►
From YouTube: DevoWorm (2021, Meeting 25): GSoC #5/6, Networks Redux, Light Environments, Models of Cell Tracking
Description
Update on Summer of Code activities (Mainak Deb, Improving DevoLearn), review of Networks 2021 activities and ANNs/BNNs research. Review of potential tasks and group-wide task board, versioning of DevoLearn. Papers on role of light in visual and developmental environment, RoboWorm paper, and cell tracking and quantification in induced pluripotent stem cells. Attendees: Akshay Nair, Bradly Alicea, Assaf Wodeslavsky, and Mainak Deb.
A
B
A
A
little
soft
okay,
so,
let's
listen
again
so
welcome
to
the
meeting
we
had
a
meeting
last
week
there
was
a
holiday
in
the
u.s
and
canada,
so
we
had-
and
I
think
dick
and
susan
are
still
out.
So
why
don't
we
get
started?
My
knock!
You
have
been
busy
with
your
gsoc
work
yeah.
So
why
don't
you
do
you
have
an
update
to
present?
B
B
A
B
B
B
Pipeline
to
see
if
the,
if
the
approach
is
viable
or
not
so
I
did
that
and
I
trained
it
for
a
couple
of
epochs,
which
is
very
low.
I
know
but
like
at
least
I
got
a
reasonably
good
score,
which
is
a
79
intersection
over
union
score
of
the
segmentation
maps,
and
these
are
the
validation
price
loss
that
I
got
and
the
learning
dates.
Of
course,
they
remained
the
same
because.
B
B
B
B
But
like
like
octane,
actually
did
the
hard.
A
B
B
Like
if
we
notice
out
here
here,
we
have
two
nuclei
and
the
same
goes
here
like
it's
sort
of
in
an
intermediate
stage.
They
are
not
exactly
two
different
separate
entities,
but
they
are
overlapping
and
the
model
also
like
it.
Basically
also.
I
was
able
to
map
it
so
which
is,
of
course,
which
is
a
good
thing.
B
A
So
this
was
the
pull
request
you
issued.
It
was
a
couple
days
ago.
A
So
that
that
looks
good,
yeah
yeah,
so
yeah.
I
like
this
image
here
that
you
just
showed.
A
A
B
A
So
maybe
yeah
go
back
up
to
your
graphs,
a
little
bit
yeah
this
one,
so
the
iou
score
is
intersection
over
union.
A
So
that
means
that
you're,
looking
at
like
there's
an
intersection
between
the
two,
like
the
prediction
and
the
and
the
in
the
mask
and
then
that's
versus
like
the
empty
set,
which
would
be
the
union
where
there's
no
overlap
at
all.
A
B
B
B
B
B
B
So
I
actually
stopped
training
only
just
to
make
sure
that
these
excess
curves
in
the
in
the
nuclei
are
not
mapped
and
are
not.
And.
A
Yeah,
you
don't
want
to
have
overfitting
yeah
yeah
right.
That's
very
good!
I
think
that's
a
nice.
You
know
you're
doing
a
very
good
job
of
applying
the
methods
to
the
problem
to
the
data
set,
and
this
is
a
challenging
data
set
and
you
know
you're.
I
think
you're
learning
a
lot
about
those
challenges
and
drawbacks.
B
B
A
Yeah,
that's
great
just
making
some
notes:
yeah
yeah,
that's
great!
So
what's
your
step
for
next
this
coming
week,
what
do
you
plan.
B
To
do
okay
in
the
upcoming
week,
I
actually
planned
to
build
the
guide
for
this
for
the
simple
online
demo
that
I
was.
A
B
B
B
B
A
A
B
A
Yeah,
that's
great,
thank
you
for
the
update,
okay,
so
that's
great
and
if
you're
watching
on
youtube-
and
you
want
to
know
more
go
to
the
github
repository
mynac
is
pushing
this
to
the
diva
worm,
repo,
the
diva
worm
directory
or
the
divorm
repository
in
the
gsoc
2021.
A
If
you
want
to
take
a
look
and
then
he's
going
to
put
this
push
this
to
diva
learn
soon
so
we'll
have
that
incorporated
and
then
I
don't
know
we
might
release
another
version,
maybe
either
the
midpoint
of
g
sock
or
at
the
end
of
g
sock.
I'm
not
sure
yet,
but
I
think
that's
good.
Now,
I'm.
A
So
you
don't
have
anything
to
worry
about
I'll,
just
fill
it
out
and
then
you'll
have
a
opportunity
to
comment
on
it
or
you
have
to
make
comments
I
think
too,
but
that
they'll
send
you
emails
on
that,
so
that'll
be,
but
that
has
to
be
done
by
a
certain
deadline.
So
just
make
sure
that
you
watch
out
for
those
emails.
A
Yeah,
okay,
so,
let's
see
now,
I'm
gonna
get
into
some
other
things
that
we're
gonna
switch
gears
and
I
want
to
go
over
the
submissions.
I
don't
think
there's
anything
pressing
in
the
submissions,
but
I
did
want
to
go
over
this
for
people
on
youtube
who
are
watching.
A
So
this
is
our
submission
document
and
it's
been
a
little
inactive.
Lately
we
haven't
had
a
lot
of
stuff
coming
up.
We've
submitted
a
lot
of
things.
A
lot
of
things
have
been
presented.
All
of
the
netsile.
All
the
network
stuff
has
been
presented,
this
growth
form
and
theory
of
deep
learning,
which
was
actually
submitted
under
a
different
title
to
net
neuro.
A
That
was
put
up
on
twitter
advertised
on
twitter
and
last
last
week,
and
then
I
actually
made
a
10
minute
video
of
this
kind
of
going
over
the
poster
for
the
attendees,
so
that
that
youtube
video
is
on
our
youtube
channel
for
this
poster.
So
if
you
go
to
our
youtube
channel
you'll
see
that
10
minute
video,
you
can
go
through
the
poster
and
it's
just
a
quick
tour,
and
so
that
was
that
was
from
net
neuro.
A
The
topo
nuts
workshop
was
also
very
good.
They
were
very
excited
about
the
work,
so
this
is
the
oiler
cycles
for
life
and
they
were
very
excited
about
that
work
and
I
think
the
next
steps
on
that
actually
are
to
think
about
the
input
data
for
other
types
of
organisms,
or
you
know
maybe
having
you
know,
maybe
next
year's
gsoc
project
might
one
of
them
might
be
to
have
to
have
input
data
and
then
have
an
interface
where
you
can
actually
do
this
type
of
analysis
on
it.
A
So
there's
you
know,
there's
a
lot
of
potential
for
this.
I
think-
and
this
again
was
like
drawing
a
skeleton
around
the
you
know:
a
cell,
a
single
cell
or
a
small
multicellular
colony
and
then
defining
a
network
in
terms
of
the
edges
instead
of
the
nodes.
So
you
have
these
boxes
in
a
you
know
in
some
sort
of
geometric
array,
and
then
you
calculate
the
euler
completeness
of
that
network,
and
then
you
determine
its.
