►
Description
DevoWorm (2021, Meeting 4): ResNet training with the EPIC dataset and pre-trained model improvement, "Killing the Winner" and "Survival of the Systems", comparing trees with metrics, a billion years of biogeography, and comparisons between evolutionary algorithms and biological evolution. Attendees: Susan Crawford-Young, Bradly Alicea, Krishna Katyal, Mainak Deb, Jesse Parent, Richard Gordon, and Shruti Raj Vansh Singh
B
A
So
yeah,
I
don't
know
susan
if
you
want
to
just
type
into
the
chat,
if
you
you
know
when
you
have
something
to
say,
I
know
your
audio
isn't
working.
A
So
yeah
susan's
audio
is
off
and
there
are
videos
on,
but
our
audio
is
off
and
vice
versa,
apparently
so
welcome
to
the
meeting.
I
don't
know
who
else
is
going
to
show
up?
Usually
people
kind
of
filter
in,
but
we
have
a
bunch
of
stuff
to
talk
about
today.
My
knock
has
some
updates
on
the
diva
learn.
So
actually
he
had
some
things
on
the
devil
learned
software.
He
was
updating
and
I
know
that
mayoq
has
released
a
new
version.
I
mentioned
it
in
the
email
we
have
0.2.1.
A
So
that's
a
new
version.
There
were
some
updates
that
I
think
my
knock
made
and
there
were
a
couple
other
updates
that
he
wanted
to
put
out
a
release,
and
I
told
him
yeah.
If
you
want
to
put
out
a
release,
we
don't
really
have
a
formal
release
schedule,
but
that's
what
so
we
have
one
and
my
what
my
not
my
knock
will
give
us
an
update
and
maybe
we'll
do
another
really
soon.
A
Well,
let
me
go
through
what
we're
going
to
do
today.
First,
yeah,
okay,
so
yeah
yeah,
I
don't
know
if
christian
is
gonna
show
up,
but
he
has
his
evolution
abstract.
A
Open
science
framework
conference
next
week,
so
it's
next
tuesday,
I'm
gonna
go
through
some
of
those
slides
and
then
I'm
gonna
we're
gonna
try
to
work
on
them
this
week
and
then
some
other
opportunities
darwin
day
is
coming
up
and
then
hi
dick
good
morning
morning.
A
We
have,
then
we
have
some
papers
and
things
like
that.
So
what
my
neck,
why
don't
you
present
your
screen?.
C
I
I
unplug
my
visual
and
I
get
audio.
A
B
B
B
Which
basically
translates
everything
into
frames
and
also
creates
a
single,
unified,
annotation
file
so
which
makes
it
easier
to
train,
but
the
downside
was
that
he
was
not
using
all
of
the
260
videos.
He
was
using
six
or
seven,
if
I
remember
correctly,
but
I
don't
think
that
was
the
right
thing
to
do
so.
What
I
decided.
B
B
B
B
So
moving
on
okay,
so
this
is
basically
this
basically
an
example
that
shows
the
comparison
between
using
less
data
and
more
data.
So
to
the
left.
We
see
these
blue
data
points
right.
The
blue
ones
are
the
data
points
and
the
red
line
is
the
model
prediction
so
to
the
left.
This
shows
a
diagram
that
basically
represents
what
happens
if
if
the
money
is
stubbed.
B
Data
we
increase
the
number
of
data
points.
We
successfully
capture
the
true
distribution
of
the
data
which
is
shown
to
the
right,
like
we
have
considerably
increased
the
number
of
blue
points,
which
is
the
data
points.
Hence
the
model
is
able
to
more
generally
classify
the
objects
so
moving
on,
so
we
all
saw
these
results
right,
so
I
actually
found
something
really
similar
going
on
when
I
trained
it
on
the
lineage
population
prediction
model.
B
So
what
I
basically
did
is
that
I
trained
two
neural
networks
with
two
with
two
different
data
sets
of
different
sizes
with
the
same
hyper
parameters
like
the
learning
rate,
the
like
the
architecture,
learning
rate
the
bad
size,
etc.
So
the
dotted
line
here.
B
C
B
So
the
architecture
which
I
used
in
both
these
cases
is
the
resonant
50,
which
in
case
you
were,
is
the
resonance
50
so
which
is
not
same
of
the
model.
So
this
can't
be
really
compared
with
the
reasons
from
the
current
version
of
debugger,
because
a
developer
uses
resonate
18,
which
is
smaller
so
anyways.
I
use
the
same
architecture
and
the
same
hybrid
parameters.
B
So,
if
there's
a
clear
indication
that
using
more
data
points
is
helping
like
like
in
the
in
the
image
below,
we
see
that
the
dotted
line
is
much
much
closer
to
the
blue
non-dotted
line,
which
clearly
indicates
that
using
more
data
points
is
helping
like
it's
just
like
this
one,
this
image,
that
is
sure
it's
basically
a
real
life
example
on
the
epic
data
set
of
in
this
game.
Yes,
so
that's
the
result.
I
got
yeah
so
as.
B
It
took
about
2.5
hours
to
train
the
last
model
that
I
had
shown
this
one
so
like
I
made
contact
with
because
he
has
a
he
has
a
better
gpu.
So
I
just
send
him
the
bigger
data
sets
and
he'll
just
train
it
for
me.
So
that's
something
that
I
plan
to
do
soon
and
yeah.
Okay.
So
that's
basically
it
from
the.
A
Yes,
oh,
could
you
put
it
share
your
screen
again,
I
wanted
to
go
back
over
some
of
the
slides
just
to
talk
about
them
a
little
bit.
B
A
Well,
let's
see:
go
back
I'll,
tell
you
when
to
stop.
A
Right
yeah,
yeah
yeah:
I
can
see
it
just
go
back
in
the
slide
deck.
So
this
one
here
I
guess
it's
slide,
four
yeah
yeah
or
a
slide
five.
We
can
start
with
that
yeah.
So
this
is
the
one
that
mayak
generated.
I
think
in
when
he
was
working
on
this,
and
I
think
it's
probably
one
of
the
notebooks.
B
Had
a
few
demos
that
had
graphs
similar
to
this,
so
I
had
refactored
some
code
to
create
a
plot
that
is
similar
to
this,
but
these
are
like
on
a
different.
