the major transitions in evolution...evolution of correlation in two environments • some evolve at...
TRANSCRIPT
How can evolution learn?
Eörs Szathmáry
Biological Insitute,
Eötvös University, Budapest
Patmenides Center for the Conceptual
Foundations of Science,Pullach/Munich
Units of evolution: a tacit
‘algorithm’
Some hereditary traits affect
survival and/or fertility
1. multiplication
2. heredity
3. variability
• Electronics and
Computer Science
• University of
Southampton
Evolution as a cognitive process?
• Gregory Bateson (1904-1980)
Bayes and Darwin
Bayes and selection (e.g. Harper,
1010)
The Hebb synapse
A Hopfield network
Unsupervised learning with
Hebbian rule
Types of learning
Cycle of the lambda phage
Genetic regulatory network
Analogue Hopfield neural
network
Potential domains of learning
New activity level of a gene
Selective environment
Fitness of a phenotype is a scalar product
Selection pressures on interaction
coefficients are Hebbian
Strong selection, weak mutation
Selection for single phenotypic
patterns
Changing interaction, two target
patterns
Evolution of correlation in two
environments• Some evolve at a constant positive rate; these arise from
pairs of traits that are positively correlated in both patterns
(e.g., genes 2 and 6 are ++ in S1 and −− in S2), likewise
negative interactions evolve at a constant rate between
pairs of traits that have opposite signs in both patterns.
• When the correlation of a pair of traits in one pattern is
contradicted by the correlation of that pair in the other
pattern (e.g., s1s2>0 in S1 and s1s2<0 in S2) the
corresponding regulatory interactions (e.g., B12 and B21)
are unable to record the correlation of either target pattern
and remain near zero onaverage
Memory with two target
phenotypes
Two target phenotypes
Can the system generalize
beyond the training set?
Empirical modularity
Fixed and modularly varying
goals
Evolution of genetic triggers
MVG “modularity language”
Modularity Varying Goals: goals change
over time but share the same subgoals
Lotka-Volterra competitive model
Experiments affect carrying
capacities
Selection for mutations affecting
wij
Caveat! Limiting assumptions
• Wij = Wji symmetry
normalization
Mechanistic equivalence between
eco-evo and learning
The flip side of the coin:
evolution IN cognition?
Hebb and Darwin
(Adams, 1998)
synaptic replication synaptic mutation
The most exciting hint from
neurobiology: structural plasticity
What could be the algorithmic advantages?
Candidate mechanisms of
“neuronal replication”
• Local connectivity copying
• Copying of activity patterns in bistable
neurons
• Path evolution
• Other?
A recurrent attractor network
Population of networks: selection
Evolution
Thanks for the invitation!
Thanks for your attention!
Paths as Units of Evolution
Mutation and Crossover of Paths
The interplay of Hebb and Darwin
DNA replication Neuronal copying
local influence non-local influence