A
What
would
I
guess,
call
developmental
state
whether
there's
this
sort
of
overlap
or
whether
there's
this
modularity
or
whether
it's
a
complete
circuit?
So
there
I
think,
there's
a
lot
of
potential
to
you
know
test
different
types
of
input
organisms.
I
know
we
have
the
basal
area.
I
don't
know
if
that
would
be
a
good
choice,
but
there
are
other
diatoms
that
where
we
can
get
microscopy
data
for
that
and
then
we
also
met.
A
We
also
talked
about
potential
sort
of,
like
sort
of
like
you
know,
computational,
models
of
the
genetics
and
things
like
that.
So
you
know
we
talked
about
binary
networks.
We
talked
about
genetic
regulatory
networks.
C
A
The
abstract
is
up
on
the
fixture
where
the
slides
are,
but
that,
but
that
you
know
that's
gonna,
be
sort
of
there's
future
work
to
do
there.
A
So
that's
that's
that
and
then
the
main
talk,
the
embryo
networks
plus
connectome's
talk
which
isn't
the
actual
title,
but
I
think
we
talked
about
that
last
time
and
that
went
well
too,
that
video
is
up
on
youtube
and
again
there's
more
work
to
do
there.
I
think,
in
terms
of
like
unders,
like
giving
a
more
formal
network
structure.
A
So
the
idea
of
this
talk
was
to
say
that
there's
this
you
know
there
are
a
bunch
of
developmental
cells,
as
you
see
from
our
embryos
and
those
cells
differentiate
and
as
they
differentiate,
they
start
to
form
two
different
types
of
networks.
They
form
sort
of
an
embryo
network
which
is
an
association
of
embryo
cells
or
developmental
cells,
and
then
the
neuronal
cells
that
start
to
differentiate.
A
And
then
you
have
those
two
networks
that
interact
and
then
you
have
other
things
like
the
germline
and
you
have
other
types
of
tissues
that
start
to
differentiate
and
form,
and
so
those
form
other
networks
as
well,
and
the
idea
was
that
there's
this
process
of
you
know
divergent
integration.
So
those
cells
diverge
in
terms
of
their
fate,
but
those
networks
are
still
integrated
at
some
level
and
so
they're
diverging,
but
there's
maintaining
some
level
of
integration,
but
not
as
much
as
they
would
if
they
were
the
same
type
of
cell.
A
Network
model,
and
then
I
don't
know
what
kind
of
data
we
can
bring
to
bear.
We
have
data
for
clns,
of
course,
but
they
have
to
like
play
around
with
the
data
to
get
it
to
work
for
the
analysis.
So
that's
that's
something
we
can
do
later.
A
This
could
just
simply
be
like
a
theoretical
paper
where
we
create
pseudo
data
where
we
have
like
you,
know,
different
cells,
and
you
know,
make
up
states
and
do
different
like
theoretical
tests
like
look
at
a
random
network
versus
like
a
structured
network,
there's
a
lot
that
we
can
do
with
that.
So
so
we
have
the
different.
So
we
have
these
other
things
that
we've
been
having
these
longer
term
submissions.
A
A
We
have
the
quantitative
comparison
of
our
kia
and
shape
droplets,
which
could
fit
in
plug
into
this
euler
cycles
for
life
project,
or
it
paper
certainly
give
us
data
to
work
with,
and
then
we
have
the
a
couple
things
here.
The
eye
of
nature
book,
which
is
something
that
we
talked
about
steve
mcgrew
and
we
talked
about
his
book
and
now
actually
dick
sent
the
book
to
us
to
me
and
I
think,
to
thomas
or
tom
portages
and
I've
been
looking
it
over.
It's
a
very
interesting
manuscript.
A
It
was
never
published,
but
it's
something
that
it's
it's
on
sort
of
like
evolution
and
how
to
teach
evolution
properly.
And
so
it's
it's
an
interesting
book.
Maybe
we
can
you
know
I
I
don't
know
we're
not
a
publishing
house,
but
it's
possible.
We
could
make
this
available
as
a
sort
of
an
ebook
or
something
in
his
honor.
A
So
we'll
look
into
that
then
there's
the
differentiation
tree
of
the
brain,
which
I
didn't
bring
the
slides
today,
but
I
have
some
interesting
information
about
this
and
some
advances.
I
I
mean
you
know
I'll.
Probably
wait
till
dick
comes
back
from
vacation
to
talk
about
that,
but
that's
something
else
and
then
the
mathematics
of
diva
worm,
which
is
this
right.
Now
it's
a
poster
and
we're
going
to
try
to
make
it
into
something
a
little
bit
more
formal,
but.
A
Equations
of
diva
worm
we
didn't
put
any.
We
don't
really
have
a
lot
in
terms
of
like
deep
learning.
We
have
a
generic
neural
network,
but
did
you
have
something
to
say
about
that?
Minok.
A
A
Okay,
oh
it's
like
actually
it's
here.
Oh
actually,.
C
A
So
that's
I
mean
that's
something
we
can
work
on
here:
mathematics,
a
diva
worm,
I'd
like
to
pull
that
up
in
the
near
future.
I've
also
been
making
some
progress
on
this
basilarian
neuronal
cognition
paper,
and
this
is
something
that
is
it's
long
suffering
and
I'm
gotten
extension
after
extension
on
this
and
I've
been
really
trying
to
work
on
it,
but
I'm
kind
of
getting
some
head
some
steam
on
this.
Some
headway
made
so
we'll
in
next
couple
weeks.
I
think
we'll
be
in
a
position.
A
We're
gonna
review
it
in
the
meeting
and
go
over
some
of
the
points
on
it
and
try
to
flush
it
out
into
something.
That's
really.
A
Looks
like
a
book
chapter
instead
of
a
set
of
notes.
So
you
know.
That
is
something,
of
course,
that
we're
drawing
from.
We
wrote
this
digital
basil
area
paper
two
years
ago.
Now,
where
we
had
the
we
did
this,
it
was
actually
and
ozmit
singh.
Who
did
this
work
with?
You
know
they
they
kind
of,
took
what
was
available
in
terms
of
microscopy
data.
A
They
analyzed
it
using
a
deep
learning
model
and
then
they
were
able
to.
A
You
know
we
were
able
to
have
like
a
digital
data
set
and
then
now
we're
looking
at
sort
of
the
behavior
of
diatoms,
and
this
doesn't
have
a
lot
of
data
behind
it,
but
we're
going
to
try
to
do
some
mathematical
modeling
instead,
and
you
know
that's
where
this
is
going,
so
we
haven't
really
been
able
to
hang
the
data
on
this
yet,
but
I
hope
that
we
can
at
least
come
up
with
something
for
for
a
mathematical,
modeling
audience
so
and
then
the
other
thing
I
want
to
mention
one
final
thing
is
this:
poster
that
was
presented
at
net
neuro
actually
has
a
lot
of
connections
to
the
artificial
neural
network.
