►
From YouTube: DevoWorm (2022, Meeting 3): Archaea Shapes, Transformers for Bio, Crumpled Biology, Dinosaur Embryos
Description
Computer vision to analyze Archaea shape variation and diversity. Transformers and machine attention for biological datasets, analysis of crumpled paper (surfaces) and fractal analysis of folded structures. Dinosaur embryos, anatomical configuration in the egg, and semi-directed late embryo behaviors. Attendees: Mainak Deb, Richard Gordon, Karan Lohaan, Susan Crawford-Young, and Bradly Alicea.
B
A
Yeah
yeah,
I
don't
know,
I
think
there
are
a
couple
people
wanting
to
present,
but
now
they're
not
here.
So
I
don't
know
what
the
situation
is.
That's
good.
I
might
present
some
things
first
and
then
see
if
they
show
up.
A
B
B
B
A
Yeah
yeah
yeah
welcome
to
the
meeting
so
yeah
we're
gonna
have
my
knock
present
and
ask
quran
if
he
wanted
to
present
something
on
transformers.
I
don't
know
if
he's
gonna
come
in,
he
might
come
in
later
and
then
we'll
do
some
other
things
as
well.
So
my
knock,
you
were
doing
this
stuff
with.
I
think
it
was
with
dick
on
the
what
what
is
the
topic
of
your
presentation.
C
A
C
D
C
Is
the
screen
visible?
Yes,
yes,
yes,
yeah
sure,
so
the
work
that
we
are
doing
it's
it's
basically
surrounded
around
the
quantification
of
shapes
of
the
rk
images.
C
So
and
it's
been
done
using
the
computer
vision,
algorithms,
so
I'll
go
ahead
and
let's
see
what
it's
about
so
the
work
that
the
part
of
it
that
I'm
working
on
it
involves
mainly
three
steps,
which
is
the
image
preprocessing
step.
Then
we
are
approximating
the
polygons
from
the
extracted
perimeters
of
the
archaea
and
then
we
are
also
quantifying
the
shapes
and
we
are
trying
to
extract
more
statistics
from
the
shapes
that
we
get
so
moving
on.
C
A
C
Yeah,
so
if
you
would
be
considering
the
example
of
this
image
and
we
would
sort
of
be
following
through
of
what
exactly
we
are
trying
to
do,
and
so
in
this
example
for
the
pre-processing
section
we
have,
we
first
had
to
invert
the
image,
pixel
values
and
then
the
scale
that
we
have
to
the
bottom
right
of
the
input
data.
We
had
to
erase
that
and
then
and
then
I
also
added
a
gaussian
blurring
algorithm
over
on
top
of
the
raw
image
to
get
rid
of
the
high
frequency
noise
and
after
okay.
C
Yes,
something
along
those
lines
could
be
done,
but
this
black
bar
after
I
remove
the
bar
that
is
actually
eliminated
after
I
threshold
the
image
so
okay,
okay,
but
yeah
yeah.
So
this
is
the
image
histogram
that
we
can
see
to
the
left
of
the
image
right
before
we
threshold
it
and
after
we
threshold
the
image,
the
pixel
values
are
shifted.
C
C
So
moving
on
in
the
next
step,
we
extract
the
centroids
of
the
of
the
archaea,
and
then
we
approximate
the
polygons
from
it.
The
from
the
perimeter
like
we
use
the
perimeter
and
using
the
raw
perimeter
we
approximate
the
polygons
from
it
and
for
approximating
the
polygons,
I'm
I'm
using
something
called
the
douglas
spiker
algorithm,
which
is
heavily
used
in
cartography.
I
think
and
there's
something
called
the
epsilon
value
which
dictates
the
sensitivity
of
the
algorithm
dictates
how
it
dictates
how
close
the
approximate
shape
is
close
to
the
curved
to
the
curved
perimeter.
C
So,
as
we
can
see,
if
the
epsilon
value
is
a
smaller
value,
it
gives
us
the
if
the
epsilon
is
smaller
than
the
approximate
polygon,
that
we
get
it's
closer
to
the
real
it's
closer
to
the
real
line.
But
if
the,
if
the
epsilon
value
is
higher,
then
we
get
the
approximated
polygon
to
be
a
bit
rough
around
the
edges
and
it
sort
of
it
it
sort
of
it
tends
to
deviate
a
bit
from
the
original
perimeter.
C
So
what
I've
done
is
that
I
am
using
a
fraction
of
the
parameter
as
the
epsilon
value,
which
sort
of
senses
the
parameter,
the
total
perimeter,
and
it
adapts
according
to
the
to
the
to
the
perimeter
of
the
shape
that
we
are
actually
dealing
with
and
in
some
cases
to
the
left.
We
have
these
large
blobs,
which
have,
of
course,
overlapping
cells,
and
we
also
have
smaller
patches,
which
are
not
really,
which
are
not
really
good.
C
The
algorithm
says
that
this
shape
right
here
it
has
four
corners
and
the
area
is
also
being
put
here
which
shows
to
be
179.5,
and
these
and
this
area
it's
in
terms
of
number
of
pixels,
which
are
there
so
there
are.
There
are
a
couple
of
examples
where
we
get
three-sided
results
like
in
this
case.
We
have
a
triangular
shape,
which
and
the
approximated
polygons.
C
That
seems
to
have
the
number
of
corners
to
be
three,
and
we
have
the
four-sided
ones
too,
as
examples
here
and
moving
on
okay.
