►
From YouTube: DevoWorm (2023, Meeting #23): Towards a biological embryo model, GSoC, TDA/GNNs and Embryo Networks
Description
GSoC Updates: Segmenting cells for Topological Data Analysis, DevoLearn maintenance, DeepLabv3 vs SAM models. Recent developments in biological embryo models (beyond cell reprogramming and embryoids). Presentation on persistent homology, lineage trees, and connective anatomy. Attendees: Sushmanth Reddy Mereddy, Jiahang Li, Himanshu Chougule, Bradly Alicea, and Jesse Parent.
A
B
Are
you
yeah
I'm
fine?
How
are
you
pretty
good,
pretty
good,
welcome
to
the
meeting
well
Jesse
so
come
on.
Did
you
want
to
give
your
update.
A
Okay,
so
my
update
of
this
week
is
like
I've,
been
working
on
the
model
that
I
proposed
last
week.
Damascus
yearn
model,
so
one
of
the
things
that
I
was
trying
to
I
was
doing
is
the
creating
the
bounding
boxes
for
all
the
cells
that
are
present
like
in
the
data
set.
A
So
I've
worked
on
that
as
well
as
I
tried
to
do
it
in
some
different
way,
like
Toyota,
without
using
a
function
and
doing
it
on
my
own
to
see
how
it
how
I
can
make
a
better
way
of
cutting
the
bounding
boxes
and
other
than
that
I
have
been
working
with
this
data
set,
which
was
always
also
used
in
the
previous
year
G-Shock,
as
well
as
the
new
data
set
which
are
shown
last
week.
So
the.
A
A
Like
also
inference
through
this
was
like
I
researched
more
about
our
data
set
and
basically
the
data
is
about
the
first
larva
stage,
L1
of
S3
elegance,
and
they
have
used
a
cons,
open
image,
stack
of
15
different
individuals
as
well
as
they
have
inference
three
and
segmented
and
allocated
558
developed
568
annotations
at
the
developmental
stage
out
of
the
357
nuclear
data
present
and
other
one,
and
they
have
also
done
like
annotations
manually
as
well
as
use
the
computational
method
on
the
data
set
as
well.
A
So
after
completing
the
model
I'll
be
running
it
on
this
as
well,
and
I
also
been
reading
about
temperature
data
analysis
and
getting
up
to
advertise
with
that
as
well,
because
I'd
be
working
on
on
it.
A
The
few
the
next
few
weeks,
like
the
one
paper
which
you
had
also
given
in
the
slack
Channel
the
paper
which
was
referenced
by
jahang
as
well,
about
the
temporal
data
analysis,
which
was
done
on
C,
elegans
and
I've,
also
read
this
tracking
everything
and
everywhere
all
at
once,
which
was
like
something
which
month.
B
Good
so
yeah
you
have
you're
working
on
the
data
sets
you're
getting
established
for
starting
data
analysis
once
you
get
the
segmentation
part
done
and
then
your
so
you're
on
your
way,
you're
doing
yeah,
yeah
I.
Think
that's
good!
Any
roadblocks
coming
up.
Do
you
think
or
do
you
think
you're
on
the
right
track?
As
far
as
you
can
tell.
B
Okay,
what
is
your
timeline
for
sort
of
getting
to
the
analysis
stage.
A
I'd
be
finishing
with
the
segmentation
instant
segmentation
part
until
this
week
and
hopefully
show
you
it
in
the
next
week.
So
that
was
the
original
timeline
and,
let's
see
I
I
guess
as
well
as
then,
and
the
way
I
had
given.
The
timeline
was
like
the
first
four
weeks
for
segmentation,
the
next
four
weeks
for
the
day
of
last
part
like
so,
there
are
better
graphic
meetings
and
the
next
like.
A
B
A
Now,
right
now,
I'm
just
using
torch
machine
and
all
the
features
it
gives
us
so
like
for
bounding
boxes
like
we
have.
A
tutorial
has
run
for
like
a
util
function
for
it
and.
A
Like
I've
been
using
torch,
Vision
now
only
yeah,
also
like
pil
and
other
libraries
for
like
image
looking
at
the
image
and
so
on.
