Transcript D(-1) - Elte

An Evolutionary Framework for
Neuronal Architectures
Eörs Szathmáry
Collegium Budapest
Eötvös University
The group
•
•
•
•
•
•
•
•
•
Zoltán Szatmáry
Péter Ittzés
Máté Varga
Ferenc Huszár
Anna Fedor
István Zachar
Gergő Orbán
Máté Lengyel
Szabolcs Számadó
programming, neuro
programming, bio
programming, elect. eng.
informatics
bio, ethol
bio, evol
biophys, Bayesian learn
neuro
bio, evol
It all started with JMS…
• „You know Eörs, we have
to consider language
seriously in the book”
• The origin of language
remains the primary
motivation behind this
work
The major transitions
(JMS & ES, 1995)
*
*
*
*
* These transitions are regarded to be ‘difficult’
Why is language so interesting?
• Because everybody knows that only we talk
• …although other animals may understand a
number of words
• Language makes long-term cumulative
cultural evolution possible
• A novel type of inheritance system with
showing “unlimited hereditary” potential
What is so special about human
language?
• Basically, it is the fact that we make sentences
using grammar
• Languages are translatable into one another with
good efficiency
• Some capacity for language acquisition seems to
be innate
• THE HOLY GRAIL IS THE EMERGENCE OF
SYNTACTICAL PRODUCTION?
Three interwoven processes
• Note the different time-scales involved
• Cultural transmission: language transmits itself as
well as other things
• A novel inheritance system
The case of Nicaraguan sign
language: something seems to be
innate
• School for deaf
children was
opened 30 years
ago.
• People range from 4
to 45 years by now.
Development of NSL
• NSL has evolved from a system of nonlinguistic
gestures into a full sign language with its own
grammar that continues to expand and mature
• The youngest children in the NSL community are
the most fluent signers
• Deaf Nicaraguan children have created their own
language independently of exposure to a
preexisting language structure.
• Language is so resilient that it can be triggered by
exposure to a linguistic input that is highly limited
and fragmented—an indication of the fundamental
innateness of grammar  language readiness
A note on semantix and syntax
• The fact that today one can dissociate
semantics from syntax does not mean that
they were dissociated throughout language
evolution
• If language is efficacious, then selection
acted on semantics
• Emerging syntax thus was semantically
constrained
Challenges: a simple experiment
(Hauser & Fitch)
• Habituation experiments
• Finite state grammar
(AB)n is recognizable by
tamarins
• Phrase structure grammar
AnBn is NOT.
• Human students
recognize both
BUT: Recursive syntactic pattern
learning in birds!
• European starlings (Sturnus vulgaris) accurately
recognize recursive syntactic patterns
• They are able to exlude agrammatical forms
• Centre-embedding is not uniquely human
Patterns are made up of naturally
occurring vocal patterns
• Learning to classify by operant conditioning
• This is NOT production!
The genetics of complex behaviour
is not easy…
• Pleiotropy: one gene affecting different traits
• Epistasis: effects from different genes do not combine
independently
• Intermediate phenotypes must be identified!
Genetic analysis of fruitfly behaviour
The FOXP2 gene is mutant in a
family with SLI
• SLI: specific language impairment
• In the KE family the mutation is a single
autosomal dominant allele
• Another individual has one copy deleted
• TWO intact copies must be there in humans!
• The mutation affects morphosyntax: Yesterday I
went to the church and talk to nanny brother
• Chromosome 7, forkhead protein
Nucleotide substitutions in the
FOXP2 gene
• Bars are nucleotide substitutions
• Grey bars indicate amino acid changes
• Likely to have been recent target of selection
FOXP2 seems even more interesting
• FOXP2 single nucleotide polymorphism (mainly
in the 5’ regulatory region) associates with
schizophrenia with auditory hallucinations
• FOXP2 is under stabilizing selection (even on
synonymous changes) in song-learning birds
(human mutations are not seen), but not in vocallearning mammals or in non-singing birds
• the human-unique substitution in exon 7 (T303N)
was flanked by two changes in both whale and
dolphin (S302P and T304A)
FOXP2 seems even more interesting II
• studies in songbirds show that during times of
song plasticity FoxP2 is upregulated in a striatal
region essential for song learning
• FOXP1 and FOXP2 expression patterns in human
fetal brain are strikingly similar to those in the
songbird
• including localization to subcortical structures that
function in sensorimotor integration and the
control of skilled, coordinated movement.