So
I
I
was
talking
about
the
epsilon
value,
so
this
is
a
gif
that
basically
shows
the
approximated
polygons
for
a
varying
epsilon
value.
So,
as
we
can
see
as
the
gif
perceives
over
time,
the
approximated
polygon,
it
comes
closer
and
closer
to
the
original.
C
The
original
curve
around
the
edges
so
yeah
so
the
so,
if,
if
the
epsilon
value
is
high,
it
gives
us
a
rough
estimate
and
if
it's
slow,
it
gives
us
a
very
fine
estimator,
but
that
also
have.
But
that
also
has
like
it's
not
really
really
viable
to
have
a
very
high
epsilon
value.
Then
we
get
the
polygons
to
have
the
number
of
corners
to
be
30
or
40.
C
What
we
are
doing
is
that
yeah
yeah,
could
you
go
back
one
I've
got.
E
A
question
about
that:
how
do
you
decide
what
the
optimal
value
is
for
the
epsilon.
C
A
F
A
The
epsilon
value,
so
I
think
the
image
on
the
right
is
that
an
animation
that
shows
kind
of
like
where
you're
increasing
the
number
of
edges
or.
C
A
So
higher
the
epsilon
value,
then
it's
becoming
more
refined
around
the
edge
or.
A
A
Yeah,
oh,
hey,
I
missed
your
answer
because
it
cut
out.
A
Okay,
yeah,
I
don't
know,
he's
cutting
out.
I
think
his.
A
Yeah
it
keeps
cutting
out,
though
oh
yeah,
the
connection
keeps
cutting
out.
Okay,
okay,.
C
Yeah
sure
so
could
you
hear
the
answer
or
should
I
go
over
it
again.
C
C
C
The
angle
should
be
the
angle
that
we
should
be
getting
in
case
of
an
ideal.
Polygon
would
be
the
red
line
here,
but
clearly
the
clearly.
There
is
some
deviation
that
we're
getting
in
this
approximate
polygon
and
the
same
goes
with
this
example
to
the
right.
C
So
that
is
something
that
we
are
doing
and,
as
dick
had
told
like,
there
are
some.
There
are
some
cells
which
contain
holes
in
them.
As
we
can
see
in
this
example,
it's
not
really
visible,
but
once
we
plot
it
in
3d,
we
can
have
like
a
depression
where
the
whole
could
be
so
that
sort
of
allows
us
to
look
into
it
from
a
different
angle
and
and
sort
of
just
see
where
the
hole
is,
that's
something
that
we've
been
doing
so
exploring
yeah
so
exploring
the
other
other
parameters.
C
E
C
So,
as
we
can
see
that
there's
there
are
certain
areas
which
are
in
a
way
preferred
like
they.
They
occur
more
and
there
are
certain
and
there
are
certain
areas
which
occur
less
and
it
seems
to
be
the
case
that
the
average
area
seems
to
be
slightly
more
than
600
pixels.
C
So
that
is
something
that
we
are
doing
and,
apart
from
that
from
the
estimated
from
the
approximated
polygons,
we
are
also
plotting
the
internal
angles
that
they
have
and
the
red
line
here.
C
It
signifies
the
average
internal
angle
of
of
the
data
that
we
have
and
this
green
line.
It
signifies
the
ideal
internal
admin,
the
ideal
internal
angle
that
we
should
be
having
in
case
every
one
of
these
polygons
were
basically
ideal.
But
that's
not
the
case.
So
we
clearly
have
a
deviation
here
and
apart
from
that,
the
parameters
that
we
had
extracted
in
the
second
step.
C
We
are
also
plotting
out
the
curvatures
of
the
of
the
perimeter,
so
this
heat
map,
it
basically
maps
the
pixel
values
of
the
parameter
according
to
how
curvature
or
of
of
how
the
curvature
is
varying.
If
it's
a
sharp
end,
then
we're
getting
something
like
a
red
part
there,
and
if
it's
not
really
sharp
and
if
it's
smooth
then
we're
getting
like
a
color
closer
to
the
background.
E
Okay
suggestion
here:
if
you
look
at
the
contour
areas,
okay,
then,
if
we
did
that
for
bacteria
and
for
eukaryotic
cells,
it
might
be
sharper.
C
C
C
A
Okay,
that's
great,
why
not
thank
you
so
yeah?
I
like
that.
I
guess
I
had
a
question
about
the
curvature
heat
map,
so
I
I
didn't
really
understand
what
was
going
on
there.
Could
you
explain
it
a
little
bit
more,
I
mean
what
are
those
circles.
A
C
Okay,
the
circuits
are
basically
the
perimeters
that
we're
getting
here.
Is
it
visible?
Yes,
so
what
we're
doing
is
that
we're
finding
the
curvatures
along
this
edge
and
we're
plotting
it
as
a
heat
map?
So
if
it's
a
very
sharp
curve,
we
get
a
larger
value
and
like,
and
if
it's
not
a
sharp,
and
if
it's
not
really
a
sharp
curve,
then
we
get
a
value
which
is
closer
to
the
background.
C
Okay
and
I
just
thought
of
like
just
to
map
the
pixel
values
as
the
curvature.
Just
you
know
just
to
get
a
more
visual
representation
of
what
we're.
A
C
A
I
see
so
the
the
scale
is
like
the
coordinates
in
the
image.
A
E
Impress
on
a
theory
of
what
determines
the
shape
it
turns
out
if
you
assign,
if
you
assume,
that
a
noun,
a
spherical
still
pancakes
known
as
flats
okay,
then,
if
you
assign
a
different
energy
to
the
flat
surface
to
an
edge
in
the
corner,
so
three
different
energies.