These
are
on
a
different
network.
We
a
different
architecture
and
it
has
different
hyper
parameters.
So
yeah.
A
A
Line
so
you're
doing
like
you're
doing
a
predicted
versus
observed
so
you're,
observing
it
and
then
you're
predicting
using
the
model
you're
predicting
it
right.
A
B
B
B
A
Yeah,
I
remember
when
was
working
on
it
last
summer
there
was
a
lot
of
because
it's
not
fully
aligned
they're
just
collecting.
You
know
microscopy
data
they're,
just
letting
it
run
across
different
embryos
and
they're
doing
the
tracking.
So
it
doesn't
align
perfectly,
and
you
can't
really
do
that.
Anyways,
just
kind
of
like
yeah.
B
A
One:
okay,
all
right
that
one
yeah!
So
that's
the
this
is
the
blue,
dots
or
data
points.
Red
lines
are
model
predictions.
So
it's
basically
like
the
blue
lines,
follow
the
red
lines,
so
it's
kind
of
like
so
we're
increasing
the
number
of
data
points
and
we're
capturing
the
model
prediction.
So
the
model
prediction
is
like
a
distribution
based
or
maybe
not,
distribution
based,
but
it
basically
simulates
the
distribution
and
you
get
like
this
prediction
and
then
the
blue
lines
match
that
eventually.
A
D
A
Yeah,
I
think
that's
a
good
lesson
so
yeah,
I
think
that's
well,
so
that's
good
and
so
go
back
to
slide
two
for
a
minute,
okay,
yeah,
so
yeah.
This
is
the
this
is
the
this.
Is
the
a
nice
sort
of
summary
of
what
we
have
here
with
this
data
set,
so
this
is
collected
on
this
group.
They
basically
put
a
fluorescent
marker
in
the
embryos
and
they
tracked
all
the
embryos
across.
A
I
don't
know
how
many
worms
like
you
know,
hundreds
I
thought,
maybe
a
thousand
or
so,
and
they
were
able
to
get
all
these
data,
and
you
know
they've
basically
made
all
these
videos
available
to
people.
A
Raw
forms
that
needs
to
be
processed
but
to
have
this
like
brought
together
as
a
annotated
file,
is
really
good
for
future
exploration
because
they
I
mean
they
have
like
a
version
released
where
they've,
like
you
know,
put
the
these
dots
they've
tracked
and
put
the
numbers
in
a
file.
But
the
thing
the
advantage
of
going
back
to
the
images
is.
You
can
extract
other
types
of
information
so
like
with
the
with
just
the
dot
with
the
locations.
As
you
know,
points
and.
F
A
A
The
advantage
to
this
is,
you
can
extract,
you
know,
like
you
know,
maybe
their
extraction
method
wasn't
complete
or
maybe
there's
some
features
across
regions
of
the
embryo
that
we
can
recognize
with
this.
So
I
mean
this
is
a
good
yeah.
This
is
a
good
upgrade.
A
Now
my
oak
isn't
here
I
don't
yeah,
so
we
can't
really
ask
him.
Well,
you
know
what
you
can
do
is
just
put
this
as
a
pull
request
and
then
try
to
see
if
he
can
help
you
with
some
processing
on
this.
I
think
the
having
processing
powers
is
going
to
be
an
issue
yeah
exactly
exactly.
A
So
yeah
so
yeah.
Let's
keep
us
posted
on
this.
This
is
good
work
yeah
and
I,
like
the
idea
you've
been
able
to
expand
the
number
of
data
points.
I
remember
in
the
gsoc
period
you
know
they're,
like
constraints
on
what
you
can
do
in
terms
of
time
and
and
you
know
getting
up
a
perfect
concept.
C
A
G
I'm
trying
to
be
a
little
more
explicit,
in
other
words,
if
you,
if
you
pretended,
even
if
the
rules
were
wrong,
that
you
simulated
an
embryo
and
got
a
whole
bunch
of
cells
with
a
differentiation
tree,
can
you
how
far
can
you
recover
that
tree
from
you
know?
In
other
words,
if
you
put
a
blob
at
each
point
in
space
each
time
you
got
a
new
cell
yeah,
can
we
cover
the
coordinates
of
that
blob
or
angles
or
whatever.
A
Yeah
I
mean
if
we
had
like
a
data
set,
we
could
like
process
it
further
and
put
it
in
a
normalized
space.
Yeah.
G
I
mean,
for
example,
you
can
make
a
simulation
and
say:
okay,
there's
going
to
be
a
physical
constraint,
which
is
the
size
of
the
embryo.
Okay,
within
this,
each
cell
is
going
to
divide
according
to
some
program
and
when
it
does
so,
it
has
to
stay
inside
that
boundary.
G
G
D
A
Yeah,
I
think
that
that's
a
good
idea,
thank
you,
sure,
yeah,
so
yeah,
that's
good
yeah,
so
we
we
can
do
yeah.
We
can
do
something
like
that
as
a
test,
because
I
know
like
a
lot
of
these
algorithms,
you
get
like
performance
metrics
for
the
data
set
itself,
yeah.
C
A
E
H
G
Suppose
that
you
knew
exactly
where
each
cell
was
at
each
time:
okay,
okay,
yeah,
yeah,
yeah;
okay,
how?
How
accurately
can
you
recover
that
information
and
what
I'm
suggesting
is
making
a
simulation
of
an
embryo
would
be
a
very
clean
simulation
that
will
include
precisely
known,
coordinates
and
angles
and.
C
H
H
An
early
stopping
system
yeah,
that's
possible
that
we,
you
know
generally
don't
train
data
to
a
certain
limit.
We,
you
know
only
stop
it
because
there
isn't
trade-off
between
the
computation
cost
and
performance.
If
you
can
say
spending
100
bucks
or
on.
G
G
G
H
Yeah,
my
grandfather,
my
grandfather
did
his
phd
from.
I
guess
he's
he's
87
now
and
he
also
tells
the
same
story.
Okay,
and
he
always
tells
me
that
if
these
things
would
have
been
in
his
time,
he
would
have
done
something
really
great.