B
Do
it
yeah,
so
that
looks
good
yeah.
Thank
you
for
the
update
and
you're
finding
those
data
useful,
so
the
data
from
the
larval
stages,
yeah
yeah,
so
yeah.
That's
that's
really
good!
Thank
you
for
that
now,
I
I
got
a
message.
C
B
You
go
so
yeah
I
got
a
message
from
Gia
Hong
this
week
and
he
was
interested
in
doing
a
paper
on
some
of
the
things
he
did
last
summer
and
some
of
the
things
you're
doing
this
summer
and
I
said
yeah
I'd
be
great,
so
I
don't
know.
If
he's
gotten
in
touch
with
you
about
that,
but
he's
interested
in
doing
a
this
is
a
little
bit
longer
term
than
you
know
in
the
few
next
few
weeks.
B
A
A
Yeah
also
John
didn't
mention
that
this
yet
I
guess
he
would
be
mentioning
with
this.
D
A
Was
that
John
didn't
DM
us
yeah
like
me
and
sushma,
so
probably
will.
B
Be
mentioning
us
this
week,
oh
yeah,
yeah
I,
don't
know
if
we
didn't
get
around
to
it
so
yeah
but
yeah
we'll
be
talking,
I,
don't
know.
Maybe
hopefully
it
comes
in
a
little
bit
later.
We
can
talk
about
it,
but
if
not
we'll
just
talk
about
in
the
slack
yeah.
B
Sounds
good
yeah,
yeah
good!
Thank
you.
Okay,
all
right
so
yeah
like
I
said
switch
month
is
out.
He
had
he
has
his
exams
he's
you
know
doing
well.
I
gave
an
update
and
we
have
a
Friday
meeting
on
open
source
and
that's
usually
for
the
orthogonal
lab
Group,
which
is
a
different
group
they're
doing
different
projects
we
usually
meet
then,
and
talk
about
like
their
projects
and
I,
usually
give
some
updates
on
like
open
source
topics,
and
so
that's
actually
on
YouTube.
You
can
look
it
up.
B
The
orthogonal
research
and
education
laboratory,
YouTube
channel
so
I
mean
that's,
that's
something
if
you,
you
feel
free
to
attend
if
you
want,
but
we'll
be
doing
your
main
updates
in
this
meeting
for
the
diva
world
of
people
so
just
to
clarify
that,
like
I
said,
Gia
hung
said
he
might
be
able
to
make
it
later
in
the
meeting.
So
we'll
reserve
some
time
for
him
to
do
that.
B
Let's
see
I'm
going
to
share
my
screen
now
and
I'll.
Give
a
talk
on
a
few
topics
here.
So
I
think
I
mentioned
this
in
the
meeting
in
the
Saturday
meeting
for
the
orthogonal
web.
B
That
I
was
going
to
go
over
some
papers
on
human
making
human
embryos,
which
is
there,
have
been
a
couple
papers
that
have
come
out
recently
on
this
I'm
gonna
go
over
the
papers,
the
understanding
that
I'm
not
like
I,
don't
know
a
lot
about
this
I
mean
I,
know
about
embryos
and
I,
know
about
like
organoids
and
you
know,
reprogramming
cells.
But
you
know
this
is
definitely
something
that
is
okay.
B
So
how
much
you
said:
I'll
try
to
attend
Friday's
meeting
again,
yeah
it'd
be
great,
but
it's
definitely
like
sort
of
a
Cutting
Edge
area.
So
this
this
was
post.
We
have
a
channel
in
our
orthogonal
lab
slack,
which
is
brain,
organized
Channel,
and
so
we
posted
these
in
there.
B
This
is
where
we
monitor
a
lot
of
the
organoid
work
as
well,
so
this
first
paper
is
called
transgene
free,
x,
utero,
derivation
of
a
human
post,
implantation
embryo
model
solely
from
genetically
unmodified
naive
pscs
or
like
stem
cells,
basically
we'll
get
into
what
that
is
in
a
little
bit.
This
is
a
number
of
authors.