A
Biological
neural
network
stuff
that
that
we
were
we've
been
talking
about
in
the
meetings
krishna
and
I
and
jesse
wrote
a
paper
on
this
a
preprint
and
there
was
actually
a
talk
in
the
net
neuro
session
on
artificial
neural
networks
and
biological
neural
networks.
A
Goulos
who's,
a
european
scientist
he's
done
a
lot
of
stuff
with
primate
brain
imaging.
I
think,
and
some
other
topics
like
developmental
neural
net.
You
know
developmental
networks
in
brains.
He
actually
has
this
technique
where
he
can
take
data
from
like.
I
think
he
was
showing
examples
of
the
primate
brain
where
they
were
able
to
take
brain
networks
from
the
primate
brain
and
map
them
to
artificial
neural
networks.
A
So
it's
really
exciting
and
I
put
some
slides
of
it
up
in
the
slack,
and
I
don't
have
the
slack
open
right
now.
But
if
you
go
to
the
slack
and
you
go
to
the
diva
worm,
channel
you'll
see
those
slides,
so
you'll
see
some
slides
posted,
not
in
the
not
too
recent
past
and
that
that's
what
that
refers
to
that
talk.
A
So
there's
a
lot
going
on
in
that
area
and
he's
actually
created
a
program
where
you
can
do
this
or
you
can
convert
the
data
and
so
and
then
actually
in
fact,
I'm
also
mentoring
for
neuromatch
and
a
group
of
mine
is
doing
something
similar
with
the
visual
system
they're
trying
to
take
the
visual
system.
I
think
it's
a
human
visual
system
from
fmri
data
and
map
it
to
a
deep
learning
model
where
the
different
layers
represent
different
parts
of
the
visual
stream.
A
So
that's
all
very
you
know
up
and
coming
stuff.
So
there's
a
lot
of
room
there
to
contribute,
and
I
don't
know
exactly
how
we
contribute
to
that.
This
paper
was
a
struggle
to
try
to
find
like
our
voice
on
that.
So
I
I
think
that,
like
though
we
can
have
a
longer
discussion
about
that
in
the
near
future,
people
want
to
be
involved
in
that,
and
that's
definitely
something
that
you
know
it's
thinking
more
deeply
about
deep
learning
networks.
A
You
know
we
talk
about
the
metrics
and
its
performance,
but
and
then
we
talk
about
this
broad
analogy
to
the
brain.
But,
like
you
know,
what
is
it
about
the
networks
in
the
brain
that
are
similar
and
there's
just
so
much
so
many
people
are
kind
of
just
kind
of
fumbling
around
trying
to
find
those
analogies
and
there's
a
lot
of
interesting
stuff
there.
But
but
there's
a
you
know,
there's
a
lot
of
room
for
people
to
do
stuff
there.
So
so
that's
the
submissions.
A
Now
I
want
to
talk
about
neuromatch
and,
like
I
said,
neuromatch
started
last
week
and
I've
advertised
it
in
the
group
here
and
we've
talked
about
neuromatch
and
how
they're
going
to
they
have
this
computational
neuroscience
course
going
on
right
now
and
in
august
they're
going
to
have
a
deep
learning
course
and
if
you,
if
you're,
not
enrolled,
and
if
you
didn't
have
time
to
enroll
or
you
didn't
want
to,
they
actually
have
the
materials
online.
So
they
have
the
syllabus
here
at
this
address.
A
So
this
is
neuromat,
github,
neuromatch
academy
course
content
and
they
have
the
course
material.
So
you
can
do
what
they
call
the
observer
track,
which
is
to
just
you
know,
do
the
notes
and
go
through
it
on
your
own
and
the
benefit
to
that.
A
A
You
know
whatever
you
find
interesting
say
you
know
about
model
types
you
know,
but
you
really
want
to
get
into
this
modeling
part
here,
it'll
be
two
and
three,
so
you
can
just
go
to
week
two
and
three,
and
do
that
and
explore
that.
So
they
have
a
lot
of
stuff
too,
not
just
deep
learning,
but
in
the
computational
neuroscience
course
they
have
bayesian
models.
A
They
talk
about
hidden
markov
models,
they
talk
about
control,
models,
reinforcement,
learning
and
then
network
causality,
all
which
are
important
in
neuroscience
and
but
they're
not
related
to
deep
learning,
necessarily
now
the
deep
learning
we'll
get
into
that
in
a
minute,
so
they
just
kind
of
give
you
some
materials
here
you
talk
about.
A
You
know
the
things
that
you're
doing-
and
this
is
all
shared
under
creative
commons
license,
so
this
is
all
free
to
interact
with
and
use
now,
if
you
go
to
the
deep
learning
course
content,
which
is
a
little
bit
different,
this
is
this:
is
the
first
year
they're
doing
the
deep
learning
course-
and
I
think
krishna
is
going
to
be
a
ta
for
this.
So
congratulations
krishna,
and
this
is
again
this
is
going
to
be.
A
They
don't
really
have
the
course
materials
worked
out
for
this
yet,
but
basically
they're
going
to
do
they're
going
to
get
really
deep
into
deep
learning,
they're
going
to
do
multi-layer,
perceptrons,
optimization,
regularization,
so
they're
going
to
start
with
kind
of
the
basics,
linear
models
and
then
doing
work
with
a
few,
with
fewer,
more
work
with
fewer
parameters,
so
they're
going
to
show
you
how
to
use
like
conv
nets
and
rnn's
and
attention
and
transformers
generative
models-
and
this
is
of
course
according
to
deep
learning.
A
So
you
know:
generative
models
are
a
lot
of
different
types
of
generative
models,
they're
going
to
focus
on
a
couple
here
that
are
gans
and
vaes,
so
variational
autoencoders
and
generative
adversarial
networks.
So
these
are
all
like
really.
A
You
know
this
kind
of
demystifies
a
lot
of
the
the
model
zoo
that
you
have
in
deep
learning
and
in
machine
learning
and
then
advanced
methods.
They
actually
do
get
into
reinforcement,
learning
unsupervised
and
self-supervised
learning,
reinforcement,
learning
for
games
and
they're.
You
know
they
have
some
big
names
here,
coordinating
this,
so
they
have
ten
little
crap
and
blake
richards
and
jane
wong
and
joshua
vogelstein.
So
these
are
people
who
have
done
a
lot
of
work
on
different
things
in
the
field.
So
this
is
a
very
good.
A
You
know
if
you're
interested
in
following
along
with
this,
you
know
the
the
notes
will
be
available.
So
I
wanted
to
make
that
I
wanted
to
highlight
that
as
it's
going
along,
and
I
have
two
groups
of
mentoring
for
the
computational
neuroscience
course
and
they're
doing
one
one
is
doing
the
a
n
b
n
comparison
and
another
one
is
doing
a
neuro
imaging
project.
So
this
is
this
is
all
something
we
can
follow
up
on.