• The specific co-localization of FoxP1 and FoxP2
found in several structures in the bird and human
brain predicts that mutations in FOXP1 could also
be related to speech disorders.
More on FOXP2
• fMRI: underactivity of Broca during word
generation
• repetition of non-words with complex articulatory
patterns: the core deficit is one of sequential
articulation of phonological units
• FOXP2 mutation could have been responsible for
the perfecting of speech
• How would it affect the mirror system?
An evaluation of selective scenarios:
Trends Ecol. Evol. in press
Selective scenarios for the emergence of
natural language
Szabolcs Számadó and Eörs Szathmáry
Collegium Budapest (Institute for Advanced Study), Szentháromság u. 2, H-1014, Budapest, Hungary
Corresponding author: Számadó, S. ([email protected]).
The recent blossoming of evolutionary linguistics has resulted in a variety
of theories that attempt to provide a selective scenario for the evolution of
early language. However, their overabundance makes many researchers
sceptical of such theorising. Here, we suggest that a more rigorous
approach is needed towards their construction although, despite justified
scepticism, there is no agreement as to the criteria that should be used to
determine the validity of the various competing theories. We attempt to fill
this gap by providing criteria upon which the various historical narratives
can be judged. Although individually none of these criteria are highly
constraining, taken together they could provide a useful evolutionary
framework for thinking about the evolution of human language.
(1) selective advantage (2) honesty (3) grounded in reality
(4) power of generalisations (5) cognitive abilities (6) uniqueness
Theories/Questions
1
2
3
4
5
6
Language as a mental tool (Jerison, 1991; Burling, 1993)
+
+
-
+
-
-
Grooming hypothesis (Dunbar, 1998)
-
+
-
-
-
-
Gossip (Power, 1998)
+
+
+
+
-
+
-
+
+
-
+
-
-
+
-
+
-
+
-
-
+
-
+
+
+
+
+
+
+
-
-
Tool making (Greenfield, 1991)
Mating contract (Deacon, 1997)
Sexual selection (Miller, 2000)
Status for information (Dessalles, 2000)
Song hypothesis (Vaneechoutte & Skoyles, 1998)
Group bonding/ ritual (Knight, 1998)
Gestural theory (Hewes, 1973)
Hunting theories (Washburn & Lanchester, 1968)
The evolutionary approach
genes
selection
development
learning
behaviour
environment
Impact of evolution on the developmental
genetics of the brain!
One method of finding out (within
ECAgents)
•
•
•
•
Simulated dynamics of interacting agents
Agents have a “nervous system”
It is under partial genetic control
Selection will be based on learning performance
for symbolic and syntactical tasks
• If successful, look and reverse engineer the
emerging architectures
• HOW GENES RIG THE NETWORKS??
Between linguistic input and
output…
Transmission dynamics in
simulated agents
The most important precedent
„the purpose of this paper is to explore how genes could
specify the actual neuronal network functional
architectures found in the mammalian brain, such as those
found in the cerebral cortex. Indeed, this paper takes
examples of some of the actual architectures and
prototypical networks found in the cerebral cortex, and
explores how these architectures could be specified by
genes which allow the networks when built to implement
some of the prototypical computational problems that must
be solved by neuronal networks in the brain”
Highly indirect genetic encoding
• There are special results with direct genetic
encoding (one gene per neuron or per
synapse)
• THIS IS NOT WHAT WE WANT
• There are around 35 thousand genes
• Only a fraction of them can deal with the
brain
• Billions of neurons, more synapses
Summary of our efforts
In: Nehaniv, C., Cangelosi, A & Lyon, C.