Jacob
pan
out
who's
done
a
lot
of
work
on
embryos
in
implantation
in
the
past.
So
this
is.
B
This
is
the
first
paper,
the
abstract
reads:
our
ability
to
study
early
human
post-implantation
development
remains
highly
limited
due
to
the
ethical
and
Technical
challenges
associated
with
this
kind
of
development,
so
the
difference
between
C
elegans
and
humans,
of
course,
is
that
c
elegans
develops
in
an
egg
and
humans
develop
in
a
in
an
amniotic
sac.
Or
what
will
you
know?
B
B
Reprogramming,
which
is
where
they
take
a
somatic
cell
and
they
put
in
a
nucleus
from
or
they
take,
a
stem
cell
or
a
germ
cell,
and
they
put
in
the
nucleus
of
a
somatic
cell.
So
this
is
the
the
egg
or
the
ovum,
and
then
they
have
an
it
has
a
nucleus.
They
take
that
nucleus
out.
Then
they
take
a
somatic
cell
from
the
from
an
adult.
Let
me
put
that
in
a
in
a
Gammy
and
then
that
now
has
a
genome,
that's
fully
adult,
but
it's
in
a
germ
cell.
B
So
it's
reprogramming
the
cell
that
gets
implanted
into
you,
know
a
uterus
and
it
grows
to
like
a
a
clone.
Basically.
So
this
is
the
human
cloning
work
that
is
done
probably
about
25
years
ago.
Now,
first
done
and
it's
been
they've
been
able
to
do
it
ever
since,
if
you
think
about
something
like
cellular
reprogramming,
which
is
where
you
take
a
self
like
a
maybe
a
skin
cell,
so
the
same
type
of
differentiated
cell
and
you
introduce
some
viral
Vector
with
some
factors.
B
C
B
Just
infects
the
cell
and
it
goes
into
the
nucleus
and
it
inserts
itself
in
the
genome,
and
then
from
that
you
get
this.
Basically
the
stem
cell.
It's
not
really
a
stem
cell
like
we
might
find
in
like
a
an
embryo,
but
it's
a
stem
cell.
It
basically
expresses
all
the
markers
of
a
stem
cell.
So
that's
that's
a
directory
programming.
This
is
indirect.
B
Which
means
that
it's
kind
of
here
putting
you
know
you're
putting
the
genome
in
the
cell
and
it's
sort
of
it's
not
integrating
it's
just.
Turning
on
some
genes,
direct
means
that
you're
putting
genes
in
the
Genome
of
a
cell
and
you're
sort
of
directly
programming
its
function.
So
it's
that's
the
distinction,
and
so
you
know
we
have
human
development.
We
have
something
like
CL
against
C
elegans,
of
course,
as
an
egg,
where
you
have
cells
that
divide
two
cell
stage
or
cell
stage
eight
cell
stage,
and
that
in.
B
B
This
is
true
of
a
lot
of
mammalian
cells,
but
a
lot
of
mammalian
embryos,
and
then
this
embryo
has
to
implant
itself
in
the
uterus
of
the
mother,
and
the
successful
implantation
leads
to
a
production
of
a
viable
embryo
that
becomes
a
young
organism
which
is
not
a
clone.
So
this
is
the
what
they're
talking
about
here
and
they're
talking
about
this
implantation
step
and
then
post
implantation,
which
is
where
it's
actually
developing
in
the
context
of
like
the
basically
it's
being
supported
by
the
mother.
B
So
there's
a
lot
of
biology
here
that
I'm
not
going
to
get
into
so
it's
it's
important
to
recapitulate
this
step.
So
what
we
can
do
now,
of
course,
is
we
can
create
organoids.
Of
course,
we've
talked
about
organoids,
and
you
know
it
could
be
something
like
you
know.
B
You
can
get
cell
signaling
between
different
cell
types
and
it's
so
that's
what
you
know.
That's
the
organoid
work
now
this
is
actually
creating
embryos
and
so
they're
talking
about
some
of
the
work
that
they're
doing
here
so
they're
able
to
sort
of
you
know
they
need
to
recapitulate
this
early
cellular
environment.