F
H
H
H
H
A
H
B
H
The
other
temperature
climate
conditions
are
more
favorable
so
which
helps
in
its
survival.
Does
it
really
wins
or
does
the
winner
get
killed
here?
It
is
goes
that,
for
example,
we
have
two
species
and
they
are
competing
with
each
other.
H
So,
due
to
some
reason,
if
the
population
of
dogs
is
rising,
then
it
will
equate
that
the
population
of
cats
would
decline,
but
this
this
isn't
concerned
that
if
dogs
are
kept
getting,
you
know
greater
in
population
that
would
aid
cats
because
their
play
is
also
getting
bigger.
So
if
the
two
species
you
know
are
you
know
dependent
on
each
other
for
their
survival.
H
Even
better
now
that
is
locked
ultra
equation.
So
it's
you
know
a
relationship
between
relationship
between
prey
and
the
radiator,
so
I
get
in
there
was
in
fox,
which
was
used
for
and
that
fox
used
to
play
on
a
certain
kind
of
rabbit.
So
when
the
when
the
population
of.
H
Rat
decreases
and
so
was
saying
that
when
the
population
of
fox
was
declining
to
a
certain
extent,
the
population
of
the
attack
increased,
but
then
it's
starting
decreasing
rapidly.
So
there
was
a
relationship.
F
H
So
here
it
is
here
you
can
see
that
if
the
various
species
are
completing
and
they
have
their
own
predator,
if
one
species
flourish
too
far
above
from
this,
then
the
predator
also
thrives
and
eats
the
winner.
So.
H
Show
that
the
combination
of
competition
and
radiation
eventually
made
all
the
species
go
extinct,
so
how
this
theory
that
you
know
just
made
perfect
sense
is
not
practically
probably.
H
H
It's
getting
more
prosperous,
then
it's
the
species
also,
you
know
modified
itself
and
you
know
it's
evolved,
but
so
does
it's.
You
can
say,
explain
it.
For
example,
if
the
cats
are
becoming.
A
That's
good,
so
this
is
the
krishna
is
trying
to
come
up
with
a
proposal
for
the
evolution
conference,
and
this
is
the
proposal.
Last
week
we
talked
about.
Maybe
how
you
would
simulate
this
to
show.
A
I
mean
there
are
a
number
of
routes.
You
could
take
your
different
software
programs,
agent-based
models
or
evolution
simulations.
You
could
even
do
like
a
numeric
simulation.
If
you
wanted
people,
you
know
if
you
wanted
to
model
like
the
lock
of
altera
equations.
G
G
This
was
first
discovered
on
using
with
deer
that
got
isolated
on
small
island.
They
ate
everything
on
the
island.
The
population
went
up
to
thousands,
then
it
crashed
through
just
a
few
individuals.
G
Okay,
the
the
other
thing
on
the
mutations
that
strikes
me
is
that
it
covers,
what's
called
the
red
queen
hypothesis,
how
many
of
you
have
read
that
allison
also
wondering.
G
Okay
and
alice
in
wonderland,
the
red
queen,
shows
alex
how
she
has
to
run
as
fast
as
possible
just
to
stay
where
she
is
okay,
and
this
has
been
turned
into
a
hypothesis
of
how
evolution
goes
on.
The
the
predator,
predator
and
clay
are
both
evolving.
They
get
better
and
better
at
running
away
from,
but
I
put
some
references
on
both
in
the
chat.
D
A
Yeah
they're
over
in
the
chat
yeah,
so
there's
some
cenarvon
svinson,
pg,
huggaway
maynard
smith,
yeah,
so
they're,
some
pretty
good
citations
in
there
yeah.
So.
G
A
A
So
yeah
yeah,
why
don't
you
do
some
literature
search
and
then
I
think
about
how
you
might
simulate
this
as
a
process?
A
I
Yeah,
I'm
not
agreeing
with
simulations
for
both
of
these,
but
there
may
be
some
fine.
A
Yeah
yeah,
I
mean
it
they're
not
going
to
be
all
simulations,
but
maybe
a
couple
of
them
have
simulations
in
them,
but
you
know
you'll
get
a
sense
of
the
problem,
though,
by
going
through
these.
A
Okay,
anyone
else
have
any
comments
about
this
talk.
A
A
So
if
you,
you
know
just
be
mindful
of
deadlines,
you
know
month
is
not
a
long
time,
but
you
have
to
plan
it
out.
So.
A
Okay,
so
that's
good
anyone
else
have
anything
else.
You
want
to.
A
G
Yeah
yeah,
the
problem
is
with
very
very
few
differentiation
trees,
but
we
could
generate
some
random
ones
and
ask
about
similarities
stuff
like
that
right.
G
The
question
is,
you
know
what
is
the
similarity
measure
between
two
trees
and
do
we
weight
lower
branches
higher
than
more
than
higher
branches,
for
example,
things
like
that
yeah.
A
So
yeah
we
we've
done
so
well.
We
actually
a
couple
years
ago.
We
did
some
work
using
the
hamming
distance,
which
is
an
information
theoretic
distance.
So
the
the
idea
is
you're,
comparing
two
binary
trees
and
you
want
to
know
like
if
you
you
have
like
each
node
is
identified
with
an
id.
So
you
want
to
know
if
you
have
two
trees
and
they're
generated
different,
you
know
using
maybe
a
different
process
or
something
you
know
what
the
sort
of
match
is
between
them
so
like.
A
If
you
have
differences
in
the
nodes,
you
know
how
they
rearrange
themselves,
and
so
it's
actually,
you
know
you
can
attach
the
pr
attack.
The
problem
in
different
ways
so
about
four
years
ago,
came
up
with
a
method
for
doing
this,
using
a
simple
information,
theoretic
metric,
but
there's
no
way
really
to
say
anything
more
than
just
the
two
trees
are
different
at
this
amount
of
distance.
A
So
that's
not
really
ideal.
Although
it's
you
know
it's
a
good
start.
Is
there
anyone
saying
that
they're
similar
up.
A
Yes,
actually
so
in
in
biogenetics,
which
is
the
study
of
evolutionary
trees.