A
So
let
me
give
you
a
quick
update
on
diva
learn,
there's
not
much
going
on
but
like
as
as
minox
said,
he
was
going
to
update,
he's
going
to
do
a
diva
or
an
update
soon,
maybe
in
the
next
couple
weeks
and
do
a
push
to
the
repository,
and
so
you
know
we'll
have
I
don't
know
if
we'll
do
a
new
version
release
on
this,
but
we
can
you
know
we
don't
really
have
a
versioning
scheme.
So
if
it's
something
that
we
can
see,
I
just
wanted
to
see
what
we
are.
A
We
have
14
contributors
right
now
we
haven't
had
any
pull
requests
issued
since
the
gsoc
application
period,
which
is
fine.
It's
just
you
know,
that's
our
sort
of
our
rush
time
for
people
issuing
pull
requests
and
so
yeah
we're
sort
of
at
our.
I
can't
remember
what
version
we're
at
right
now,
zero
point.
A
A
A
Done
but
that's
okay
yeah,
so
there
are
some
things
outstanding,
like
getting
data
from
certain
people
and
recruiting
people
to
work
and
as
contributors
and
this
oh,
this
oiler
pants
for
life
paper.
So
there
are
there's
things
that
need
to
be
updated
presentation.
That's
done!
A
I
might
next
week
I
might
present
on
the
stuff
that
was
done
at
networks
2021,
but
I'm
not
quite
sure
I
didn't
have
time
this
week
to
do
it.
But
and
then
we
have
a
bunch
of
action
items
here.
So,
if
you're
interested
in
like
following
up
on
some
of
this,
if
you
want
to
contribute,
let
me
know:
if
you
have
things
to
add,
let
me
know
we
can
put
them
in
them
in
the
to
do
and
we
can
move
them
around
the
board.
A
So,
let's
see
we
have
some
things
in
the
chat.
Let
me
come
back
to
this
okay,
so
I
may
not
shared
his.
I
think
this
was
his
pull
request.
That
was
merged.
This
is
the
okay.
Actually,
sir,
hello,
sorry
I
joined
light
had
some
summer
school
procedures.
No
problem
he
actually
asked
might
now
cause
your
gsoc
project
going.
Is
it
going
good
so
far?
A
The
neuromatch
academy
course
seems
interesting
yeah
and
then
that's
this
week's
blog
post
for
minox
gsoc
progress.
This
is
week
five.
So
if
you're
following
along
and
week,
four
is
also
on
that
blog.
It's
just
a
different
length
week,
four,
instead
of
week
five.
So
if
the
blog
post
is
usually
just
a
recap
of
what
he
presents
here
so
see,
I
put
I
put
a
link
to
it
in
the
slack
as
well
on
one
of
the
channels
so
definitely
check
that
out.
C
Yeah
hello,
yeah
yeah,
so
I
actually
wanted
to
work
on
the
a
n
b
n
paper.
I
guess
this
is
a
favorite.
A
Yeah
we
we
have
a
pre-print
on
it,
but
it's
you
know,
we're
gonna
do
another
version
of
it.
I
can
loop
you
into
this
yeah.
A
Well,
I
mean,
I
guess
what
I
could
do
is
I
could
send
you
some
of
the
materials
that
we
have
now
and
then
you
could
give
some
feedback
on
that.
So
we
have
a
preprint
and
we
have
a
poster
and
I'll
send
you
those
links
after
the
meeting
and
then
you
can
maybe
give
some
feedback
on
what
what
you
think
might
improve
it
or
you
know
what
direction
we
might
go
and
it
might
you
know
yeah,
but
yeah,
so
it
might
fit
into
the
paper.
A
It
might
be
some
new
thing
that
we
do,
but
I'm
gonna
keep
driving
it
forward.
C
A
Yeah,
I
don't
know
where
she
is
on
that
I
think
she
was
busy
with
her
exams
and
her
yeah.
C
That's
fine
yeah,
but
you
you
can
send
me
the
links
like
on
slack
yeah.
A
A
Yeah,
I
will
all
right
and
then
yeah
anything
else
might
knock.
B
B
But
I
think,
having
us
a
one
single
release
that
would
which
would
have
all
the
upgrades
would
be
better.
I
guess
so
yeah,
that's
it.
Yes,
yeah.
A
Yeah,
I
think
so
too
I
mean
you
know
it's
like.
We
don't
really
have
a
versioning
scheme,
but
it
should
be
something
that's
consistent
like
you
know,
if
we
just
have
like
a
little
bit
of
an
upgrade,
it's
not
that
desirable
to
have
it
in
in
place
before
we
have
the
whole.
A
You
know
till
it
all.
It's
all
sort
of
integrated,
yes
yeah,
so
yeah,
we'll
we'll
figure
that
out
and
so
yeah,
but
but
everyone
be
on
the
lookout
I
mean
the
energy
sock
is
coming
up
sooner
or.
A
Later
and
then
I
yeah-
I
don't
know
what
we'll
do
in
terms
of
promoting
the
new
version.
I
think
it's
definitely
worth
like
kind
of
you
know.
I've
been
doing
these
presentations
to
different
groups
on
evilern
and
I
don't
know
how
many
people
are
really
getting
interested.
I
presented
an
incf,
so
I
don't
know
who
saw
it,
but
we
have
materials
and
I'll
do
maybe
another
short
update
presentation,
maybe
put
it
on
youtube
or
we
can
do
some
other
promotion
of
it.
A
So
you
know
just
I
guess
you
just
have
to
keep
getting
out
there
and
updating
and
people
will
start
to
use
it.
But
it's
it's
yeah.
So,
okay,
that's
good!
Now
let
me
go,
I'm
gonna
do
papers
and
I
have
some
things
papers
and
then
maybe
some
other
things
in
the
folder
as
well.
So,
let's
see
so,
we
have,
let's
see
so
I
have
this.
I've
never
visited
a
question
that
I
think
dick
brought
up
several
weeks
ago.
A
A
If
it's
just
like
me,
walking
down
the
street,
you
know
if
it's
a
sunny
day
or
if
it's
raining
there
are
different,
all
sorts
of
different
inputs,
depending
on
where
I
am
in
the
world
and
whatever
so
you
know,
that's
that's
how
it's
it's
very
vague
in
that
sense,
but
there
are
actually
two
things
we
came
up
with
after
that
question,
and
I
think
this
is
mainly
my
response,
but
I
think
we
talked
about
it
a
little
bit
offline.
A
The
first
is
that
environment
is
permissive
and
that
that
means
is
that
it
allows
for
things
to
happen
where
it
enables
things
to
happen,
and
so
general
information,
such
as
changes
like,
if
it's
really
hot
out
versus
not
really
hot
out
or
stresses,
which
is
this
heat.
But
it's
you
know
at
a
very
high
level
of
tolerance
for
the
organism
or
intensities
like
such,
as
you
know,
like
really
strong
light
or
really
high
temperatures.