They
need
to
recapitulate
this
reproductive
environment,
and
so
how
do
they
do
that?
That's
been
the
big
stumbling
block
to
you
know
being
able
to
basically
clone
organisms
so.
C
B
Can
do
this
with
like
sheep
and
goats
and
cows,
and
things
like
that
they've
been
able
to
do
indirect
reprogramming,
but
a
lot
of
this
is,
you
know
hard
to
do,
especially
in
humans,
and
so
you
need
to
be
able
to
recapituate
post-implantation,
Epi
blast
development
within
the
context
of
these
extra
embryonic
compartments,
most
naive
flurry,
pollutant
stem
cells
or
pscs.
So
these
are
pluripotent
stem
cells
and
we've
talked
about
plural
potency
before
which
just
means
that
you
have
a
stem
cell
that
can
become
multiple
types
of
mature
cell
or
adult
cell.
B
B
B
Floridian
stem
cells
have
recently
been
shown
to
give
rise
to
embryonic
and
extra
embryonic
stem
cells
capable
of
self-assembling
into
post-gastrulation,
which
is
a
stage
in
development
Mouse
sem's,
while
bypassing
and
sem
is
this
stem
cell
direct
embryo
model,
while
bypassing
blastocyst
stage
and
eventually
initiating
organogenesis
X
utero.
So
this
means
that
they
are
able
to
have
these
stem
cells.
They
give
rise
to
different
types
of
embryonic
next
embryonic
stem
cells,
so
they
transition
to
another
type
of
stem
cell
and
they're
able
to
self-assemble
into
these.
B
These
different
models
that
this
type
of
model
that
they're
trying
to
build.
It
bypasses
some
of
the
traditional
stages
of
development
in
an
issue
it
initiates,
organogenesis,
meaning
that
you
can
build
organs
in
one
of
these
stem
cell
direct
embryo
models.
B
So
that
means
is
that
you
have
you:
have
the
mouse
model
that
they've
been
able
to
do
successfully
they've
been
able
to
kind
of
get
around
this
step
of
implantation
and
get
you
know
they've
been
able
to
sort
of
fast
forward
through
development
of
it
and
get
some
of
these
structures
that
we're
interested
in.
So
this
is
different
than
your
typical
organoid.
B
In
that
the
organoid
isn't
you
know
it
can
generate
some
of
the
cell
types
of
organs,
but
it
doesn't
necessarily,
you
know,
recapitulate
some
of
the
compartments
that
we
need
for
early
post-implantation
stage
human
embryos,
so
some
of
these
include
epiblast
hypoblast,
extra
embryonic
mesotherm
and
trophoblast
surrounding
the
latter
layers.
So
you
can
see
these
layers
of
tissue
in
the
form
it
start.
You
start
to
get
the
formation
of
tissues
in
this
in
these
type
of
models.
B
The
organized
human
sem's
recapitulate
key
Hallmarks
of
post-implantation
stage,
embryogenesis
up
to
13
to
14
days,
post-fertilization
such
as
bilaminar
disc
formation,
that'd
be
blessed
luminogenesis
angiogenesis,
so
these
are
just
different
that
again
amniogenesis,
which
is
the
Genesis
of
an
amniotic
sac
by
laminar
disc
formation,
anterior
posterior
symmetry
braking.
So
this
is
where
you
get
differences
between
the
front
and
the
back
end
of
the
embryo
primary
and
secondary
yolk
Sac
formation,
an
extra
embryonic
mesoderm
expansion
that
defines
a
cavity
in
a
connective
stock,
the
chorionic
cavity.
B
So
these
are
all
things
that
you
see
in
embryos
posting
Plantation.
They
lead
to
different
structures
in
the
in
the
in
utero
form,
and
you
know
so
they
can
basically
take.
They
can
build
a
model
of
an
embryo
and
they
can
fast
forward
past
implantation
and
they
can
get
sort
of
the
origins
of
some
of
these
structures,
and
this
is
of
course,
useful,
because
you
can
do
this
in
a
lab.
You
can
look
at
you
know
some
of
these
structures,
how
they
form
and
you
can
modify
them
as
needed.
