Transcript The Major Transitions in Evolution

Evolution IN the brain?
Eörs Szathmáry
Munich
Collegium Budapest
Eötvös University
The neuronal replicator hypothesis
Chrisantha Fernando, Eörs Szathmáry
University of Sussex AND Collegium Budapest
Units of evolution
1. multiplication
2. heredity
3. variability
Some hereditary traits affect
survival and/or fertility
The importance of cumulative
selection
• Natural selection is a non-random process.
• Evolution by natural selection is a
cumulative process.
• Cumulative selection can produce novel
useful complex structures in relatively short
periods of time.
The extended evolutionary synthesis
• Extending the Modern Synthesis in depth:
e.g. questions of evolvability and evolution
of development
• Extending the Modern Synthesis in width:
e.g. chemical, linguistic and neuronal
replicators
• Forthcoming Book from the Altenberg
meeting (Konrad Lorenz Institute)
Von Kiedrowski’s replicator
What makes us human?
• Note the different time-scales involved
• Cultural transmission: language transmits itself as
well as other things
• A novel inheritance system
Evolution OF the brain
Fluid Construction Grammar
with replicating constructs (with Luc Steels)
• selective amplification by linked replication
• mutation, recombination, etc.
William James, 1890
• Every scientific conception is, in the first instance,
a 'spontaneous variation' in someone's brain. For
one that proves useful and applicable there are a
thousand that perish through their worthlessness.
Their genesis is strictly akin to that of the flashes
of poetry and sallies of wit to which the instable
brain-paths equally give rise. But whereas the
poetry and wit (like the science of the ancients)
are their own excuse for being ... the 'scientific'
conceptions must prove their worth by being
'verified'. This test, however, is the cause of their
preservation, not of their production…
Baldwin, 1889
• And how far the method of law called by Darwin
"natural selection" goes, what its range really is,
we are now beginning to see in its varied
applications in the sciences of life and mind. It
seems to be--unless future investigations set
positive limits to its application--a universal
principle; for the intelligence itself, in its
procedure of tentative experimentation, or "trial
and error," appears to operate in accordance with
it.
Monod, 1971
• For a biologist it is tempting to draw a parallel
between the evolution of ideas and that of the
biosphere. For while the abstract kingdom stands
at a yet greater distance above the biosphere than
the latter does above the nonliving universe, ideas
have retained some of the properties of organisms.
Like them, they tend to perpetuate their structure
and to breed; they too can fuse, recombine,
segregate their content; indeed they too can
evolve, and in this evolution selection must surely
play an important role. I shall not hazard a
theory of the selection of ideas. But one may at
least try to define some of the principal factors
involved in it. This selection must necessarily
operate at two levels: that of the mind itself and
that of performance.
Dawkins 1971, 1976
• Selective neurone death as a possible
memory mechanism (1971)
• Conceptualization of the meme as the
cultural analogue of a gene that
spectacularly evolve in a population of
human brains (1976)
• There is a „neuronal genotype of memes”
Campbell, 1974
• A blind-variation-and-selective-retention process
is fundamental to all inductive achievements, to all
genuine increases in knowledge, to all increases in
the fit of system to environment…
• in going beyond what is already known, one
cannot but go blindly. If one can go wisely, this
indicates already achieved wisdom of some
general sort.
Changeux, 1973
• There is selection, but without the capacity
for the modification of a heuristic search
(permitted by the full natural selection
algorithm).
• This fundamentally important limitation is
admitted by the authors who write "an
organism can not learn more than is
initially present in its pre-representations."
Variation and selection in neural
development (Changeux)
• There is vast
overproduction of
synapses
• Transient redundancy is
selectively eliminated
according to functional
needs
• The statistics and the
pruning rules for the
network architecture are
under genetic control
Edelman, 1987
• Edelmans theory of neuronal group
selection proposes that a primary repertoire
of neuronal groups within the brain compete
with each other for stimulus and reward
resources.
• This results in selection of a secondary
repertoire of behaviourally proficient groups
Crick on Edelman (1989)
Crick, continued…
There are no units of evolution here!
• We propose that the algorithms of Edelman and
Changeux fundamentally consist of a population
of stochastic hill-climbers.