Those
are
all
kind
of
interrelated
are
transduced
into
the
biological
system.
A
So,
for
example,
if
you
take
like
a
drosophila
egg
or
drosophila
set
of
drosophila
eggs,
and
you
raise
them
at
a
high
temperature,
you
can
get
them
to
mutate
in
different
ways
and
it
doesn't
mutate
the
genome
and
mutates
the
phenotype
there's.
This
thing
called
a
reaction
norm
that
happens
in
some
organisms
that
actually
changes
the
way
the
phenotype
is
expressed.
A
Another
example
is
like
neuroplasticity,
so
in
you
know,
in
like
a
small
mammal,
maybe
you
have
in
in
their
when,
as
they're
being
raised
after
they're
born,
they
have
to
interact
with
their
environment
to
kind
of
get
a
sense
of
the
world,
and
so
they
have
these
things
called
enriched
versus
impoverished
environments.
A
A
A
So
you
know
there
are
different
patterns
of
like
you
know
if,
if
there's
a
period
of
starvation,
for
example,
and
there's
no
input
of
food
or
energy
into
the
organism
that
can
affect
the
sort
of
the
expression
of
the
phenotype
in
development.
A
For,
moreover,
you
can
have
you
know
some
sort
of
you
know,
you
can
have
some
sort
of
information
in
the
environment.
A
You
know
like
light
source
that
can
affect
how
say
the
visual
system
develops
and
we'll
talk
about
that
in
a
minute,
but
that
you
know
those
are
all
things
that
are
sort
of,
and
this
is
again
at
a
very
general
level.
So
you
want
to
get
drill
down.
You
can
use
these
sort
of
two
points
to
kind
of
think
about
how
your
specific
system
develops
and
what
kinds
of
information
specifically
is
affecting
in
these
ways.
A
So
that
leads
us
to
this
paper
here
by
danny
nilsson
and
jokin
smoka
and
they're
biologists
and
they're
interested
in
this
idea
of
quantifying
biologically
essential
aspects
of
environmental
light,
so
the
environment
for
them
they're
focusing
and
are
just
light
and
they're
asking
the
question:
what
are
the
essential
biological
aspects
of
this
and
then
how
do
we
quantify
this?
So
we
can
measure
it
measure
the
differences
in
intensity
or
an
amount,
and
how
does
that
affect
the
developmental
process?
A
Or
I
guess
in
this
case,
just
generally
so
the
abstract
is
quantifying
and
comparing
light
environments
are
crucial
for
interior
lighting
architecture,
visual
ergonomics.
So
these
are
things
in
the
human
world,
but
you
can
see
how
they
they
they're,
also
important
in
animal
development
and
plant
development
and
so
on.
A
Yet
current
methods
only
catch
a
small
subset
of
the
parameters
that
constitute
a
light
environment,
an
early
account
for
the
light
that
reaches
the
eye.
Here
we
describe
a
new
method,
the
environmental
light
field,
method
or
elf,
which
quantifies
all
essential
features
that
characterize
a
light
environment,
including
important
aspects,
have
previously
been
overlooked.
A
The
elf
method
uses
a
calibrated
digital
image
sensor
with
wide
angle,
optics
to
record
the
radiances
that
would
reach
the
eyes
of
people
in
the
environment,
so
here
they're
talking
about
radiances
and
they're,
actually
specifically
focused
on
how
we
perceive
light
with
our
eyes
now.
This
is
interesting
in
development,
because,
although
that
could
have
a
developmental
relevance,
you
know
a
lot
of
the
things
in
development
involve.
You
know.
A
Maybe
exposure
like
in
plants
exposure
to
light
the
angle
of
the
light,
the
radiance,
the
the
frequency
of
the
white
or
the
color,
and
then
you
know
also
heat
energy.
So
there
are
a
lot
of
things
we
can
quantify
and
it's
maybe
even
broader
in
terms
of
development,
but
you
know
in
in
visual
development,
where
you
know,
when
you're
trying
to
figure
out
what
to
look
at
in
development
as
a
child
or
in
your
developmental
period.
This
is
also
important
in
the
development
of
eyes.
A
So
so
they
use
these
techniques,
these
measurement
techniques
to
sort
of
get
a
sense
of
what
the
white
is
and
then
what
you
know
optimal
light
is
and
then,
as
a
function
of
elevation
angle.
It
quantifies
the
absolute
absolute
photon
flux,
its
spectral
composition
in
the
red,
green
blue
resolution,
as
well
as
its
variation
or
contrast
span.
So
now,
they're
breaking
down
light
into
these
different
aspects
of
its
spectral
properties,
and
so
now
they're
trying
to
make
a
statement
and
people
do
this
with
plants
I
think
as
well
or
in
in
some.
A
Well,
they
don't
do
it
in
like
exposure
to
like
light
radiation
or
something
like
that,
but,
like
they'll,
do
I
think
they're
in
if,
in
fact,
in
photosynthesis,
the
spectral
composition
is
important
for
driving
that
process.
A
But
you
know
that's
something
that
people
measure,
but
this
is
just
the
way
that
I
kind
of
think
about
how
to
quantify
this
together.
These
values
provide
a
complete
description
of
the
factors
that
characterize
a
light
environment,
the
eof
method.
This
offers
a
powerful
and
convenient
tool
for
the
assessment
and
comparison
of
light
environments.
So
you
can
compare
light
environments.
A
We
also
present
a
graphic
standard
for
easy
comparison
of
light.
Environments
show
that
different
natural
and
artificial
environments
have
characteristic
distributions
of
light.
So
this
is.
This
is
important
for
visual
ecology,
environmental
psychology,
lighting
science,
that's
sort
of
their
intended
audience
and
then
they
kind
of
get
into
the
mathematics.
They
actually
look
at
the
cie
palette,
which
is
what
a
lot
of
people
use
to
look
at
colored
differences.
A
So
when
you
think
of
like
how
to
define
a
color
it
like
in
a
in
a
computational
image,
often
times
we
use
rgb
but
there's
a
cie
standard,
which
is
like
a
color
map
which
breaks
it's
like
the.
I
think
it's
the
subtractive
aspect
of
color
instead
of
the
additive,
but
those
sorts
of
things
still
have
problems.
A
So
another
problem
with
current
methods
for
measuring
environmental
light
comes
from
a
standardized
spectral
sensitivity
on
which
units
such
as
lux
and
candela
are
based
in
lighting.
The
spectral
sensitivity
is
generally
taken
as
representing
human
vision,
although
it
is
in
fact
only
the
spectral
sensitivity
of
the
retinal
channel
for
achromatic
contrast
measured
in
the
retinal
center.
It
ignores
the
blue
cones
in
the
retinal
channel
for
channels
for
color
vision.
A
A
So
there's
this
they're
they're
kind
of
making
the
point
that,
like
the
methods
that
we
have
now
are
really
you
know
not
up
to
the
task
of
what
we
want
to
do
to
characterize
environment.
