Greg Detre
1/3/01
Cliff, in psychological AI Box (University of Sussex)
evolving robots, Neuroethological manifesto
Harvey co-author
description of GAs and genes to code NNs
genes
number of neurons
number of connections/outputs
proportion that loop
standard deviation of distance
weighting � SD
mutation of genes???
rate of synaptic decay
Hebbian learning function/weight adjusting
rate of formation of new connections
proportion of inhibitory/excitatory connections
symmetricity
growth process
rule-following vagueness
proportion of input neurons (environment variable � non-gen)
how specify position absolutely with genes???
shape of body
interaction between � higher level
how do you get genes to affect different areas of the network
split genes (mitosis) like a stretched jumbo jet
variable number of genes???
Karl Sims � genetic graphs
evolution to the power of evolution???
how change the number of inputs to a NN???
Julian Jaynes
consciousness emerges out of social hierarchy??? that you hear voices of even when you�re away from it???
cultures �/span> different types of consciousness???
has to be recently in history
why/how worldwide so fast??? not evolutionary???
e.g. children nowadays rewiring visually from TV and computers
if 1 person revives, forces their friends too
PROLOG � inductive learning instead of controller NN
mathematical patterns in Evoids???
localised neurotransmitters
specified by controller NN??? by genes
big NNs as comprising interconnected smaller NNs
isn't there a sort of fitness function operating for these smaller NNs to survive???
have we evolved the critical window for language learning � yes
children in India � surrounded by lots of languages
(depends how acoustically similar they are)
absolute pitch � babies have it, lose it at age 2 or 3
learn Chinese??? use it to learn the language sounds
how expand evolutionary space, because programmer has to specify the structure of the genes???
is our evolutionary space limited??? our genes only code for types of proteins � cannot build titanium cells, for instance???
can our genes act on themselves � y
but they can only really operate on the genetic mechanisms within a saltatory theory???
Hofstadter � gene viruses with different/incompatible code � we�re safe
need to change loads of processes, just to change the chemical in 1 base
requires a huge, simultaneous/circular jump
so we�re stuck with the transcription mechanism, from billions of years
what if the controller was to alter the genotype/phenotype relationship???
genes can either be a recipe or a map
what if the genes can code for the growth process itself???
need to actually grow each NN from scratch???
that growth is modified by/also affected input
how different are people�s brains???
how would you measure that???
grow from >1 neuron at once???
like an embryo � grow a bit, then split + develop separately
in a growing NN, how do the neurons find each other
pre-determined paths
if a connection dies, the two neurons move apart
neurons sense other nearby neurons and connect probabilistically
what problems should I feed it???
something you can calculate a fitness for
stimulus categorisation
coloured nouns, parts of speech, plurals, prime numbers, English past tense, doubles, divisible by 4 or 3, addition, face recognition
translational shifts, place coding
generalisation � or just association
represent words � word/N, normalised spelling letter/N, syllables binary number/N
how represent the number � rate of firing, finite APs
XOR problem
reward system??? +ve neuromodulator
output as < or > than specified rate