• Each neuronal group is randomly initialized, and
those groups that are closest to a good solution
obtain a greater quantity of synaptic resources
allowing them to ‘grow’ and/or ‘change’.
• Thus groups become strengthened but not
replicated.
A crucial limitation
• Replication has the advantage of leaving the
original solution intact, so that a non-functional
variant does not result in loss of the original
solution.
• Unless the neuronal group has the capacity to
revert to its original state given a harmful
variation, in which case it is effectively behaving
as a 1+1 Evolutionary Strategy (Beyer 2001),
there is the potential that good solutions are lost.
• We shall hazard a theory of the selection of
neuronal replicators. One may at least try
to define some of the principal factors
involved in it.
The copying problem of neuronal
connectivity
DNA
neuronal network
An elementary copying circuit
Maps and copies
• Hebbian synapse:
neurons that fire
together wire together:
CORRELATIONS
• Spike-time dependent
plasticity (STDP): if A
fires before B, then the
connection from A to
B is strengthened,
otherwise it is
weakened:
CAUSALITY
With error correction and sparse
activation
Cycles of copying
and assessment
1. The circuit to be copied exists in the lower
layer L0. The black connections in L0
show the original circuit.
2. Horizontal UP connections are activated,
e.g. by opening neuromodulatory gating.
These are the equivalent of the h-bonds in
DNA copying.
3. A copy of the topology of L0 is made in L1,
using STDP and error correction.
4. The layers are functionally separated by
closing neuromodulatory gating of the UP
connections. The fitness of each layer is
assessed independently.
5. The layer with the lowest fitness is erased or
reset, i.e. strong synaptic connections are
reduced. In the above diagram we see that
L1 fitness greater than L0 fitness, so L0
experiences weight unlearning.
6. DOWN vertical connection gates are
opened.
7. STDP in layer 0 copies the connections in
L1.
8. After DOWN connections are closed,
fitness is assessed and the cycle continues.
Why is natural selection better
than stochastic search?
• Population diversity allows parallel
exploration with resources biased on current
success.
• It allows the evolution of evolvability (cf.
sex)
Dynamical Neuronal Replicators
• The minimal unit of activity
replication consists of two
reciprocally coupled bistable
neurons (black circles), gating
by two associated inhibitory
neurons (red circles).
• 8 of these units can be chained
together to form the
experimental setup The gates
controlling the output of parent
0 for example can be switched
on and off together.
A Copying Event
The Hierarchical IF and only IF
(HIFF) problem
• A rugged fitness landscape that is difficult to
climb
Recombination
•
•
•
Instead of a one to one reciprocal
circuit, there is a many to one
reciprocal circuit. Instead of a
gating vector, there is a gating
matrix. The top diagram shows the
minimal unit from which the
system is composed. It consists of a
population size of 3, with genome
length = 1.
Gating can determine which of the
parental states are to be copied to
the offspring. The bottom figure
shows 5 of these units combined
together forming 3 parents with
genome length = 5.
Recombination is simply
undertaken by opening for example
gates, A1, B1, C1 and D2, E2.
Once the offspring has been formed
and its fitness assessed, it is
recopied back to the parent that
most resembles it.
Functions
• Structured search in rugged problems, e.g.
insight problems, working memory
• Memory consolidation from dynamical to
topological replicators
• Function passing rather than data passing
• Repair of connectivity patterns
• A neuronal basis for causal inference
Insight problems
• How to connect all the
dots with three straight
lines only?
• Sleep helps
• Lots of spontaneous
intrinsic neuronal
activity!
What is neuronal fitness?
• We know that reward systems DO exist in
the brain
• Maximize learning progress (i.e. 1st
derivative of predictability)
• Maximize mutual information between
future and past
We have just scratched the surface!
• Hypothetical mechanisms how it MIGHT work
• The brain is a breeder, „evolution” in the brain
could be more similar to artificial than natural
selection!
• Encourage more competent people to come up
with other variants of the idea
• Test the ideas against experimental evidence:
behaviour, dynamic neuronal phenomena at a high
resolution
• Select for or against the idea
GRAZIE…