So
this
is
where
you
have
you
know,
common
ancestors
and
then
speciation
events,
and
you
have
these
trees
that
are
generated
over
long
periods
of
time,
but
they're
inferred,
using
like
common
ancestors.
A
Let
me
see
if
I
can
illustrate
this,
I'm
going
to
share
my
screen,
because
this
is
this
is
actually
this
is
made
for.
So
I'm
going
to
use
a
tablet
here
to
see
if
it's
working
today,
okay
good,
so
you
have
a
binary
tree
right
and
let's
trace
this
and
start
over.
A
A
Okay,
so
our
problem
is,
we
want
a
map.
We
have
an
id
for
each
node,
so
we
might
give
it
a
number
like
you
know.
This
is
one
two
three
four
and
then
this
is
five
through
eight,
and
so
the
idea
is
that
if
we
had
another
tree
where,
for
example,
these
things
flipped.
So
this
were
three
and
this
were
four
then
we'd
want
to
know
the
distance
between
that
tree.
But
if
we
do
the
single
distance
measure,
you
know
we
can't
differentiate
between
the
layers.
A
You
get
these
trees
where
you
get
the
same
structure,
but
the
differences
are
characterized
by
characters.
So
you
usually
put
marks
of
characters
on
the
branches
like
this,
so
you
have
a
common
ancestor
here
right
and
then
there's
a
speciation
event,
so
you
have
two
different
descendants
and
the
descendants
are
different
by
a
certain
amount.
So
this
common
ancestor
changes
in
this
daughter
species
by
two
characters
and
this
other
daughter
species
by
three
characters.
A
And
so
then
you
can
measure
this
distance
across
the
tree
and
the
same
holds
for
these
two
branches
here,
so
you
can
actually
characterize
and
these
these
characters.
You
know
they
could
be
dna
sequences,
which
is
rather
objective
or
you
could
use
like
some
phenotypic
character
like
eye
shape,
which
is
more
subjective
because
you
have
to
you
know,
figure
out
what
the
coding
is
for
the
shape
and
everything.
A
But
in
any
case
you
get
these
differences
in
terms
of
the
number
of
characters.
Difference
between.
A
A
A
Yeah,
so
we
could,
we
could
do
that
we
could
map.
We
could
map
that
to
a
different
problem
domain.
I
hadn't
done
this
in
a
while
voice.
I
was
kind
of
fascinated
by
these
at
one
time,
like
the
the
idea
of
using
you
know,
sort
of
the
similarity
between
like
claudistics
phylogenetic
trees
and
binary
trees,
and
I
never
really
got
you
know
much
going
on
that,
but
that
that
was
something
a
long
time
ago,
but
yeah
you
can
take
these
types
of
trees
and
map
them
to
other
domains.
A
A
Imagine
if
you
had
two
trees
where
you
had
these
differences,
and
this
will
often
happen
with
like
if
you
have
a
gene
tree
versus
a
tree
of
of
phenotypic
traits
where
the
each
source
of
data
gives
you
a
different
answer
or
you
might
search
through
a
search
space
for
an
optimal
tree,
and
you
have
a
number
of
candidates
that
are
more
or
less
the
same
optimality,
but
not
really
the
same
thing.
So
you
get
this.
A
You
know
you're
going
to
have
differences
in
the
number
of
characters
on
each
branch,
so
you
can
actually
compare
those
two
and
the
method
they
usually
use
in
phylogenetics
is
consensus,
which
is
just
like
an
average,
but
there
are
other
ways
to
do
this.
Of
course
you
could
characterize
you.
Could
you
know,
count
the
number
of
characters
on
a
branch
and
like
find
a
difference
or
find
an
average
or
you
know,
do
it
for
each
section
of
the
tree
or
map
it
to
some
other
type
of
tree.
A
That's
similar
in
structure
so
yeah
I
mean
there
are
a
lot
of
ways
to
do
this.
I
I
mean
I
don't
have
an
answer
but
yeah.
I
think
that
that
would
be
something
we
could
explore
in
further
detail.
I
think
in
in
in
future,
in
the
future.
I
think
there's
a
lot
of
potential
application
for
it.
I
think
deck
is
looking
primarily
at
like
the
well.
We
have
the
issue
on
complexity
measures
and
it's
sort
of
a
thing
that
you
know
we
haven't
really
talked
about.
A
We've
talked
about
it
on
and
off,
but
we
haven't
really
done
any
heavy-duty
work
on
it.
So
we
have
some
good
references
here:
meeks
and
scott
spanning
trees
and
the
complexity
of
flood
filling
games.
It's
a
theoretical
computer
science
paper
and
the
complexity
of
networks
to
the
set
complexity
of
edge
colored
graphs.
A
So
there
yeah
there
are
a
number
of
problems,
because
binary
trees
are
really
just
partitions
of
space
and
partitions
of
other
types
of
things.
So
I
mean
we
can
like
look
at
it
that
way
too.
We
can
look
at
it
as
like.
A
well
like
kd
trees
are
a
partition
of
a
space
where
you're
dividing
a
space,
and
you
know
into
halves,
recursively
and
then
you're.
The
tree
is
a
record
of
that
those
subdivisions.
A
A
But
yeah
I
mean
there's
a
lot
of
there's
a
lot
of
ground
to
explore
here.
I
don't
really
know.
I
H
It's
quite
off
the
topic,
but
I
once
read
an
article
and
it's
described
the
human
hierarchy
with
the
help
of
phylogenetic
trees
and
that
you
know
date
back
to
the
when
we
and
lobsters
were
having
the
same
ancestors
right.
I
can
send
them.
It's
not
exactly.
You
know
very
biological,
it's
much
more
into
psychology,
but
yeah.
H
It's
you
know
how
you
can
one
can
use
phylogenetic
trees
and
how
our
sun,
after
strengthening,
secretion
and
lobsters
threatening
situation
you
know,
was
linked
due
to
some
common
stress
in
ancestor
and
how
that
it
explained
human
psychology
and
human
hierarchy.
I
G
Okay
yeah
list
of
those
in
jack,
okay,.
B
A
Says
yes,
can't
talk
much
right
now,
complexity,
so
jesse's
interested,
but
he
can't
talk
by
speech
right
now.
So
that's
that's!
Okay.