Can
you
go
through
those
energies?
You
can
explain
the
political
shapes
but,
of
course,
nobody's
measured
those
energies
so
until
they're
measured,
we
can't
be
sure.
Theory
matches
the
actual
archaea,
but
you
do
get
polygons
when
you
do
that.
E
E
A
So
great,
thank
you
once
again.
So
quran
are
you
there.
A
G
H
A
So
dick
emailed
me-
and
he
asked
me
about
like
transformers-
and
he
wanted
to
know
more
about
it-
and
I
asked
quran
if
he
could
provide
a
quick
overview
of
them.
So
this
is
an
area.
You
know
it's,
it's
a
up-and-coming
machine
learning
model.
So
you
know
it's
one
of
these
things.
The
last
couple
years
has
emerged
and
it's
you
know,
sort
of
a
contender.
A
A
So
I
actually
attended
a
talk
last
week
on
like
comparing
attention
transformers
which
are
where
they
select
different
parts
of
a
data
set
like
image,
information,
text,
information
and
so
forth,
and
put
it
together
the
attention
being
like
the
spotlight
on
different
things
and
then
comparing
that
with
actual
like
biological
attention
in
the
brain.
So
it
was
kind
of
an
interesting
session.
They
they
had
a
lot
of
experts
from
ibm
and
other
area.
A
G
Okay,
you
can
see
your
scroll
is
the
screen.
Yes,
okay,
so
coming
to
platform,
is
they
make
I'll
start
with
islands
they're?
Currently
human
networks,
because
they're
kind
of
based
on
the
original
idea.
G
Yeah,
this
is
the
structure
of
a
very
general
transformer.
Okay,
it
has
like
an
encoder
part
and
a
decoder
one.
The
decoder
part
of
the
pre-processing
part
is
kind
of
very
similar
to
what
an
iodine
is
like.
G
So
essentially,
what
it
does
is,
there's
like
a
vector
space
of
birds
with
similar
meanings
like
the
transformer
is
used
for
generally,
you
know
for
like
nlp
natural
language
processing,
in
which
you
know,
we
have
like
a
vector
space
of
words
which,
like
that
vector
space,
is
based
on
the
meaning,
the
inherent
meaning
of
the
words
itself.
So
it's
like
there
are
words.
G
So
what
transformers
specifically
was
created
to
you
know,
modify
link.
It
was
based
on
the
rm
architecture
and
it's
like
a
modified
sequence
to
sequence
model.
So
what
it
will
do
is
it
will
take
an
input
sequence
of
vectors
and
it
will
output
again
a
sequence
of
vectors
compared
to
like,
let's
say,
image,.
G
Or
image
recognition
does
there
also
like
we
have
when
we
usually
use?
You
know
a
big
eye
or
like
a
matrix
or
another
any
sort
of
sequence,
and
we
usually
get
a
playing
analysis
of
that
thing,
and
you
know
we
determine
whether
that
thing,
whether
there's
an
object
in
that
thing,
so
that
would
be
like
that
would
be
more
of
a
sequence
to
let's
say
a
linear.
G
D
G
G
The
two
unique
aspects
of
transformers
that
have
kind
of
you
know.
G
Made
them
better
at
whatever
task
that
they
were
supposed
to
do
so
when
it
comes
to
like
a
general
vector
space
like
I
was
talking
about
how
you
know,
words
have
different
values
in
that
vector
space
and
they're.
You
know
mapped
like
each
words
can
be
mapped
either
by
zero,
like
the
relationship
with,
let's
say,
or
any
basic,
like
two
words,
let's
say
or
like
the
screen
and
the
white
space
behind
the
screen,
like
the
image
that
you've
seen
right
now.
G
On
the
white
screen,
let's
say
this
is
this:
is
the
sentence
so
here
the
there
like
four
five
words
here
white
screen?
Is
there
so
we
kind
of
get
the
context,
the
global,
contextual
meaning
from
the
global
embedded
space,
which
is
the
electric
space?
Where
you
know,
we've
mapped
a
relationship
between
different
words,
and
you
know
how
far
there
because
computers-
don't
they
don't?
G
The
transformer
uses
it
kind
of
creates
a
separate
vector
space
for
localizing,
or
you
know,
gathering
more
contextual,
meaning
of
these
sentence
itself
like
if
you
have
like
a
very
general
relationship
mapping,
you
know
between
words,
you'll
come
out
with
more
maybe
general
ideas.
You
know
that
won't
be
very
context
specific
like
if
I
have
a
general
mapping
like
let's
say
somebody
gave
me
a
dictionary
and
told
me,
you
know
the
words
are
related
in
this.
This
particular
matter.
G
G
Given
me,
some
contextual
information,
so
I'll
have
to
you,
know
to
either
base
my
predictive
sentences
I'll
have
to
use
that
contextual
information
to
either
draw
more
inputs,
because
if
you,
you
know,
keep
on
adding
more
sentences,
you
kind
of
come
to
realize
that
just
using
your
global
vector
space,
you
know
won't
be
enough,
for
you
know,
drawing
useful
deductions.
G
G
Now
it
could
go
into
like
the
practical,
how
to
you
know
practically
implement
it
using
the
tensorflow
library
and
bison
hair
they've,
given
a
very
good
description,
and
you
know
how
to
get
straight
away
with
doubling
time
with
the
one,
but,
conceptually
speaking
the
differences
arise
with
like
these,
these
two
areas,
what
is
positional
encoding
and
the
attention?