B
So
this
is
a
sort
of
I,
don't
know
if
it's
really
an
ethically
gray
area.
It's
it's
not
like
we're
trying
to
actually
clone
humans,
but
we're
like
trying
to
sort
of
recapitulate
some
stages
of
early
human
embryogenesis
to
see
if
we
can
form
tissues
and
we
can
do.
C
B
Things
like
that,
so
this
new
platform
constitutes
attractable
stem
cell
based
model
for
experimentally,
interrogating
previously
in
accessible
Windows
of
human
Peri
and
early
post-implantation
development.
So
they
really
kind
of
gotten
to
the
point
where
you
know
we
can
actually
the
organoids
are
these
3D
cell
cultures,
where
you're
starting
to
get
differentiated
tissues
you're
starting
to
get.
You
know
these
sort
of
tissues
living
in
proximity.
This
is
Maybe
One,
Step,
Beyond
organoids,
where
we
have
some
of
these
features
of
early
development,
and
so
this
is,
you
know
they
do
some
they've
done
this.
B
Some
of
this
in
Mouse,
now
they're
doing
it
with
human
embryos,
and
so
there
are
a
lot
of
genes
that
you
need
to
express
in
some
of
these
cells
to
get
them
to
move
to
certain
Fates
and
so
they're
able
to
do
that.
Some
of
that
can
be
done
chemically
or,
if
you're,
a
trans
gene
or
something
like
that,
but
you
can
also
do
it
just
by
having
tissues
and
Associated
Asian
with
one
another
and
getting
cells
to
differentiate
and
form
like
layers
or
structures
or
compartments,
as
they
call
them.
B
So
the
second
paper
is
the
stem
cell
derived
Mouse
embryos
develop
within
an
extra
embryonic
yolk
Sac
to
form
anterior
brain
regions
in
a
beating
part.
So
this
is
about
Mouse
embryos.
This
is
a
different.
This
is,
if
we
back
up-
and
we
say
your
mouse
embryos-
that
we
can
develop
where
these
Mouse
models
stem
cell
derived
embryonic
models.
You
know
we
can
do
this
in
Mouse
again.
Mouse
development
is
very
similar
to
human
development,
in
the
sense
that
there's
the
same
Plantation
step
it.
You
know
it
develops
an
amniotic
sac.
B
It
develops
in
a
uterus,
and
you
know
that
stuff,
of
course,
is
also
necessary.
So
Mouse
embryo
was
developed
within
an
extra
embryonic
yolk
Sac
to
form
anterior
brain
regions
in
a
beating
part.
So
we're
able
to
recapitulate
this
ability
to
embed
this
thing
in
a
embryonic
extra
embryonic
yolk
Sac.
So
this
is
extra
embryonic.
This
is
outside
the
embryo.
B
You
have
a
yolk,
Sac
you're
able
to
form
these
different
structures,
so
in
this
case
you're
forming
anterior
brain
regions
or
things
at
the
front
of
the
brain
like
the
forebrain
and
then
a
beating
heart.
So
it's
not
only
your
forming
heart
cells
that
differentiate
your
forming
the
structure
that
starts
to
beat
and
you
can
get
a
viable
organ
in
that
way.
So
you
get
organogenesis
basically
in
this
kind
of
model,
so
the
abstract
reads.
B
So
this
is
the
authors
here
make
the
one
as
there
gets,
who
is
a
famous
person
in
this
field
and
a
couple
other
people
here,
so
the
abstract
reads:
embryo-like
structures
generated
from
stem
cells
can
achieve
varying
developmental
Milestones,
but
none
have
been
shown
to
progress
through
gastrulation
correlation
and
organogenesis.
So
these
are
again
stages
of
development
that
we
expect
in
an
embryo
in.
B
Embryo
here
we
show
that
etix
Mouse
embryos,
the
etix,
is,
is
I,
don't
know
if
they'll
take
it
into
the
definition
of
this.
Probably
in
the
paper
established
for
embryonic
stem
cells,
aggregated
with
trophoblast
stem
cells,
an
inducible
extreme
organic
embryonic
endoderm
stem
cells.