We
can
follow
up
on
it
later
so
yeah.
Thank
you
we'll
we'll
look
into
that
further.
I'm
gonna
make
a
list
of
these
references
and
we'll
probably
talk
about
it
next
week.
A
More
maybe
this
week
we'll
explore
a
little
bit
but
it'll
be
well.
That's
this
direction,
we'll
think
about
in
this
direction.
So
actually
I
can
also
send
out
the
paper
that
we
did
in
2016
on
this.
It's
not
like
again,
it's
just
a
small
proof
of
concept,
but
it
gives
you
an
idea
of
what
what
the
problem
is
in
terms
of
like
looking
at
comparing
trees.
A
So
let
me
go
over
some
other
things
here,
the
last
maybe
30
minutes.
If
you
have
to
leave
it
the
top
of
the
hour.
That's
fine!
I
just
wanted
to
go
through
some
things
here.
So
the
first
thing
I
wanted
to
point
out
for
people
if
you're
interested
is
that
lewis
wolpert
died.
This
last.
C
A
He's
a
famous
developmental
biologist-
and
he
you
know
he
did
a
lot
of
had
a
lot
of
landmark
publications
in
the
literature,
and
I
found
this
web
of
stories
site
where
he
talks
about
some
of
his
stories
here
like
where
he
was
doing.
I
think
he
studied
civil
engineering
at
one
time.
A
Then
he
had
a
wild
youth
and
he
switched
to
cell
biology
and
he
did
some
work
on
the
fluid
membrane
model.
He's
known
for
the
positional
information
hypothesis.
I
A
Involves
like
the
position
of
cells
and
development,
and
so
you
know
it's
it's
a
pretty
key
concept
in
embryogenesis.
So
if
you
want
to
read
that,
if
you
want
to
look
over
the
web
of
stories,
I
put
it
in
the
chat
and
you
can
also
look
up
some
of
his
papers.
I
mean
lewis
will
search
for
that
and
your
favorite.
B
A
Tool
and
you'll
find
it
another
thing:
that's
maybe
coming
up
in
the
next
couple
weeks
is
every
year
on
this
blog
that
I
do.
I
do
a
a
post
for
darwin
day.
It's
darwin
is
february
12th
and
it's
just
a
day
where
people,
usually
you
know,
put
together
things:
social
media
posts
and
things
on
darwin,
you
know,
allows
people
to
think
about.
You
know,
evolution
and
promote
evolution
in
that,
so
this
year's
darwin
day
we
need
a
post
for
it,
and
I
have
some
ideas,
but
if
you're.
A
On
an
idea
for
this
I'd
be
open
to
hearing
about
it.
Last
year
I
did
a
post
on
speary
darwin
and
the
evolution
of
reference
frames,
so
this
goes
back
to
the
neuroscientist
roger
sperry
and
he
did
these
experiments
with
where
they
took
a
toad
in
development
and
they
took
the
eye
and
they
cut
the
eye
out
and
they
rotated
it
90
degrees
or
180
degrees.
So
they.
C
A
Where
it
was
and
then
they
let
the
toad
sort
of
behave,
and
the
idea
here
is
that
the
toad
when
the
eye
is
aligned
as
it
is
when
it
develops
normally,
it
can
target
insects
and
catch
them
with
its
tongue.
But
when
you
do
this
surgery
to
the
eye
and
you
move
the
eye
around,
it
actually
flicks
its
tongue
in
the
opposite
direction
from
the
fly.
F
A
Suggesting
that
there's
like
this
sort
of
automated
response
in
the
brain
that's
going
on,
and
but
it's
that,
because
you
shift
that
spatial
reference
frame,
it's
actually
not
registering
the
prey.
It's
doing
exactly
what
it's
supposed
to
be
doing.
But
it's
just
not
registering
the
prey
in
the
right
place
and.
C
A
Is
an
interesting,
it's
called
the
chemosensory
hypothesis
and
kind
of
goes
through.
I
kind
of
walk
through
that
experiment
and
then
how
this
might
change
over
in
evolution
talking
a
little
bit
about
scaling
laws
and
other
things
it's
it's
not.
It
doesn't
come
up
with
a
great
insight.
It's
just
you
know
talking
about
these
things
that
are
related,
and
so
then
I
have
some
references
here
and
that's
I
mean
that's
basically
what
I
would
be
looking
for
in
a
blog
post.
So
if
you
have
ideas
about
that,
that's
fine!
C
A
A
Okay
and
then
yeah,
so
then
that's
okay,
that's
good!
So
if
you're
interested
in
this
and
doing
a
blog
post,
let
me
know
you
know
I'll
be
probably
doing
it.
If
no
one
has
any
ideas,
that's
fine!
We
might
do
something
on
like
lewis,
wolford's
work,
it's
possible
so
yeah.
So
then
I
didn't
get
really
get
into
this
flash
talk.
This
is
a
talk
that
I'm
giving
on
behalf
of
the
group
next
week,
next
tuesday.
A
So
maybe
next
monday
I'll
have
like
a
full
dry
run
of
this
presentation,
but
this
is
the
it's
a
five-minute
flash
talk.
It's
to
this
group.
It's
open
opens
science
framework,
because
it's
an
organization
that
advocates
open
science.
They
run
the
open
science
framework
repository
and
they
have
a
conference
on
open
source
education.
A
So
this
is
a
talk
to
open
source
educators
and
this
is
a
supposed
to
be
a
five
minute
flash
talk,
which
means
it
just
has
to
have
a
lot
of
images
and
you
have
to
go
through
the
slides
quickly,
but
this
is
going
to
be
on
the
diva
learn
platform,
so
I've
got
a
bunch
of
slides
here.
A
I
think
I
have
enough
slides.
The
question
is
how
to
organize
this
so
that
it's
effective,
so
I've
got
a
couple
of
you
know.
I
have
like
the
open
source
contributors
that
we've
had
on
the
platform,
so
we've
had
about
14
contributors
over
time.
A
lot
of
those
were
during
hacktoberfest,
but
some
of
them
were
just
you
know.
Like
people
like
people
come
to
the
meetings,
krishna
jesse
and
you
know,
they've
been
contributing
on
some
of
the
other
parts.
