro

up

Unit

Link

Layer

Link Val

ue

Value

Value

Wewant to make sure that emerging properties are those of the model and not thoseof the so�ware running the model.

What is a decision ?(Rougier& Hutt, 2011)

A dual particle system

(1) x = −αx + (x − y)(1− x) + αIx(2) y = −αy+ (y− x)(1− y) + αIy

with ∀t, x(t) > 0, y(t) > 0

Each particle tends to inhibit the other

Three fixed pointsUsing Ix = Iy = 1→ x = 1, y = 1→ x = 1− α, y = 1→ x = 1, y = 1− α

0.0 0.2 0.4 0.6 0.8 1.0x

0.0

0.2

0.4

0.6

0.8

1.0

y

50 trajectories of a dual particle system (x,y)

The decision is influenced by noise, initial conditions and evaluation order.

Dynamic Neural Fields(Wilson & Cowan, 1972)

EquationLet u(x,t) be the membrane potential at position x and time t, f a transfer functionandw a kernel of lateral interaction. The temporal evolution of u(x,t) is given by:

τ

time constant

·∂u(x, t)∂t

= −u(x, t)

leak term

+

∫ +∞

−∞w(x − y) · f(u(y, t))

lateral interactions

dy + I(x)

input

+ h

resting potential

Some (remarkable & remarked) properties

From distributed computations to decision(Rougier & Vitay, 2006)

Visual decision

• A simple neural field• Decision making• Robustness to noise & distractors

Saliency

Input

Focus

Competition

Embodiment of the modelIs there an objective motor behavior ?

Curious (visual tracking) ?

Saliency

Input

Focus

Competition

Camera

We can connect the model to a pan-tilt camerasuch that it follows a given stimulus

Shy (visual avoidance) ?

Saliency

Input

Focus

Competition

Camera

We can connect the model to a pan-tilt camerasuch that it looks away from a given stimulus

The actual behavior of the model depends on the links to its body. Ultimately, themodeler is the one who decide on the behavior.

Embodiment of the modelIs there an objective sensory behavior ?

The superior colliculusSuperficial layers are sensory-related, and receive input from the eyes as well as other sensory systems.

Intermediate layers integrates multi-sensory cells and motor properties.

Deep layers are motor-related, capable of activating eye movements as well as other responses

Many computational models(Droulez & Berthoz, 1991), (Optican, 1995), (Trappenberg,2001), etc.

Superior ColliculusA shared decision model (Taouali et al., 2014)

The anatomy of the retina makes the projection to be magnified (rostral to caudal axis) and transformed. Theprojection of a physical line is not a “neural” line. A sensori-motor perspective is necessary (O’Regan & Noë,2001).

Superior ColliculusA shared decision model (Taouali et al., 2014)

Shared decisionThe express saccade decision depends on the projection of the stimulus onto the superior colliculus. Rostralstimuli appears bigger and are more likely to be selected a�er the competition process.

The actual behavior of the model depends on the body (retina) and the outer world.

From decision to behaviorExploring the visual world

Attention (Vitay & Rougier, 2005)Working memory allows to keep tracked of whathave been already seen

Bias/GatingSaliency

Memory

Update

Focus

InputCompetition

Memorization

Anticipation (Fix et al., 2007)Anticipation allows to anticipate saccadeconsequences on sensory input

Bias/GatingSaliency

Memory

Anticipation

Update

Saccade

Bias

Focus

Prediction

InputCompetition

Memorization

Toward the organization of visual behaviorA bottom-up sequential exploratory behavior can emerge from distributed & numericcomputations. No need of a central executive nor homunculus.

Roger the robotRoger loves orange, or does it ?

What is Roger actually doing ?Does it even “know” what it is doing ?

From an external point of view, we cannot know what’s Roger “thinking”

So what ?

Back to modelingWhat is a model ?

Supposons qu’un être (ou une situation) extérieur(e) X présente un comportement énigmatique, etque nous nous posions à son sujet une (ou plusieurs) question(s). Pour répondre à cette question,on va s’e�orcer demodéliser X, c’est-à-dire, on va construire un objet (réel ou abstrait)M, considérécomme l’image, l’analogue de X :M sera dit lemodèle de X.