So
these
are
monosumbrio
models
that
are
established
from
different
types
of
stem
cells,
so
they're
just
putting
all
these
stem
cells
in
association
with
one
another
and
they're.
Seeing
what
happens
this
aggregate
can
develop
through
gastrulation
and
Beyond.
So
it
can.
B
You
can
take
that
aggregate
through
different
developmental
stages,
and
you
can
do
this
to
undertake
neural
induction,
which
is
a
stage
where
you
get
the
induction
of
neurals
like
the
neural
tube
and
the
neural
crest
and
generate
the
progenitors
needed
to
create
the
entire
organism.
So
you
can
get
this
sort
of
neural
differentiation
and
from
that
you
get
progenitors
that
you
need
to
create
things
like
the
forwarding
of
the
of
the
or
the
forebrain.
B
The
head
folds
of
the
etix
embryos
show
interior
expression
of
otx2,
defining
four
brain
and
midbrain
regions
that
resemble
those
in
the
natural
Mouse
embryo
etix
embryos
also
developed,
beating,
Hearts
current
structures
comprising
a
neural
tube
and
somites
tailbuds
containing
neuromesodermal
progenitors
and
primordial
germ
cells
and
gut
tubes
derived
from
definitive
vendor
a
fraction
of
ETI
X
embryos
show
two
neural
tube
abnormalities,
so
that
what
that
means
is
that
you
get
these
etix
models.
B
You
get
head
folds,
which
show
an
anterior
expression
of
this
Gene.
That
we
know
is
a
marker
for
neural
development.
This
defines
the
forebrain
and
midbrain
regions,
so
you
get
this
distinct
region
formation.
So
there's
these
differences
that
you
can
identify
in
these
models.
This
looks
like
natural
Mouse
embryo.
If
you
compare
them,
then
they
also
developed
beating
herds
trunk
structures,
which
is
the
sort
of
the
formation
of
a
neural
tube.
B
So
you
get
these
areas
that
are
sort
of
analogous
to
the
mouse
embryo,
where
you're
getting
neural
development
you're,
also
getting
the
neural
tube,
which
is
where
you
get
this
tube
of
cells.
That
kind
of
form
what's
going
to
be
the
nervous
system
and
the
and
the
spinal
cord,
and
things
like
that,
so
you
get
this
too
long,
two
bit
forms
in
the
embryo,
so
you're
starting
to
get
that
as
well.
You're,
getting
somites
tail
blood,
tailbud
containing
neuromesome
progenitors,
so
you're
getting
these
tailbud.
B
So
it's
like
the
tail
of
the
mouse
and
primordial
germ
cells,
which
are
different
types
of
germ
cells
that
you'll
need
for
different
types
of
stem
cells.
So
that's
that's
all
there
and
a
fraction
of
etix
embryosian
little
tube
abnormalities.
B
So
a
fraction
of
these
kind
of
embryos
that
are
being
generated
actually
show
different
abnormalities.
So
they're
not
all
perfect.
They
don't
all
come
out
the
same
Sometimes
they
come
out
with
errors
or
abnormalities.
What's
interesting,
is
you
can
partially
rescue
these
by
treatment
with
a
metabolic
whack
to
form
a
folic
acid,
so
this
actually
resembles
common
birth
defect
Therapies
in
humans?
So
they
do
this
with
birth
defects.
They
treat
the
embryolith
folic
acid
and
they
can
maybe
reverse
or
arrest
some
of
these
birth
defects.
B
Normally
etix
embryos
also
develop
a
yolk
Sac
with
blood
Islands,
so
they
actually
start
to
develop
blood
and
a
circulatory
system.
Perhaps
we're
not
really
quite
there
with
with
this
end
with,
as
well
as
with
the
yolk
sac,
but
it's
you
know,
you're
getting
the
precursors
of
that
overall
etix
embryo
is
uniquely
recapitulate
natural
embryos
developing
further
than
any
other
stem
cell
derived
model.
There
are
multiple
post-implantation
stages
and
with
an
extra
embryonic
membranes.