This
is
multi-attention
here,
it's
it's
again,
like
they've
kind
of
explained
it
properly
in
other
ways
yeah.
G
So
these
two
specific
you
know
there's
like
a
function
here
which
uses
qkv
q
stands
for
the
query.
Vector
k
is
the
value.
G
So
when
it
comes
to
practically
implementing
something
to
you
know,
give
a
lot
of
like
give
at
least
a
higher
priority
to
that
localized
vector
space
rather
than
the
globalized
vector
space.
You
know.
So,
let's
say
a
dog
is
walking
on
the
street,
so
you
you
know
trying
to
otherwise
conceptually
this
is,
you
know
essentially
what
it
means.
Otherwise,
if
you
get
into
what
it's
actually
doing,
it's
like
it's
generating
an
array
of
values.
You
know
like
dog,
will
have
some
particular
value
in
that
vector.
G
Space
walking
will
have
some
particular
value
in
that
vector.
Space
and
street
will
have
some
particular
value
of
that
into
space,
and
so
it's
like
an
array
of
values
that
range
from
depending
on
how
how
we
trained
on
what
we
embedded
space
is
like.
So
it
will
have
a
range
of
values
in
that
particular.
G
And
that
range
of
values
will,
you
know,
go
through
these
functions?
One
of
them
is
positional
encoding
which,
like
it
will
again
not
only
draw
like
that.
Particularly
let's
say
dog
is
assigned
the
value
of
1.6
and
walking
is
saying
that
some
other
particular
value,
let's
say
y
x
and
y,
let's
take
x
and
y,
so
those
particular
values.
G
Through
positional
encoding
and.
G
These
key
points
within
that
vector
space
over
the
iterations
that
it
goes
through.
H
G
Translation
algorithm
slightly
more
so
it
kind
of
like
it
changes
the
relationships
between
those
particular
key
points
in
that
vector
space.
You
know
this.
G
B
G
Exactly
perfectly
close,
but
something
that
is
very
accurate
to
the
actual
model,
for
you
know
where
things
lie
in
the
contextual
vector
space
so
like.
B
G
You
know
it
has
gone
through
all
sorts
of
blogs
that
are
there
on
the
internet
to
all
sorts
of
if
you're,
specifically
training
it.
G
I
just
want
to
train
it
on
research
papers.
You
know
the
kind
of
language
that
is
used
in
research
papers
on
that
kind
of
data
and
it
will
generate
a
global
sort
of
you
know
key
point
vector
space
where
each
word
relation
to
another
word
is
kind
of
map.
It's
it's
a
lot
of
data,
it's
a
lot
of
data,
so
that
is
like
that
would
be
like
the
global
attention
vector
space
that
I
would
be
using
for
a
model
like
this,
whereas
the
local.
G
G
B
G
G
Predicting
certain
things
or
like,
in
this
case,
the
model
of
a
protein
structure,
or
you
know.
B
G
Names
that
is
given
to
you,
so
how
do
you
it's
like,
having
a
very
basic
sequence
and
generating
the
3d
computational
3d
model
with
that
that
is
more
computationally
complex.
So,
if
you
see
compare
different.
G
G
Current
ai
lacks
a
lot.
Is
this
general
intelligence
thing
that
you
know
humans
have
this
tendency
to
have
like
if
we
see
some
very
basic
other,
you
know
that
has
been
committed
by
and
yeah.
We
just
you
know
understand,
but
we.
G
G
Amber
this
is
like
a
big
database.
Sequence
of
you
know
protein
sequences
and
how
they
fold.
So
this
was,
I
think,
written
quoted
in
portrait
and
see
by.
B
Some
professors
in
some.
G
G
G
Percent
accuracy
on
most
of
the
tests
that
they
were
given
giving
it
to
you
know
for,
like
they
have
like
predicting
these
structures
in
which
the
ways
in
which
proteins
fall.
They
have
these
competitions,
gas
competitions
in
which
they
have
to
predict.
You
know
what
this
structure
will
hold
like.
It
will
finally
be
like,
so
they
had
performed
quite
well.
You
know.
G
G
But
they
kind
of
you
know,
give
a
brief
overview
about
the
things
that
they
considered
by
doing
something.
Similarly,
because
when
it
comes
to
your
like
they're,
like
molecular
physics.
G
That
has
stack,
and
you
know.
G
G
G
Some
key
points:
you
know
that
they
had
taken
into
consideration
when
coming
up
with
this,
but
essentially
what
they
were
doing
is
again.
They
were
taking
structures
amino
acids
two
at
a
time,
and
you
know,
mapping
out
their
relationships,
then
mapping
your
maybe
taking
them
batches
in
batches
and
then
mapping
out.
You
know
how
they,
you
know,
relate
to
the
overall
structure,
whether
like
proteins
such
as
from
alpha
helixes,
and
you
know
beta
sheets
and
crazy,
crazy
combinations
of
structures.
So
how
you
know
those
relationships
like
any.
B
G
G
So
this
is
this
is
how
you
know
you
deal
with
any
problem
of
this
sort.
So
I
think
it
has
a
lot
of
potential
for
computational
biology,
and
you
know
things
that
our
body
has
been
covered,
but
the
thing
is
you're
using
the
transformer
with
words
is
easy,
like
it's
it's
very
you
can
still,
you
know,
use
intuitive
understanding
of
the
network
and
the
simple
relationship
between
words-
and
you
know
the
language
english
language
itself
to
you
know,
maybe
generate
a
model
that
is
very
accurate.