So
there
are
a
lot
of
things
here
that
they
kind
of
go
through.
They
go
through
their
own
approach
to
this,
and
then
this
is
the
single
scene,
and
this
is
how
they
kind
of
show
this
visual
space
or
this.
A
This
light
space
they're
doing
these
they're
doing
these
extractions
and
they're
building
images
out
of
them,
so
they're
quantifying
quantifying
environmental
light
with
the
environmental
light
field
method,
so
they're
acquiring
data
with
a
camera
they're
using
a
fisheye
lens
they're,
using
the
mounted
bubble
to
standardize
the
images
and
then
they're
generating
these
images
that
are.
C
A
Oh
well,
I'm
not
sure,
maybe
I'll
ask
you
mean
just
in
general,
like
on
the
different
images
that
she's
acquiring.
A
A
So
yeah
that
does
curve
that
does
provide
curvature
to
the
image
in
some
ways.
So
there's
that
aspect
of
it
yeah.
So
then
this
is
the
so
they're
going
to
go
over
their
method.
Here
they
show
some
results
and
then
that's.
Their
argument,
for
the
paper
is
that
this
is
how
you
can
quantify
the
visual
world
or
the
light
world
in
in
this
way.
So
this
is
an
interesting
paper.
A
I
wouldn't
have
thought
to
approach
it
this
way,
but
it's
definitely
one
so
you
know,
maybe
in
the
future,
we'll
find
more
papers
like
this
thinking
about
other
aspects
of
the
environment
that
we
can
quantify,
and
it's
really
good
to
go
through
that,
like
critical
view
of
it.
A
Another
thing
that
we
have
here
is
this
paper
on.
This
is
a
new
paper.
I
put
it
in
the
if
you
go
to
the
slack
and
you
go
to
the
new
papers,
tab
or
I
think
I
think
it's
new
papers
or
current
papers
or
something
like
that.
There's
a
channel
where
we
feature
papers.
I
put
this
paper
in
this-
is
towards
a
living
soft
micro
robot
through
optogenetic
locomotion,
control
of
c
elegans.
A
So
it's
like
this
is
by
a
chinese
group
and
they're
doing
this
work
with
they're
trying
to
build.
I
guess
a
a
microbot
based
on
c
elegans
motion,
so
the
abstract
breathes.
Learning
from
the
locomotion
of
natural
organisms
is
one
of
the
most
effective
strategies
for
designing
microbots,
and
these
are
small
micro
robots
that
you
know
move
they
want
to
use
biological
mechanisms
to
move
them
because
those
are
well
established
at
that
scale.
A
However,
the
development
of
bio-inspired
micro
robots
is
still
challenging
because
of
technical
bottlenecks
such
as
design
and
seamless
integration
of
the
actuation
mechanisms
and
the
energy
source.
So
they
have
to
you,
know,
put
a
battery
and
they
have
to
make
the
accurate
actuation
work
at
that
scale.
It's
very
hard
to
get
those
things
well,
design.
You
know
design
them
pretty
well
and
get
them
optimized,
so
they
want
to
directly
harness
the
activation
energy
and
intelligence
of
living
tissues.
A
A
A
So
a
living
worm
is
engineered
through
optogenetic
and
biochemical
methods
to
shut
down
the
signal
transmissions
between
its
neuronal
and
muscular
systems.
What's
muscle,
cells
cells
still
remain
optically
excitable,
so
they're
using
these
up
genetic
either,
basically
trying
to
shut
down
certain
parts
of
like
the
neural
circuit,
the
connection
between
the
neural
circuit
and
the
muscles
and
then
they're
trying
to
measure
this
and
then
transform
transfer
that
information
into
the
into
the
robot
through
dynamic,
modeling
and
experimental
verification
of
the
worm
crawling.
A
So
optogenetic
excitation
is
where
they
take
light
of
a
certain
frequency
and
they
stimulate
the
tissue
where
they
stimulate
the
neurons,
depending
on
how
they
use
it,
and
these
optogenetics
in
the
brain
and
there's
a
they.
They
inject
a
protein,
it's
a
light-sensitive
protein
and
they
put
it
into
the
tissue
that
they
want
to
measure
from
or
the
neurons,
and
then
they
use
this
optogenetic
beam.
That
is
at
a
certain
frequency
that
stimulates
this
protein.
A
A
So
you
can
actually
turn
on
and
off
neurons
or
turn
on
and
off
muscle
in
a
very
controlled
way,
and
so
so
they're
using
optogenetic
excitation.
We
emulated
the
major
worm
crawling
behaviors
in
a
controllable
manner.
Furthermore,
with
real-time
visual
feedback
of
the
worm
crawling,
we
realized
closed
loop
regulation
in
the
movement
direction
and
destination
of
single
worms.
A
This
technology
may
facilitate
more
information
about
the
biophysics
and
neural
basis
of
crawling
locomotion
and
c
elegans.
A
They
they're
using
this
optogenetics
technique
to
sort
of
get
up
how
the
worm
moves,
decoupling
muscle
from
the
neural
circuit
and
then
kind
of
reverse
engineering.
That
and
then
be
you
know,
building
this
into
the
robot.
So
it's
it's
a
really
interesting
approach,
let's
see
if
they
have
any
okay.
So
this
is.
This
is
how
they
deliver
the
optogenetic
signal.
They'll
use
a
like
a
light
source
and
a
microscope.
A
They'll
kind
of
you
know,
hit
different
parts
of
the
tissue
where
they've
injected
the
protein
and
then
they'll
stimulate
those
areas
with
that
light
frequency.
So
you
can
see
that
there's
they
have
their
experimental
setup
and
then
they
show
how
it
moves
with
the
laser
beam
activation,
and
then
they
have.
This
thrust
force
analysis.
A
And
then
they're
inducing
movements
and
then
they're
able
to
take
that
take
that
information
and
incorporate
it
into
the
design
of
this
robot.
It's
really
interesting
work.
I
I
don't
know
much
more
about
it.
I'm
not
really
deeply
familiar
with
optogenetics
other
than
I
know,
basically
what
it
does,
but
this
is
interesting
and
then
they
actually
move
the
rover
worm
through
a
maze
so
they're
able
to
build
this
maze
and
move
the
robot
worm
using
this
information.
A
So
this
is
a
paper
on
deep
learning
but
applied
not
to
embryos,
but
pluripotent
cells.
So
this
is
so
they
they
come
up
with
on
the
summary.
Lineage
tracing
is
a
powerful
tool
in
developmental
biology
to
interrogate
the
evolution
of
tissue
formation,
but
the
dense
three-dimensional
nature
of
tissue,
and
this
is
like
in
in
vivo.
This
is
like
your
muscle,
tissue
and
c
elegans,
for
example,
limits
the
ability,
the
assembly
of
individual
cell
trajectories
into
complete
reconstructions
of
development.
A
A
That
the
the
nuclei
are
visual,
you
know
visualized
and
then
they
contract
the
cells
using
that
that
visual
signal.