A
A
I
might
modify
this
a
bit,
but
this
is
a
model
of
the
diva
learn
program.
I
might
actually
update
that
to
reflect
maybe
the
whole
platform,
but
this
is
actually
a
schematic
of
the
whole
platform,
so
this
is
actually
now
0.2.1
and
then
this
is
this.
So
this
is
the
divo
learn
standalone
software,
and
then
these
are
the
two
other
major
things,
although
I
might
add
in
the
data
science
tutorials
as
well.
A
This
is
the
species
specific
models
in
the
devo
zoo,
so
you
know
there's
this
umbrella
of
things
that
we
have
for
education
and
for
research
and
then
yeah.
So
I
have
a
lot
of
visuals.
I
don't
really
have
it
organized
yet
I'll
have
to
do
that
this
week,
but
I
think
it's
coming
along.
I
don't
think
you
know.
I
think
we
have
a
pretty
good.
I
should
get
some
good
feedback
here,
hopefully,
and
then
that
will
help
us
promote
the
the
organization
as
well
as
get
some
ideas
for
for
future
development.
A
So
I
mean
that's
it's
something
that
yeah
we'll
see
how
that
goes.
So
then.
Finally,
I
have
some
other
things
here.
I
wanted
to
finish
up
on.
I
have
this
really
nice
tectonic
drift
simulation,
so
this
is
a
simulation
that
was
just
released.
A
It's
of
the
continents
as
they
sort
of
formed
as
creighton's
about
a
billion
or
so
years
ago,
and
the
as
the
plates
on
the
earth
surface
have
shifted.
These
creightons
have
come
together
to
form
continents
and
then
form
the
continents
that
we
know
today.
So
you
can
kind
of
see
that
things
are
starting
to
become
recognizable
as
a
map
of
the
glow
of
the
globe.
As
we
know
it
today
that
looks
like
a
bunch
of
ice
cubes
moving
around
on
the
surface
of
a
lake,
or
you
know.
A
Today,
in
their
relative
positions,
so
that's
from
this
paper
here
extending
full
plate
tectonic.
I
A
F
E
A
Like
the
continents
are,
as
as
it
were,
you
know
really
weren't
there
in
the
same
way
that
they
are
today.
So
you
know
we
kind
of
know
from
the
simulation
that
there
were
a
bunch
of
pieces
of
land
scattered
around
the
globe
and
that
you
know
it's
much
different
than
thinking
about
like
the
world
today,
the
biogeography
of
the
world.
Today,
where
you
have
continents-
and
you
have
you-
know
the
oceans
where
they
are
see,
this
is
the
way
the
so
yeah
this
is
so
this
might
have
some
relevance
to
the
boring
billion.
A
Into
deep
time
you
can
read
more
in
this
earth
science
reviews
paper.
It's
brand
new.
A
A
A
A
Yeah,
so
that's
yeah!
So
that's
interesting.
Let
me
go
back
up
to
the
main
folder
here,
okay,
so
the
next
thing
here
is
well
there's
this.
Let's
see
this
paper
here
called
survival.
The
systems-
and
this
is
a
new
paper
in
trends
in
ecology
and
evolution,
and
I
found
this
interesting.
The
abstract
says
since
darwin,
individual
individuals
and
more
recently
genes
have
been
the
focus
of
evolutionary
thinking.
A
The
idea
that
selection
operates
and
non-reproducing
higher
level
systems,
including
ecosystems
or
societies,
has
been
met
with
skepticism,
but
research
emphasizing
natural
selection
can
be
based
solely
on
differential
persistence,
invites
reconsideration
of
their
evolution.
Self-Perpetuating
feedback
cycles
involving
biotic
as
well
as
abiotic
components,
are
critical
to
determining
persistence.
A
The
evolution
of
autocatalytic
networks
of
molecules
is
well
studied,
but
the
principles
hold
for
any
self-perpetuating
system.
So
what
they're
getting
into
here
is
they're
thinking
about
systems
at
a
very
broad
level
and
they're
thinking
about
self-perpetuating
feedback
cycles
and
they're
thinking
about
auto
catalytic
networks
of
molecules,
so
they're
kind
of
thinking
about
the
sort
of
self-surviving
aspect
of
a
system,
but
at
a
very
general
level.
A
So
you
know
they
talk
about
ecosystems
and
societal.
Examples
and
persistence-based
selection
of
feedbacks
can
help
us
understand
how
ecological
and
societal
systems
survive
or
fail
in
a
changing
world.
So
you
have
systems
that
are
very
general
that
they're
talking
about
they're
interested
in
what
makes
them
persist
over
time.
You
know
because
you
think
about
like
we
think
about.
Maybe
a
a
system
of
feedbacks
or-
and
I
know
jesse
is
interested
in
the
cybernetic
systems,
where
you
have
these
perpetuating
feedbacks
and
other
types
of
regulation.
A
But
you
know
some
of
these
systems
have
been
in
place
for
millions
and
millions
of
years
and
if
you're
talking
about
like
ecosystems
like
we
just
saw
in
the
simulation
of
the
different
pieces
of
land
and
moving
around,
you
know
those
systems
can
survive
for
a
while
and
eventually
they
change,
but
what
allows
them
to
survive
for
long
periods
of
time.
What
what
allows
them?
Because
they're
not
static,
they
change
a
lot
internally,
but
why
do
they
stay
in
a
certain
state?
And
so
this
is
kind
of
what
they're
getting
at
here.
A
Based
on
solely
and
differential
persistence
of
biological
entities,
without
the
need
for
conventional
replication,
this
cause
for
reconsideration
of
how
ecosystems
and
social
socio-ecological
systems
can
evolve,
based
on
identifying
systems-level
properties
that
affect
their
persistence
and
so
feedback
cycles
have
irreducible
properties
arising
from
the
interactions
of
unrelated
components
and
are
critical
to
determining
ecosystem
and
social
system
persistence,
and
so
they
get
into
self-perpetuating
feedbacks
and
cycles
built
from
the
byproducts
of
traits,
naturally
selected
at
lower
levels
avoid
conflict,
and
so
they
go
through
this.