(2005) Origin of Communication, in press.
Springer-Verlag
Software architecture
519
classes
99267
lines of
C++ code
ECAgents: project founded by the Future and Emerging Technologies program
(IST-FET) of the European Community under EU R&D contract IST1940.
Population dynamics and agent lifecycle
ECAgents: project founded by the Future and Emerging Technologies program
(IST-FET) of the European Community under EU R&D contract IST1940.
Ontogenesis of a neuronal network
ECAgents: project founded by the Future and Emerging Technologies program
(IST-FET) of the European Community under EU R&D contract IST1940.
A note on the importance of
topographicity
• For each tropographical net, one can construct an
equivalent topological net
• The nature of variation is very different for the two
options
• Genes obviously affect topographical networks
ECAgents: project founded by the Future and Emerging Technologies program
(IST-FET) of the European Community under EU R&D contract IST1940.
Parameter values
Population dynamics and games
• Population size: 100.
• Time steps: 500 (200 for the cloning test).
• Number of games played per time step per agent:
100.
• Death process: least fit (5).
• Mating process: roulette wheel.
• Number of offspring: Poisson with Lambda=5.
ECAgents: project founded by the Future and Emerging Technologies program
(IST-FET) of the European Community under EU R&D contract IST1940.
Parameter values 2
Neurobiological parameters
• Number of layers: randomly chosen from the range [1,3]
(mutation rate: 0.008).
• Number of neuron classes: randomly chosen from the range
[1,3] (mutation rate: 0.2).
• Number of neurons: randomly chosen from the range [10,30]
(mutation rate: 0.2).
• Number of projections: randomly chosen from the range [1,3]
(mutation rate: 0.02).
• Rate coding with linear transfer function [-1 , 1].
• Hebbian learning rules.
• Reward matrix is same as the pay-off matrix of the given game
(below).
• Brain update: 10 (same for listener and speaker).
ECAgents: project founded by the Future and Emerging Technologies program
(IST-FET) of the European Community under EU R&D contract IST1940.
Task: A two-person game
There are:
• two kinds of environments, E={-1,1},
• three types of cost-free signals S=[-1, 1, else],
• three types of possible decisions D=[-1, 1],
where values other than –1 or 1 mean no signal and no
response respectively.
ECAgents: project founded by the Future and Emerging Technologies program
(IST-FET) of the European Community under EU R&D contract IST1940.
A Coordination Game
Speaker
Listener
Population
Signal
-1/1
Environment
Decision
Decision
ECAgents: project founded by the Future and Emerging Technologies program
(IST-FET) of the European Community under EU R&D contract IST1940.
Different types of game
Environment -1
Coodination game (Coop)
 Division of Labour (Div)
 Prisoners’ dilemma (PD)
 Hawk- Dove game (SD)
Coop (-1)
Coop (1)
Div
Div
PD (-1)
PD (1)
SD (-1)
SD (1)
PD (-1)
Coop (-1)
PD (-1)
CoopRev (1)
SD (-1)
Coop (-1)
SD (-1)
CoopRev (1)

Coodination game
Division of Labour
D(1)
D(-1)
D(1)
D(-1)
D(1)
5
1
D(1)
0
5
D(-1)
0
0
D(-1)
5
0
Environment 1
Prisoners’ dilemma Hawk - Dove game
D(1)