So
they
they
put
some
sort
of
fluorescent
marker
in
the
nucleus
and
they
track
the
nucleus
around
the
embryo
as
it's
moving
and
they
require
the
image
and
slices
so
they're,
putting
the
slices
together
they're
using
those
that
provenance
to
sort
of
tell
you
know
where
in
the
embryo,
these
things
are
and
then
you.
C
A
A
A
So
you
don't
have
to
have
it
in
a
tissue,
and
you
have
the
cells
just
in
this
in
this
monolayer
and
you
can
observe
these
processes
of
change,
and
so
it
can
recapitulate
aspects
of
developmental
processes
providing
an
in
vitro
platform,
and
that
means
outside
the
body.
That
means
in
a
dish
to
assess
the
dynamic,
collective
behaviors
directing
tissue
morphogenesis
it
reaching
an
ensemble
of
neural
networks
to
track
individual
hipscs
in
time-lapse,
microscopy
generating
longitudinal
measures
of
cell
and
cellular
neighborhood
properties
on
time
scales
for
minutes
to
days.
A
Our
analysis
reveals
that,
while
individual
cell
parameters
are
not
strongly
affected
by
pluripotency
maintenance
conditions,
which
is
where
the
cells
have
when
they
transform
to
these
pluripotent
cells,
which
means
that
they
can
you
know
it's
a
state
from
which
they
can
take
another
number
of
paths.
So
a
pluripotent
cell
is
something
that
is
like
a
stem
cell.
So
it
doesn't
really
have
a
fate,
a
terminal
fate,
but
it
can
become
like
a
muscle
cell
or
a
brain
cell,
depending
on
where
you
know
where
it
is
in
its
pluripotency.
A
So
today
potent
cell
is
a
cell
that
can
reconstruct
an
organism.
It
can
just
become
any
cell
type
and
if
a
totipotent
cell-
and
this
is
like-
maybe
one
of
the
four
or
eight
cells
and
first
four
or
eight
cells
in
the
embryo-
if
they
divide,
if
they
divide
enough,
they
can
differentiate
in
recreating
an
entire
organism.
A
A
pluripotent
cell,
on
the
other
hand,
has
a
restricted
set
of
fates,
but
they
can
still
become
a
number
of
different
things,
depending
on
where
they
are
and
they're
in
the
cell
lineage.
So
you
know,
oftentimes
you'll
have
pluripotent
cells
that
become
like
a
number
of
different
neural
cell
types
like
neurons
or
muscle
or
glia.
A
They
won't
necessarily
become
like
say,
liver
cells
or
germ
cells,
but
they
can
make
they
can
have,
they
can
take
on
a
number
of
fates,
and
so
so,
okay,
so
we
were
here.
Regional
changes
in
celebrate,
behavior
predict
sulfate
and
county
organizations.
So
now
they're.
Looking
at
regional
changes,
which
I
assume
they
mean
spatial
change,
spatial
changes
or
what
their
neighbors
are,
or
whatever,
by
generating
complete
multicellular
reconstructions
of
hi
psc
behavior,
our
tracking
pipeline
enables
fine-grained
understanding
of
morphogenesis
by
elucidating
the
role
of
regional
behavior
in
early
tissue
formation.
A
So
this
is
so
in
the
paper
here.
They
talk
about
a
little
bit
about
in
vitro
versus
in
vivo
models,
so
they
talk
about
the
c
elegans
data
set
here.
This
is
about
2006
data.
This
is
basically
a
very
similar
thing
to
the
data
that
the
data
set,
that
mynac
is
working
with,
so
automated
tracking
of
cell
migration
within
whole
embryos
in
vivo
has
been
limited
both
in
size
to
small
organisms
such
as
c
elegans.
A
So
again
we
work
with
small
embryos,
because
that's
easy,
but
larger
embryos
are
much
harder,
so
they
actually
get
into
the
idea
of
turing
patterns
from
morphogenesis,
and
we've
talked
about
that.
Where
you
have
a
bunch
of
cells
that
get
signals
from
the
environment
and
depending
on
what
signals
they're
exposed
to
they
form
different
types
of
you
know
the
difference
in
different
types
and
then
you
get
boundaries
between
signals
and
that's
where
you
get
like
stripes
and
other
things
where
you
have
this
different.
A
You
know
you
have
these
gradients
that
are
different
yeah
this
differentiating
capacity,
and
this
is
how
you
get
patterns
in
morphogenesis.
So
it's
not.
You
know
it's
something
that
we're.
We
always
try
to
model,
but
it's
it's
hard
to
observe
in
in
real
biological
systems.
A
So
similar
multicellular
organizational
events
have
been
observed
in
vitro
with
human
induced
pluripotent
stem
cells.
So
we've
observed
these
kind
of
turing
patterns
in
pluripotent
stem
cells,
so
we've
modeled,
we
we've
seen
examples.
I
think,
in
the
g
stock
period
how
these
can
be
modeled
using
cellular
automata,
and
you
can
actually
see
these
in
these
pluripotent
stem
cell
cultures
revealing
their
heterogeneous
differentiation
potential
due
to
global
positional,
cues
cell
population
boundaries
or
cell
cell
interactions.
A
A
The
differentiation
process
doesn't
happen,
randomly
they
happen
in
these
colonies
and
the
colonies
form,
and
so
you
can
see
and
actually
pick
colonies
of
these
reprogrammed
stem
cells
and
transfer
them
to
another
place
that
well.
You
know
that
the
implication
there
is
that
they're,
just
these,
they
exhibit
these
kind
of
turing
patterns
where
they
have.
A
You
know,
instead
of
being
at
random,
all
over
the
place
they're
in
these
colonies
that
are
defined,
they're,
not
stripes,
but
they're
kind
of
like
blobs
and
you'll,
see
them
in
a
it's
very
recognizable
in
particular,
because
cell
phase
and
function
are
strongly
influenced
by
local
interactions
within
multicellular
networks.
Coordinated
morphogenetic
processes
exhibit
scale-free
connectivity.
Oh
that's
interesting!
They
go
into
that
area.
A
Cell
behavior
is
coordinated
through
a
central
hub
of
influential
cells,
indicating
that
pop
small
populations
of
cells
established
by
highway
connected
organizing
centers
can
exert
a
large
impact
on
the
final
composition
of
the
developing
tissue
so
actually
get
into
networks
here,
which
I
didn't
expect
them
to.
But
that's
interesting
in
light
of
the
work
we
have
on
embryo
networks.
So
we'll
put
that
aside
for
now,
but
recent
advances
in
machine
learning,
particularly
deep
neural
networks,
would
have
demonstrated
superhuman
performance
and
image
segmentation.
So
they
go
through
that
it
looks
like
they're
using.
A
I
guess
they
talk
about
cnns
units.
So
what
are
they
using
here
looks
like
they're
using
a
unet
cnn,
because
that's
what
they're
reviewing?
Oh!