G
So
this
thing
was
a
big
breakthrough.
In
that
sense,
you
know,
because
I
mean
they're,
creating
the
3d
structure
from
sequences,
so
yeah
this
way,
it's
kind.
B
G
A
A
Yeah,
you
can
also
show
us
yeah
there
we
go
so
transformer,
okay
and
that's
okay.
I
put
that
in
the
chat
and
they
know
I
might
not
put
the
link
to
the
other
paper
in
the
chat.
So
the
other
paper
was
a
so
yeah
he.
It
was
like
a
language
based
paper,
so
yeah
it
looks
like
you
know
you
can
do
it
doesn't
necessarily
differentiate
between
like
say
languages
like
spoke
natural
languages
and
say
like
protein
sequences
or
gene
sequences.
A
It
kind
of
treats
it
in
the
same
way
and
that's
a
lot
of
like
there
are
a
lot
of
computational
methods
to
do
that.
But
I
think
it's
interesting
how
these
models
first
of
all,
can
extract
a
space
like
that
and
then
secondly,
use
these
different
attentional
mechanisms
to
you
know
to
to
sort
of
search
that
space
and
find
some
sort
of
solution.
A
E
Yeah,
okay,
carrie.
Can
I
introduce
a
challenge
for
you:
yeah
yeah,
yeah?
Okay,
of
course,
all
of
those
proteins
are
made
of
l
amino
acids.
E
G
E
Yeah,
you
see,
obviously,
if,
if
you
have,
if
all
the
amino
acids
were
d,
all
you'd
have
to
do
is
take
the
mirror
image.
G
E
Okay,
okay,
you
know,
if
you
put
a
glycogen,
which
is
a
chiral
for
if
you
put
a
d
in
in
place
of
the
mill,
presumably
it
will
break
the
alpha
chain,
yeah,
okay,
but
how
important
is
that,
in
terms
of
hydro
overall
hydrophobicity,
I'm
thinking
in
particular
about
membrane
peptides,
which
are
around
20
amino
acids,
at
least
the
sec?
The
segment
of
the
membrane
protein
is
typically
about
that
are
actually
embedded
inside
the
membrane.
G
Like
to
look
into
you,
the
exact
relationship
between
hydrophobicity
and
the,
I
think
this
is
more
rooted
in
you
know
molecular
physics
and
I'll.
A
Well,
thank
you
again,
quran
for
that
presentation
and
susan
said,
thank
you
to
both
of
you
for
your
presentations
today.
So
that's
great
quran.
Could
I
share
my
screen
yeah.
A
So
a
couple
things?
Okay,
so
that
I
want
to
talk
a
little
bit
about
something.
That's
a
little
bit
different
from
what
we
were
talking
about
in
the
last
two
presentations.
This
is
an
archive
paper
called
the
geometry
of
crumpled
paper,
and
so
why
do
we
care
about
crumpled
paper,
it's
kind
of
trivial
but
as
it
turns
out
crumpled
paper,
if
yeah,
if
you
look
at
like,
say
the
surface
of
the
human
brain,
for
example,
you
notice
that
there
are
these
sort
of
convolutions
or
what
they
call
guarification.
A
So
you
see
all
these
folds
and
it
turns
out
that
the
crumpled
break
yeah
yeah.
So
it
turns
out
that,
like
the
brain,
the
neocortex
originates
as
a
sheet
and
it
grows
in
size
and
then
it's
sort
of
wrapped
around
the
rest
of
the
brain
on
the
top
and
as
it
grows
in
development,
it
you
know
has
to
fit
into
the
brain
case,
so
it
crumples
up
and
there's
a
there's.
A
There
are
certain
species
in
mammals
who
have
this
verification
ability,
where
the
that
neocortical
sheet
folds
up
and
in
humans,
it's
kind
of
its
extreme
form
where
it's
really
folded
up,
and
so
you
get
tissue
that
is
sort
of
the
surface
is
very
convoluted
and
it's
crumpled,
and
you
actually
see
this
in
other
biological
materials
or
organs
as
well,
where
you
have
folds
and
and
crumples,
and
things
like
that.
E
Mention
an
exception
yeah.
If
you,
there
has
been
at
least
one
case
of
a
person
who
had
hydrocephalus
the
size
of
the
skull
enlarges,
yeah
and
the
brain
basically
is
not
crumpled,
it's
flat
against
the
skull,
and
yet
the
person
had
normal
intelligence.
E
A
Okay
yeah,
but
the
question
is,
then,
you
know:
can
we
like
look
at
these
this
crumpling
and
measure
it
and
understand
it
a
little
bit
better
and
and
of
course
this
relates
to
like
fractal
dimensions
if
you're
familiar
with
fractals,
you
know
that
fractals
are
just
kind
of
like
you
know,
structure
at
finer
and
finer
scale,
so
you
can
look
at
something
in
a
very
broad
scale
and
you
notice
that
it
has
a
certain
shape
and
then,
as
you
zoom
in
you
see
like
that,
the
shape
is
is
variable
so
that
you
know
you
have
different,
folds
and
and
and
curves,
and
things
like
that.
A
A
A
A
The
curvature
distribution
follows
an
exponential
form,
which
means
it
follows
a
sort
of
distribution
with
a
long
tail
where
you
get
a
lot
of
smaller
values
and
some
larger
values
in
the
mass
of
the
distribution
with
regions
of
high
curvature
localized
along
ridges.