D(-1)
D(1)
1
5
D(-1)
0
3
D(1)
D(-1)
D(1)
-1
5
D(-1)
0
3
Div/Div
Coop/Coop
3000
5000
4500
2500
4000
S:[-1] E:[-1] DL:[-1];DS:[-1]
3000
S:[0] E:[-1] DL:[-1];DS:[-1]
2500
S:[0] E:[-1] DL:[1];DS:[-1]
2000
S:[0] E:[1] DL:[-1];DS:[1]
decisions
decisions
3500
S:[0] E:[1] DL:[1];DS:[1]
1500
S:[-1] E:[-1] DL:[-1];DS:[1]
2000
S:[-1] E:[-1] DL:[1];DS:[-1]
S:[-1] E:[1] DL:[-1];DS:[1]
1500
S:[-1] E:[1] DL:[1];DS:[-1]
S:[1] E:[-1] DL:[-1];DS:[1]
1000
S:[1] E:[-1] DL:[1];DS:[-1]
S:[1] E:[1] DL:[1];DS:[1]
1000
S:[1] E:[1] DL:[-1];DS:[1]
500
S:[1] E:[1] DL:[1];DS:[-1]
500
0
0
1
101
201
301
401
501
601
701
801
901
1
101
201
301
401
time
PD/PD
601
701
801
901
SD/SD
2000
4500
1800
S:[-1] E:[-1] DL:[-1];DS:[1]
1600
S:[-1] E:[-1] DL:[1];DS:[1]
1400
S:[-1] E:[1] DL:[-1];DS:[-1]
4000
3500
S:[-1] E:[1] DL:[1];DS:[-1]
1200
S:[0] E:[-1] DL:[-1];DS:[1]
1000
S:[0] E:[-1] DL:[1];DS:[1]
S:[0] E:[1] DL:[-1];DS:[-1]
800
S:[0] E:[1] DL:[1];DS:[-1]
600
S:[1] E:[-1] DL:[-1];DS:[1]
400
S:[1] E:[-1] DL:[1];DS:[1]
3000
decisions
decisions
501
time
S:[-1] E:[1] DL:[-1];DS:[-1]
2500
S:[-1] E:[1] DL:[1];DS:[-1]
2000
S:[-1] E:[1] DL:[1];DS:[1]
1500
S:[1] E:[-1] DL:[-1];DS:[-1]
S:[1] E:[-1] DL:[-1];DS:[1]
1000
S:[1] E:[1] DL:[-1];DS:[-1]
200
S:[1] E:[1] DL:[1];DS:[-1]
S:[1] E:[-1] DL:[1];DS:[1]
500
0
1
101
201
301
401
501
601
701
801
901
time
0
1
101
201
301
401
501
601
701
801
time
other-reporting signals
self-reporting signals
dishonest signals
uninformative signals
no signal
901
Why is there communication in
SD/SD?
• There is conflict of interest in the game, BUT:
• There is mixed ESS: it pays to be the reverse of
the opponent!
• Speaker sees the environment, chooses the selfish
strategy and and informs the listener about it in the
„hope” that the other behaves complementarily.
The other has no real choice but to „believe” in it.
• Mixed ESS AND changing environments AND
informational asymmetry RESULT IN
communication
PD/Coop
SD/Coop
2500
3500
3000
2000
S:[-1] E:[-1] DL:[-1];DS:[-1]
2500
S:[-1] E:[-1] DL:[-1];DS:[1]
1500
S:[-1] E:[1] DL:[-1];DS:[-1]
S:[0] E:[-1] DL:[-1];DS:[-1]
S:[0] E:[-1] DL:[-1];DS:[1]
1000
decisions
decisions
S:[-1] E:[-1] DL:[-1];DS:[-1]
S:[-1] E:[-1] DL:[-1];DS:[1]
S:[-1] E:[1] DL:[-1];DS:[-1]
2000
S:[0] E:[-1] DL:[-1];DS:[-1]
S:[0] E:[-1] DL:[-1];DS:[1]
1500
S:[0] E:[1] DL:[-1];DS:[-1]
S:[0] E:[1] DL:[-1];DS:[-1]
S:[1] E:[-1] DL:[-1];DS:[1]
500
S:[1] E:[-1] DL:[-1];DS:[-1]
1000
S:[1] E:[-1] DL:[-1];DS:[1]
S:[1] E:[1] DL:[-1];DS:[-1]
S:[1] E:[1] DL:[-1];DS:[-1]
500
0
1
101
201
301
401
501
601
701
801
0
901
1
101
201
301
401
time
501
601
701
801
901
time
PD/CoopRev
SD/CoopRev
3000
2500
2500
1500
S:[-1] E:[-1] DL:[1];DS:[1]
1000
S:[-1] E:[1] DL:[1];DS:[1]
decisions
decisions
2000
2000
S:[-1] E:[-1] DL:[1];DS:[-1]
S:[-1] E:[1] DL:[1];DS:[1]
1500
S:[0] E:[-1] DL:[1];DS:[-1]
S:[0] E:[-1] DL:[1];DS:[0]
1000
S:[0] E:[1] DL:[1];DS:[1]
S:[0] E:[-1] DL:[1];DS:[1]
S:[0] E:[1] DL:[1];DS:[1]
500
S:[1] E:[-1] DL:[1];DS:[-1]
500
S:[1] E:[-1] DL:[1];DS:[1]
S:[1] E:[-1] DL:[1];DS:[1]
S:[1] E:[1] DL:[1];DS:[1]
0
1
101
201
301
401
501
601
701
801
901
time
S:[1] E:[1] DL:[1];DS:[1]
0
1
101
201
301
401
501
601
701
801
901
time
other-reporting signals