A
This
list
I
mean
it's
a
lot
of
there's
a
lot
of
words
here
that
I've
gone
through
but
like,
basically,
they
look
at
a
number
of
these
things.
It's
a
very
theoretical
paper
feedback
cycles
is
units
of
persistence,
based
selection,
so
they're
looking
at
feedback
loops,
they're
interested
specifically
at
something
called
auto
catalytic
networks,
which
are
like
networks
of
molecules
that
are
sort
of
self-reproducing
and
self-regulating.
A
A
Another
is,
let's
see,
disturbance,
enhancement,
so
those
are
the
two,
I
guess
and
then
feedbacks
involving
diversification
and
specialization.
They
get
into
that.
This
is
a
nice
figure
showing
how
self-promoting
cycles
support
diversification.
A
So
there's
this
relationship
between
size
and
performance
and
diversification,
so
systems
will
grow
and
change,
but
they're,
based
on
these
rules
that
they're
trying
to
sort
of
propose
here
finding
stability
is
also
interesting,
so
systems.
You
know
with
a
lot
of
feedback,
they're
very
dynamic,
but
they
have
to
be
stable,
or
else
they
fall
apart
and
so
they're
going
to
give
you
the
they
kind
of
go
into
that
part
here
and
then
the
concluding
remarks,
the
outstanding
questions.
A
So
they
have
a
number
of
questions
here
that
are
possible
to
to
come
out.
You
know
come
up
with
from
this,
so
this
is
a
good
good
paper.
If
you're
interested
in
this
topic,
I
thought
it
was
interesting.
I
wanted
to
go
over
it
a
little
bit
and
give
people
a
taste
of
it.
There's
this
other
thing
here
this
paper
on
deep
learning
for
reconstructing
continuous
processes.
A
So
this
is
a
shorter
paper.
It's
a
tour
data
science,
blog
post.
How
deep
learning
is
revolutionary,
revolutionizing
cell
biology?
A
A
So
they
talk
about
this
in
the
context
of
convolutional,
neural
networks
and
cell
biology
presents
us
with
a
very
different
problem
from
some
of
the
benchmark
data
sets,
and
that
is
that
the
classes
that
we
would
like
to
categorize-
often
very
not
so
much
in
shape
but
in
size,
and
so
that's
something
to
keep
in
mind,
and
this
gives
you
a
lot
of,
I
think
kind
of
frames.
A
lot
of
when
we
talk
about
deep
learning
in
the
type
of
data
that
we're
dealing
with
in
this
group.
A
A
A
I
would
say
that
there's
a
lot
of
complexity,
a
lot
more
complexity
than
that.
I
think,
like
we've
talked
about
meta
features
in
images,
and
those
are
things
that
exist
across
cells
across
cell
types
that
maybe
give
us
clues
to
like
motion
or
other
types
of
processes.
So
you
know
there's
a
lot
more
to
it.
I
think,
but
I
think
this
gives
you
a
good
handle
on
some
of
these
issues.
A
A
A
I
think
she's
involved
with
santa
fe
institute
and
in
that
area
of
science
and
they've
published
this
new
article
on
evolutionary
computation
from
a
biological
perspective.
So
sometimes
we
talk
about
evolutionary
computation
in
this
group,
so
evolutionary
computation
is
inspired
by
the
mechanisms
of
biological
evolution
with
algorithmic
improvements
and
increasing
computing
resources.
A
Evolutionary
computation
has
discovered
creative
and
innovative
solutions
to
challenging
practical
problems.
So
evolutionary
computation
is
where
you
have
like
a
set
of
genes:
they're
they're,
computational
genes,
so
they're,
usually
strings
of
binary
digits
and
those
binary
digits
then
are
used
to
encode
information
that
would
be
encoded
in
a
genome,
and
then
you
use
these
strings.
You
evolve
them
by
replicating
them
and
mutating
them
and
over
time
they
give
you
solutions
to
problems,
and
so
you
apply
this
algorithm
you
select
and
the
best
you
know
strings
you
mutate
them.
A
You
do
all
the
things
that
evolution
does
and
you
end
up
with
these
solutions.
Now
they
tend
to
be
very.
They
tend
to
take
very
long
time
to
evolve
solutions
to
to
converge
on
a
solution,
but
that's
one
of
the
things
that,
but
their
actually
makes
them
very,
that
same
property
makes
it
very
good
for
finding
sort
of
creative
solutions
to
problems,
and
so
people
have
played
around
with
these
for
at
least
20
years
now
and
they're
not
in
the
mainstream,
but
you'll
see
them
mentioned.
A
So
there
are
a
number
of
things
in
biological
evolution
that
so
they're,
not
exactly
a
biological
evolution,
but
we
want
to
always
think
about
in
terms
of
biological
evolution,
and
that's
sort
of
the
promise
of
evolutionary
computation
is
that
it
has
a
lot
of
the
sort
of
characteristics
of
biological
evolution.
So
in
biological
evolution
there
are
things
like
open-endedness,
which
is
like
where
you
don't
really
have
a
goal:
you're
just
evolving
things,
and
you
end
up
at
a
certain
place.
A
A
Organism
to
a
multi-cell
organism:
that's
that's
a
major
transition
in
organizational
structure,
neutrality
and
genetic
drift.
So
not
everything
in
evolution
is
a
product
of
natural
selection.
There's
actually
a
lot.
That
is
the
product
of
neutral
processes
like
the
drift
in
certain
genes
to
certain
populations
and
they're,
I'm
not
going
to
get
into
what
those
are.
But
you
can
look
them
up,
but
those
are
features
that
you
don't
necessarily
get
with
a
machine
learning
algorithm.
A
You
can
also
model
multi-objectivity,
which
is
to
have
two
or
more
goals
in
the
same
algorithm.
So
if
I
wanted
to
model
optimize
for
maybe
three
or
four
different
things
at
once,
you
know
you
could
use
the
evolutionary
computation
to
do
that
complex
genotype
to
phenotype
mappings,
which
are
where
you
basically
have
this
genome.
That
I
told
you
about
and
you
map
it
to
some
maybe
set
of
shapes
which
are
the
phenotype.
And
so
you
can
do
these
type
of
mappings
and
evolve
them
and
get
a
sense.