So
actually
we
saw
a
little
bit
of
that
in
minox
presentation,
where
you
get
the
curvature
isn't
evenly
distributed
across
the
edge
of
the
thing
you
know
in
this
case.
E
E
A
E
A
E
Flattened
polygons,
but
they're
too
surfaced
you
have
to
think
of
taking
something
that
can
ball
and
flatten
here
right
right,
because
the
polygon,
the
edges,
will
have
a
high
curvature
in
the
vertical.
You
know
perpendicular
to
the
flattened
direction.
Okay,
okay,.
F
C
A
A
So
what
they
mean
by
that
is,
they
can
take
this
surface
and
they
can
find
the
ridges
and
then
they
can
fi
build
a
network.
So
they
just
build.
Like
a
you
know,
a
set
of
nodes
or
or
like
the
set
of
lines
that
result
from
the
crumples
as
edges
on
a
network
and
then
where
they
intersect,
they
get.
E
F
A
Right
so
yeah-
and
I
don't
know
if
they
unfolded
the
paper,
but
so
they're,
basically
able
to
use
this
first
exponent,
which
is
a
technique
for
looking
at
like
these
exponentially
distributed
structures,
these
distributions
and
then
that's.
They
say
it's
in
contrast
with
previous
results.
So
people
have
done
this
a
number
of
times
so
much
of
the
theory
of
crumpled
membranes
and
shells.
So
they
actually
look
at
they've
looked
at.
A
We
talked
about
seashells
last
week,
so
they
they're
looking
at
things
like
polymerized,
vesicle,
membranes,
crumple
zones
and
automobile
bodies,
and
I
think
also
seashells,
and
things
like
that.
So
they've
actually
looked
at
people.
A
Other
people
looked
at
this
in
terms
of
equilibrium,
transformations
which
were
a
couple
of
citations
here:
nelson
1988,
the
elastic
properties
of
macroscopic
sheets
kramer
and
whitton
97,
certain
maha,
evita
2003,
and
then
a
number
of
studies
treat
the
geometry
of
develop
a
developable
cones,
which
form
the
geometric
foundation
for
structures
found
on
elastic
sheets
are
subject
to
point
deformations.
A
So
that's
when
you
have
a
sheet
that
can
stretch-
and
you
have
one
point
that
goes
up
into
it
and
it
doesn't
go
through
it.
But
it
pushes
it
up
and
then
you
get
this
cone,
and
so
that's
the
point
deformation
for
that
cone.
It
just
creates
the
cone
and
people
have
looked
at
that
as
well,
and
so
few
studies
exist
that
directly
explore
the
crumpled
state
and
elastoplastic
deformations
occur.
So
this
is
where
they're
doing
this
so
they've.
A
You
know
there
are
some
experiments
where
they
just
they
focus
on
the
dimension,
fractal
dimension
of
a
ball
that
circumscribes
a
flat
sheet
once
crumpled.
So
this
is
gomez
89.
These
are
a
lot
of
physics
papers.
I'll
put
this
in
the
chat,
so
you
can
follow
up
on
it
if
you
would
like-
and
this
is
just.
A
Okay-
and
so
you
know
they're
looking
at
a
lot
of
different-
there
are
a
lot
of
different
ways.
You
can
look
at
this
and
they
do
so.
This
is
their
experiment
here,
where
they
have
this
crumpled
piece
of
paper
they
on.
I
guess
they
unfold
it
here
and
then
they
measure
with
a
laser
the
distance
across
the
sheet.
A
It
shrinks
a
bit
if
you
measure
across
the
sheet,
but
that
x
that
area
or
that
that
distance,
that
it's
shrunken
by
is
embodied
in
these
folds,
and
so
this
is
what
they're
trying
to
measure
out
and
of
course,
if
you
measure
different
scales,
you
get
a
little
bit
different
measurement,
because
the
ridges
are
fractal,
as
you
can
see
here,
and
so
then
they're
they
take
a
camera
image
of
this
measurement
and
they
can
approximate
these
distances
across
the
crumpled
sheet.
A
A
So
we
observed
that,
as
expected,
the
curvature
of
the
fold
increases
with
an
increasingly
applied
force.
There
is
always
plastic
deformation
present.
The
data
is
linear
over
two
orders
of
magnitude
and
applied
force.
A
So
here
are
some
of
their
force
measurements
and
then
the
distribution
here,
the
first
exponent
they
used
to
measure
the
fractal
dimension
of
it.
So
that's
what
that
measurement
comes
into
play
and
it
again
it
relates
to
the
distribution
of
here
yeah.
So
this
is
a
nice
paper
you
can
get
into
it.
If
you'd
like
a
little
bit
further
looks
like
susan
put
some
things
in
the
chat
about.
Thank
you
for
the
presentation.
That's
unusual
hydrocephalism
usually
results
in
retardation
and
death.
A
That's
what
she's
talking
about
with
the
the
smooth
brain
example
that
we
were
talking
about
before,
so
my
knock.
Did
you
have
something
to
that?
You
want
to
bring
up
or.
C
Yeah,
so
I
next
week
I'll
be
having
another
I'll
be
showing
something
else
that
I'm
working
on
so
could
that
be
done.
A
All
right
that
sounds
great,
that's
all
great
all
right!
So,
if
you
need
to
go,
you
can
go,
I'm
going
to
talk
about
one
more
thing
and
then
that'll
be
that
so
thanks
for
attending,
if
you
have
to
go,
if
not,
I
have
one
more
thing
I
want
to
get
into
so
I'm
going
to
get
into
this
interesting
set
of
papers
that
came
up
over
the
holiday.