self-reporting signals
dishonest signals
uninformative signals
no signal
Early brains (t:10)
Scenario: E1: complementary, E-1:same
Visual input
Audio input
Mixed colours indicate
input mixing.
Const input or unconnected
ECAgents: project founded by the Future and Emerging Technologies program
(IST-FET) of the European Community under EU R&D contract IST1940.
Advanced brain (t:750)
Scenario: E1: complementary, E-1:same
Visual input
Mixed colours indicate input mixing
Audio input
Constants input or unconnected
ECAgents: project founded by the Future and Emerging Technologies program
(IST-FET) of the European Community under EU R&D contract IST1940.
Is there inheritance, despite highly
indirect genetic encoding?
• Scatter plots for
AudioIn, AudioOut,
Const, Vision and
Decision neurons
• Experiments on clones
ECAgents: project founded by the Future and Emerging Technologies program
(IST-FET) of the European Community under EU R&D contract IST1940.
Measuring the Heritability of
Neural Connections
in ENGA-Generated Communicating Agents
The central issue with indirect encoding is whether one
can find heritability of the simulated, evolved neuronal
networks. If our biomimetic, indirect encoding is
successful; this should be the case.
Input/output
neuron
h2
AudioIn
0.8689
AudioOut
0.8708
Const
0.8696
Decision
0.8123
Vision
0.8428
Estimated heritability values (h2) of the number
of connections of the given input/output neurons
(right).
This is a proof that ENGA works as we hoped:
despite indirect encoding, there is hereditary
variation between indivudal phenotypes on which
simulated natural selection can act.
ECAgents: project founded by the Future and Emerging Technologies program
(IST-FET) of the European Community under EU R&D contract IST1940.
Is there inheritance, or only
council of the elders?
•
•
•
•
The increase with age of time
The code of individuals of time
Green lines: individual living still the end of the simulation
Red: birth events
ECAgents: project founded by the Future and Emerging Technologies program
(IST-FET) of the European Community under EU R&D contract IST1940.
Details of learning/heritability
experiment
•
•
•
•
•
Individuals are taken from an equilibrated Coop game
All are newborn, no close relatives
Smart and stupid individuals are included
Individuals were educated in a testbed
You see the average of the reward received in 1010
turns
• Convention carved into pieces: two environments x
two types of input (audio and visual), measure the
signal or the decision
ECAgents: project founded by the Future and Emerging Technologies program
(IST-FET) of the European Community under EU R&D contract IST1940.
Is there inheritance of behaviour?
ECAgents: project founded by the Future and Emerging Technologies program
(IST-FET) of the European Community under EU R&D contract IST1940.
What next?
• For example, do the Fitch-Hauser experiment
• Select for networks that do finite state grammar and
that do central embedding
• If successful, look at the networks
• What is an ‘easy’ evolutionary path?
ECAgents: project founded by the Future and Emerging Technologies program
(IST-FET) of the European Community under EU R&D contract IST1940.