A
Maybe
sometimes
a
lot
of
people
apply
it
to
biological
problems,
but
you
could
imagine,
maybe
designing
different
types
of
car
body
or
different
program
interfaces
using
the
same
method
where
you
just
encode
all
the
sort
of
encode
templates
into
the
genome
and
then
express
it
in
different
ways
and
you
get
a
phenotype.
A
So
I
mean
you
know
and
then
co-evolution,
which
is
the
evolution.
I
think
krishna
talked
about
co-evolution
in
his
presentation,
where
you
have
two
species
evolving
and
affecting
each
other's
fitness
and
their
evolution,
and
so
you
can
model
all
these
things
using
evolutionary
algorithms,
evolutionary
computation
exhibits.
Many
of
these
to
some
extent,
but
can
be
more
more,
can
be
achieved
by
scaling
up
with
available
computing
and
emulating
biology.
More
carefully,
in
particular,
evolutionary
computation
diverges
from
biological
evolution
in
three
key
respects.
A
It
is
based
on
small
populations,
and
strong
selection
typically
uses
direct
genotype
to
phenotype
mappings,
and
it
does
not
achieve
major
organizational
transitions.
These
shortcomings
suggest
a
roadmap
for
future
evolutionary
computation,
research
and
point
to
gaps
in
our
understanding
of
how
biology
discovers
music,
major
transitions
and
advances
can
lead
to.
A
You
know,
computer
algorithms,
that
approach
the
complexity
and
flexibility
of
biology
and
maybe
serve
as
an
executable
model
of
biological
processes.
So
this
is
a
nice
comparison
of
the
two
areas.
A
You
know
you
definitely
have
the
algorithmic
approach
on
the
one
hand
and
then
the
life
approach
on
the
other
hand,
and
this
kind
of
lays
out
what
evolutionary
computation
looks
like.
So
I
described
it
very
roughly,
but
this
is
a
good
summary
of
it.
A
And
they
talk
about
a
comparison
of
biological
and
computational
evolution
across
several
characteristic
dimensions,
and
so
they
talk
about
the
different
ways
that
you
can
simulate
these
different
dimensions.
There
are
actually
programs
that
you
can
use
so
hyperneat,
for
example,
can
model
genotype
to
phenotype
mappings
mulch
multi-objectivity.
A
You
can
use
a
quality
diversity,
algorithm
neutrality
and
drift.
You
can
use
novelty
search
open-endedness,
you
can
model
with
the
evita
program
and
so
forth,
and
these
are
all
programs
that
are
available.
They've
been
developed
in
computation
in
biology,
so
these
different
things
are
sort
of
their
presence
in
evolutionary
computation
today,
so
major
transitions
are
very
minor
in
computation,
but
very
transformative
in
biology,
and
so
that's
where
we
need
to
catch
up.
A
So
this
is
an
opportunity
here
so
open-endedness.
We
need
to
scale
up
open-endedness
in
algorithms
and
major
transitions.
We
need
to
develop
that
capability
neutrality.
We
need
to
emulate
biology
more
same
thing
with
genotype
to
phenotype,
mappings
and
coevolution.
We
need
to
scale
up,
and
so
this
is
a
nice
road
map
to
how
evolution
the
opportunities
in
evolutionary
computation,
and
so
there
are
also
some
opportunities
and
challenges.
A
So
ecs
have
actually
been
around
about
60
years
and
they
started
with
john
holland,
who
was
a
researcher
at
the
university
of
michigan,
and
he
did
a
lot
of
the
early
work
on
ecs,
but
really
for
the
last
20
years.
They've
been
a
big
deal,
but
if
you
want
to
go
back
and
look
at
some
of
the
literature,
I
recommend
john
holland's
book
and
then
there
are
other
people
who
have
done
a
lot
of
groundbreaking
work
in
these
areas
but
yeah.
A
So
if
you're
interested,
there's,
there's
some
references
in
here,
but
also
you
have
you
know,
there's
a
lot
of
new
and
exciting
work
going
on.
So,
if
you're
interested
in
this,
I
would
follow
up
with
this
article,
but
also
look
in
the
literature
a
bit,
and
so
I
think,
krishna's
talk.
You
know.
I
know
he.
He
gave
a
talk
on
where
we
gave
a
talk
on
the
difference
between.
A
Biological
neural
networks
and
artificial
neural
networks
at
narrow
match,
and
so
I
think
this
this
is
a
nice
follow-up
to
that.
Looking
at
the
difference
between
like
evolutionary
algorithms
and
evolution
so,
and
you
can
see
that
there
are
a
lot
of
maybe
similar
issues
there
in
terms
of
the
differences.
A
A
Okay,
get
a
chat,
oh
carry
on.
Okay,
I've
been
using
screen
print
for
later
reference,
so
I
mean
yeah.
Actually
let
me
put
the
talks
in
a
let's
see.
I
can
do
a
link
to
the
papers
and
that's
going
to
go
away
in
a
little
bit,
but
you
can
make
and
I
can
send
out
the
link
and
I
can
actually
send
up
this
entire
list.
We've
had
a
lot
of
different
references
today
on
different
things.
Why
don't
I
make
a
list
after
the
meeting
and
I
can
send
it
out
that
way?
A
Peop
everyone
has
it
and
they
don't
have
to
worry
about
whether
I
captured
the
screen
or
not
so
so,
yeah,
that's
that's
it
for
today.
I
think
any
comments
or
questions
and
anything.
A
H
I
guess
in
the
previous
meeting
I
told
you
about
your
your
presentation
on
foreign
factor
like
the
euclidean
cycle
of
life,
so
I
can.
I
think
we
can
also
consider
it,
for
you
know,
for
darwin's
day
or
for
the
evolutionary
conference
that
how
you
know,
structures
that
are
geometrically
more
stable.
Are
you
know?
Organisms
evolve.
H
Like
how
the
bones
forms
structure
and
about
the
golden
ratio
in
evolution,.
A
A
Okay,
well
thanks
for
attending
next
week,
we'll
be
back
at
the
same
time,
and
I
don't
know
if
we
have
any
presentations
if
we
want
to
do
we'll
keep
tabs
on
krishna's
abstract.
Maybe
you
can,
you
know,
develop.