A
They
found
a
nice
dinosaur
embryo
and
I
know
we've
talked
about
paleo
embryos
before,
but
this
is
something
it's
like
found
that
was
almost
perfectly
preserved.
This
is
in
this
chinese
assemblage
that
we
talked
about
earlier
one
of
the
places
where
they
found
some
of
the
first
embry,
but
they
think
are
the
first
embryos.
A
So
this
there's
a
couple
of
pictures
here,
the
you
know
they.
They
always
put
these
news
stories
out
and
they
kind
of
I
don't
know
if
they've
oversold
it,
but
these
are
some
of
the
images
that
they've
had.
So
this
is
one
that's
where
they
show
the
sort
of
the
embryo
in
in
the
fossil
fossilized
form.
So
you
can
see
a
lot
of
detail
on
the
top
here.
A
You
have
some
it's
really
kind
of
out
of
focus
on
this
end,
but
you
can
see
some
of
the
parts
of
the
embryo
here
this
one
here
has.
This
is
showing
how
the
dinosaur
is
kind
of
growing
up
within
a
within
the
shell
of
the
egg,
and
so
it's
folded
up
like
this.
So
you
can
see
that
there's
a
lot
of
folding
of
the
you
know
the
appendages
around
the
head
and
then
eventually
it's
going
to
hatch.
A
You
know
it's
like
kind
of
like
a
bird
or
an
an
amphibian
where
they
hatch
out
of
their
egg
and
they
have
their
little,
but
they
you
know.
I
guess
this
is
like
a
lunch
box
for
the
embryo
here
and
then
here's
a
fossil
embryo
here.
So
you
compare
these
two
things
where
it's
folded.
This
is
sort
of
an
artist's
rendition
of
the
folding
of
the
embryo
up
in
this
egg.
A
Until
it
hatches-
and
this
is
you
know-
we
have
nice
pictures
of
c
elegans
eggs
where
the
c
elegans
is
coiled
up
inside
the
egg.
So
this
is
something
that
happens
within
eggs
where
you
know,
as
the
organism
grows
and
differentiates
it's
still
kind
of
folded
up
like
this,
and
then
it
has
to
unfurl
when
it
hatches
from
the
egg.
A
So
that's
the
artist's
rendition,
and
then
this
is
these
are
fossil
remains
of
the
egg
and
the
embryo,
and
so
you
can
see
that
there's
it's
kind
of
hard
to
see
the
folding,
but
you
can
see
the
similar
sort
of
setup
where
you
have
the
head
here
and
then
it
wraps
around
and
then
the
appendages
are
kind
of
folded
over
the
head.
A
So
this
is
the
paper
in
eye
science
and
exquisitely
preserved
in
ovo.
That
means
within
the
egg,
so
they
didn't
have
to
like
piece
it
together
from
different
things.
It
was
the
the
fossil
was
right
there
and
that
must
be
extremely
rare.
You
know
to
find
something
that
will
preserved
all
in
one
place:
theropod
dinosaur
embryo
sheds
light
on
avian,
like
pre-hatching
postures.
A
A
You
start
with
this
pre-tucking,
where
you
know
you
have
this
tucking
of
the
head
and
and
then
there's
this.
It
kind
of
shifts
around
a
little
bit
then
there's
membrane
penetration,
then
there's
piping,
what
they
call
piping,
which
is
where
the
legs
start
to
stick
out
and
the
head
folds,
as
you
know,
into
the
upper
appendages
between
the
upper
appendages
and
then
there's
climax,
where
there's
hatching
and
they
have
to
get
their
way.
A
So
that's
that's
what
they're
trying
to
look
at
here
so
they've
really
kind
of
never
yeah.
So
over
100
past
100
years,
they've
discovered
many
dinosaur,
eggs
and
nests,
but
finding
these
articulated
in
oval
embryos
are
very
rare.
So
it's
very
hard
to
find
good
data
sets
that
show
comparisons,
which
is
why
they
have
to
compare
with
extant
organisms
and
that's
not
necessarily
a
solid.
You
know
comparison,
but
that's
what
they
have
to
work
with
here.
I
report
an
exceptionally
preserved
articulated
over
overtroid.
A
I
I
can't
pronounce
some
of
these
words.
It's
heard
of
your
of
europe
torrid
embryo
inside
in
the
longitude
along
a
tooth
lit
egg
from
the
white
cretaceous
eku
formation
in
southern
china.
The
headline
is
ventral
to
the
body
with
the
feet
on
either
side
and
the
back
curled
up
along
the
blunt
pole
of
the
egg
and
a
posture
previously
unrecognized
in
a
non-navy
and
dinosaur.
A
So
this
is
something
that,
like
this
is
a
new
posture.
They
haven't
seen
this
type
of
posture
before
and
they
want
to
sort
of
figure
out.
You
know
they
want
to
figure
out,
maybe
the
biomechanics
of
it
or
how
you
know
this
is
possible
in
in
the
course
of
development.
A
They
say
it's
reminiscent
of
a
late
stage,
modern
bird
embryo.
So
again,
there's
the
comparison
comparison.
Another
late-stage
ovaroptory
embryo
suggests
that
pre-hatch
over
torrids
develop
avian-like
postures,
latent
incubation,
which
is
modern.
Modern
words
are
related
to
coordinated
embryonic
movements
associated
with
tucking,
so
now
we're
making
this
parallel
with
modern
birds
that
they
have
these
first
coordinated
embryonic
movements.