R. Thom,Modélisation et scientificité, 1978

Question (Q')

Question (Q)

Being (X)

Analogy

Model (M)

Answer (A')

Answer (A)

To an observer B, an object A* is a model of an object A to the extent that B can use A* to answerquestions that interest him about A.

M. Minsky,Matter, Mind and Models, 1965

The role of the observerThere is no such thing as objective observations

The standard modelprovides a direct access to the question

The computational modelquantifiable objective propertiesmany possible interpretations of behavior

The embodied modelobjective & quantifiable performances

The cognitive modelobservation through interactionqualification of behavior depends on behavior

MODEL OBSERVER

The role of the observerThere is no such thing as objective observations

The standard modelprovides a direct access to the question

The computational modelquantifiable objective propertiesmany possible interpretations of behavior

The embodied modelobjective & quantifiable performances

The cognitive modelobservation through interactionqualification of behavior depends on behavior

Brain

MODEL OBSERVER

Brain

The role of the observerThere is no such thing as objective observations

The standard modelprovides a direct access to the question

The computational modelquantifiable objective propertiesmany possible interpretations of behavior

The embodied modelobjective & quantifiable performances

The cognitive modelobservation through interactionqualification of behavior depends on behavior

Brain

Body

MODEL OBSERVER

Brain

Body

The role of the observerThere is no such thing as objective observations

The standard modelprovides a direct access to the question

The computational modelquantifiable objective propertiesmany possible interpretations of behavior

The embodied modelobjective & quantifiable performances

The cognitive modelobservation through interactionqualification of behavior depends on behavior

Brain

BodySENSORY/MOTOR

World

MODEL

OBSERVATIONINTERACTION

OBSERVER

Brain

BodySENSORY/MOTOR

World

What kind of cognition for an artificial being ?The Geminoid project, far beyond the uncanny valley

Ultra-realistic, high expectations, poor interactions

http://www.geminoid.jp

What kind of cognition for an artificial being ?The Pinokio project, not even started the uncanny journey

A simple lamp, low expectations, great interactions

http://www.pinokio-lamp.com

Behavioral Turing testDiscarding the verbal dimension

Perceptual interactions in a minimalist virtual environment“How in real-life or through the use of technical devices can we recognize thepresence of other persons and under what conditions can we di�erentiate them fromobjects?” [M.Auvray, C.Lenay, J.Stewart, 2009]

Modeling the Minimal Newborn’s Intersubjective Mind“... what could be the minimal neural core responsible for the development of theneonate social brain” [A.Pitti, Y.Kuniyoshi, M.Quoy, P.Gaussier, 2013]

Infants Learn to Imitate by Being Imitated“the development of imitation in human infants relies on associative learning ratherthan on innate knowledge.” [S.S.Jones, 2007]

Where is mymind ?Toward an embodied and social theory of mind

The Eye of the BeholderPart of the cognition we lend to others may be rooted in our own cognition.

The Body of the BeholderSuch process may depend on the intersection for the sensory andmotor apparatusthat constrain interactions. If there is no common grounds, we cannot interact. Theprocess is broken.

ConclusionTomodel or to make cognition ?

Major cognitive and behavioral functions emerge from adaptive sensorimotor loopsinvolving the external world, the body and the brain. We study, model and implementsuch loopsand their interactions towarda fully autonomousbehavior. With sucha sys-temic approach, we mean that such complex systems can only be truly apprehendedas a whole and in natural behavioral situation.

Mnemosyne project-team, INRIA Bordeaux, 2012

Brain, body & cognition

Embodiment constrains cognitionEmotions play a critical roleBehavior is motivated

New questions arise

How to understand our ownmodels ?How do we identify behavior ?What is the critical mass ?

ConclusionTomodel or to make cognition ?

Trying to understand perception by studying only neurons is like trying to understand bird flightby studying only feathers: It just cannot be done. In order to understand bird flight, we have tounderstand aerodynamics; only then do the structure of feathers and the di�erent shapes of birds’wings make sense. D.Marr, Vision, 1982

But there are other ways to fly...