No
in
this
study,
we
overcame
the
challenges
of
dense
cell
tracking
by
developing
an
ensemble
of
three
neural
networks,
fcrnb
conception
and
a
residual
unit.
A
Nuclei
displacements
were
then
connected
between
sequential
frames
of
a
time
series
enabling
high
spatiotemporal
resolution
of
hipsc
behaviors
over
relevant
developmental
time
scales
of
hours.
To
days
this,
then
cell
tracking
pipeline
revealed
distinctive
cell
behaviors
based
on
localization
or
location
within
the
colony,
so
it
had
originated
in
response
to
extracellular
signaling
molecules,
so
they
were
able
to
track
this.
They
used
human
annotators
to
sort
of
score
the
selected
nuclei,
so
they
actually
used
human
annotators
on
the
loop
here.
A
So
I
think
they
identified
things
through
tracking
and
then
they
did
some
manual
annotations
of
of
these
cells.
So
that's
interesting,
I
don't
know
like
human
annotation
is
both
good
and
bad
for
a
number
of
reasons,
but
they
don't
necessarily
have
labeled
data
here.
So
that's
what
they're
gonna
have
to
work
with,
and
so
it's
it's
interesting
to
see
how
they're
dealing
with
this
problem.
A
So
this
is
an
example,
so
they
have
aggregate
cells,
they're
unlabeled.
They
have
some
that
are
unlabeled
and
some
that
have
the
gfp
label
right.
So
they
have
labels
for
the
colonies
and
these
are
the
colonies,
so
the
colonies
will
have
labels
in
them.
A
The
cells
that
don't
reprogram
along
the
edges
won't
have
labels,
so
you'll
see
you'll,
be
able
to
identify
the
cells
in
this
colony.
But
then,
within
the
colony
we
don't
know,
there's
no
labels
to
the
data.
We
don't
know
anything
about
these
cells
other
than
that
they're,
either
pluripotent
or
they're.
Not
so
you
image
these
and
you
see
these.
A
You
can
see
the
fluorescence
here
and
so
and
then
in
this
case
they
show
that
you
have
human
raiders
that
are
actually
picking
out,
because
sometimes
you
get
things
like
autofluorescence,
so
some
cells
that
aren't
pluripotent
will
have
fluorescence.
A
You
have
to
use
a
threshold
because
sometimes
you
know
the
label
will
express,
but
it's
incompletely
pluripotent.
So
there
are
a
lot
of
things
going
on
in
these
data
that
are
hard
to
really
understand,
especially
when
you
don't
have
labels
and
so
the
we
have
the
human
readers
sort
of
creating
labels
for
them
and
annotating
them.
A
They
call
them
heterotypic,
neural
network
ensembles
and
so
they're
able
to
generate
human
quality
segmentations,
so
they're,
actually
segmenting
them
and
then
they're
actually
doing
the
human
radar
stuff.
So
they
have
the
humans,
like
kind
of
labeling
that
and
they're
using
this
ensemble
model
and
they're
matching
the
results.
A
Okay,
so
now,
let's
see
if
they
have
anything
else,
this
is
a
good
paper.
I
don't
know
okay,
now
they're
doing
spatio
temporal
linkage
of
detections,
and
this
enables
long-term
single
cell
tracking
so
now
they're
doing
this
they're
sort
of
building
a
model
of
a
migration
and
they're
getting
this
they're
doing
a
delaney
triangulation,
which
is
like
a
graph
that
they
build.
A
This
is
sort
of
a
spatial
graph
that
they
use
to
determine
the
spatial
relationships
between
cell
nuclei.
So
this
is
the
dilani
triangulation.
It's
just
like
a
network.
That's
it's
a
graph
that
they
use
to
determine
the
sort
of
it's
like
a
series
of
triangles
basically,
and
it
determines
a
spatial
like
sort
of
the
coincidence
of
of
the
cells,
how
close
they
are
in
space,
so
they
were
able
to
calculate
cell
neighborhoods
from
these
dalani
triangulations.
A
A
Data,
segmented
and
then
they
were
able
to
look
at
packing
and
migratory
behaviors
of
undifferentiated
pluripotent
stem
cells
in
these
colonies.
So
now
they're
able
to
do
we
compared
standard
pluripotency
maintenance
conditions
using
the
cnn
tracking
algorithm.
This
is
to
look
at
heterogeneous
behavior
of
the
colonies.
A
It
and
so
then,
they're
also
able
to
look
at
lineage
tracing
of
selfie
decisions
or
an
early
morphogenetic
induction,
and
they
show
that
here
they
look
at
the
density
of
different.
So
the
other
thing
you
can
do
is
look
at
different
marker
genes,
so
opt
for
eomes
and
sox2
are
all
marker
genes.
They
kind
of
tell
you
something
about
how
pluripotent
in
the
cell
is,
if
they're
expressing
high
levels
of
oct4
or
sox2,
that's
very
pluripotent.
A
If
it's
expressing
very
little
oxford
stocks
too,
that's
less
pluripotent,
and
so
using
this
technique,
you
can
use
different
markers.
You
can
use
like
a
fluorescent
marker
for
oct4
or
sox2,
and
you
can
actually
get
an
assessment
of
the
level
of
of
expression
of
those
genes
because
you
use
like
a
you,
know,
a
control
and
you
can
actually
get
the
full
difference
in
expression.
So
you
just
you
know,
use
a
an
index.
Then
you
can
plot
it
out
in
a
way.
A
That's
you
know
you
can
use
that
data
to
plot
out
not
only
the
amount
of
expression
but
their
movement
in
other
migration
and
their
position
in
space.
So
that's
kind
of
what
they're
doing
here,
they're
kind
of
like
giving
you
know,
looking
at
a
time
course
they're
looking
at
these
different
properties
of
the
cell
they're,
looking
how
much
they're
expressing
these
marker
genes
and
then
they're
looking
at
how
they
move
around
the
colony.
A
A
Advanced
stuff
compared
to
when
I
was
working
on
this
yeah,
so
that's
all
we
have
for
that
paper,
and
I
would
definitely
look
at
that
if
you're
interested
in
how
people
are
integrating
these
different
markers
and
and
machine
learning
and
microscopy
data,
so,
okay,
so
can
you
share
the
link
to
the
deep
end
and
pluripotency
paper?
A
Well,
let
me
share
the
meeting
folder
and
then
I
will
also
put
these
up
in
the
slack.
A
So
that's
the
folder,
where
the
deep
end
and
pluripotency
paper
is
and
I'll
put
the
individual
paper
in
the
slack.
When
I
get
the
link
to
the
thing.
Okay,
thanks.
Well,
thank
you
for
attending
my
knock
and
knock
shay
was
here.
Thank
you,
akshay
for
attending
and,
if
you're
on
youtube
I'll,
be
putting
the
links
to
the
papers
in
the
slack
and
the
recording
and
thank
you
for
attending
and
have
a
good
week
see
everyone
next
week.