A
They
have
a
nervous
system
in
in
that
phase
of
development,
so
they're
able
to
start
moving
around
and
they're.
You
know
so.
This
tucking
behavior
is
controlled
by
the
central
nervous
system.
It's
critical
for
hatching
success,
but
it's
something
that's
generated
from
the
central
nervous
system.
Now,
in
c
elegans,
a
lot
of
the
pre-hatch
movements
are
triggered
by
muscle
twitching,
so
it's
not
controlled
necessarily
by
the
central
nervous
system.
It's
just
controlled
by
the
muscle.
A
That's
our
that's
sort
of
forming
in
place.
It
starts
twitching,
and
you
see
these
movements
around.
These
are
coordinated
by
the
central
nervous
system,
they're-
probably
sort
of
automatic
movements
that
occur,
but
this
is
something
that's
necessary
for
hatching
success,
so
they
have
to
get
in
position
to
hatch
and
then,
if
they
don't
get
in
that
position,
they
don't
hatch
properly
and
that's
a
problem.
I
don't
really
know
much
about
that,
though.
A
It
just
kind
of
goes
through
a
lot
of
this,
like
they
go
through
the
fossil
that
they
found,
so
they
have.
They
were
able
to
identify
a
lot
of
the
anatomy
here,
which
then
allowed
them
to
reconstruct
the
with
the
the
sort
of
the
pose
of
the
embryo,
and
then
they
wanted
to.
You
know,
compare
it
to
modern
birds
to
see
what
you
know
hypothesize
about
what
those
behaviors
are.
A
You
know
how
they
may
be
generated,
or
what
they're
trying
to
do
in
that
in
that
case,
so
yeah
they
go
through
it's
a
lot
of
anatomy.
They
have
a
lot
of
pictures
of
different
parts
of
the
embryo
anatomy.
Here
this
is
a
pretty
late
version.
You
know
it's
a
pretty
late
stage
in
development
here
you
have
their
mapping
from
the
fossil
embryo
down
to
these
skeletons
that
they're
trying
to
reconstruct.
A
So
you
can
see
that
they're
reconstructing
that
image
that
I
showed
you
before,
where
they
had
it
all
sort
of
an
artist's
rendition.
This
is
what
the
data
looks
like
you
just
have
these
different
parts,
these
different
parts
of
the
skeleton
sort
of
in
these
different
positions
and
they're.
Trying
to
piece
together
what's
happening
so.
B
A
A
Then
they
did
a
phylogenetic
analysis
to
find
out
where
this
organism
fits
in
in
phylogeny,
so
that
they
could
make
some
more
informed,
comparisons
to
extent
organisms,
and
then
this
posture
here
they
just
make
some
brief
comments
about
that,
and
they
make
some
comparisons
to
some
of
the
other
remains
that
they
found
so
they
found
some
other
eruptorate
eggs
in
china
and
a
couple
other
places
and
they've
talked
about
some
of
the
things
that
they
found
so
they've
compared
with
other
fossils
and
some
of
the
things
that
they
found
fossil
embryos
and
some
of
the
other
things
they
found
there
and
how
to
make
comparisons
and
say
you
know
this
is
this
is
what's
going
on?
A
And
then
the
limitations
of
the
study,
so
they
actually,
they
talk
about
different
types
of
imaging
techniques
that
they've
used.
So
more
in-depth
comparative
study
in
the
embryonic
posture
of
non-av
and
dinosaurs
and
extant
archosaurs
are
hindered
by
the
scarcity
of
available
images
and
scans
attempts
at
imaging
one.
Such
sample
have
been
made
using
commun
computed
tomography
and
micro
ct,
but
because
of
the
high
density,
minerals
and
lack
of
contrast
between
bone
and
matrix,
it
doesn't
provide
useful
anatomical
data.
A
So
that's
one
of
the
problems
is
you
have
them
in
this
mineral
matrix
and
it's
hard
to
get
them
differentiate
between
what
is
preserved
bone?
And
what's
that
you
know
sediment,
that's
sitting
in
there.
So
it's
it's
from
a
imaging
perspective.
It's
hard
to
do
so,
and
that's
largely
because
a
lot
of
cases
you
have
iron,
rich
sediments
or
other
types
of
things
that
interfere
with
the
imaging
modality.
A
E
A
E
The
imaging
they
obviously
use
what
are
called
k
edges
k
edge
is,
if
you
have
a
particular
element,
it
absorbs
x-rays
differently
above
and
below
a
particular
wavelength,
and
then
you
can
subtract
those
two
pictures
and
get
tremendously
increased
contrast,
and
since
the
bones
presumably
have
calcium
and
you're
an
iron,
it
might
be
possible.
E
Okay,
yeah,
okay,
okay,
second
comment
is
musing
about
100
years
ago,
when
the
first
dinosaur
eggs
were
discovered,
the
people
who
paid
for
the
expeditions
to
find
them
were
rewarded
by
being
given
an
egg.
F
E
Okay,
so
it
could
be
that
the
only
purpose
of
the
central
nervous
system
is
to
wiggle
things,
and
that
would
be
similar
to
the
nematodes
so
that
they
slip
into
a
configuration.
E
B
E
Good
question
is
whether
it's
directed
or
random
wiggling
that
produces
the
configuration.
E
B
A
Yeah,
well,
I
think
that's
it
for
today.
So
thanks
everyone
for
attending
and
presenting
next
week.
You
know
we'll,
do
I
don't
know
if
anyone
wants
to
print
well,
my
knock
wants
to
present
next
week
and
yeah.