B
So
again
we
get
this
model.
This
is
in
Mouse,
not
in
human,
but
we
can
do
this
in
Mouse.
We
can
sort
of
see
that
you
can
move
Beyond
like
organoids.
You
could
move
Beyond.
Even
you
know
other
types
of
model
where
you
have
these
complex
tissues
that
form.
So
it's
a
really
interesting
sort
of
set
of
advances
towards
this
goal
of
like
maybe
creating
a
creating
an
artificial
embryo
or
creating
an
artificial
human
or
Mouse.
And
again
you
know
this
isn't.
This
is
basically
just
taking
stem
cells.
B
You're,
not
reprogramming
them,
necessarily
like
I
described
before
you're
taking
these
stem
cells
you're
putting
them
in
association
and
you're
doing
some
other
things
to
to
elicit
some
of
these
changes.
But
what
you
get
at
the
end
is
this
highly
structured?
You
know
system
that
actually
resembles
development.
It
is
in
development
itself,
but
it
resembles
development
which
is
important
to
remember
because
you
get
you
know
there
are
a
lot
of
ways
you
can
assess
what
you
have,
and
you
know
it
may
not
actually
be
developing
embryo.
B
It
might
be
something
that
resembles
it,
but
it
may
not
be
something
we
could
actually
take
to
a
full
tournament,
so
esc's
embryonic
stem
cells,
which
are
derived
from
the
upper
blast
to
a
remarkable
ability
to
form
embryo-like
structures
upon
aggregation,
so
we
have
ese
Aggregates
called
gastroides,
which,
if
you
embed
them
in
nature
gel,
which
is
this
medium,
that
they
often
use
to
do.
3D
cell
culture
can
form
Trump
light
structures
with
somites
a
neural
tube
and
a
gut.
B
So
these
are
all
like
things
that
we're
looking
for
in
the
embryo
that
you
can
actually
recapitulating
gastroyed
models
which
are
different
than
this.
What
we're
talking
about
here,
they're
a
bit
different
than
embryoids,
because
they
have
this
gas
they're
at
the
stage
of
gastrulation
and
development,
which
is
where
you
start
to
see
some
of
the
structure
pull
the
anterior
neural
development
can
be
promoted
in
gastroints
by
inhibiting
the
initial
burst
of
interactivity,
which
is
a
gene
that
gets
expressed
early
in
development,
and
it's
important
for
developmental
signaling.
B
They
do
not
Faithfully
represent
the
anatomy
of
natural
embryos,
so
embryoids
generated
from
naive
escs
or
these
embryonic
stem
cells
aggregated
with
an
ectopic
morphogen
signaling
Center,
which
is
where
you
have
this
chemical
signaling,
which
promotes
morphogenesis,
can
develop
a
posterior
midbrain
neural
tube
cardiac
tissue
in
the
gut
tube.
So
you
can
actually
do
this
in
these.
In
these
gasteroid
models,
you
can
build
or
embryoid
models
which
you
can
actually
go
beyond
the
gastro
wide
model
to
develop
some
of
these
tissues.
B
These
models
form
forms
of
epiblast
structures
de
novo,
meaning
that
they
do
this
within
the
model.
You
don't
need
to
really
induce
them,
but
do
not
unlike
natural
epiglass,
they
do
not
undergo
gastrulation.
B
So
you
have
these
different
types
of
models.
You
have
the
so
you
have
the
stem
cell.
B
B
Then
we
have
these
organoids.
B
B
B
So
you
start
to
get
you
know
a
beating
heart,
you
get
this.
You
know
you
get
blood
clots
things
like
that,
and
then
we
get
these
models
that
we're
talking
about
today.
So
this
is
this
model
here.
B
Let's
see
the
CM,
you
know
these
acronyms
are
somewhat
gonna
all
blend
together.
At
some
point.
B
Start
over
so
they
sem,
which
is
this
model
of
another
model
of
embryogenesis,
and
so
these
are
relevant
to
mammalian
systems
like
Mouse
and
human.
B
So
this
is
different
from
C
elegans.
By
far
this
is
you
know,
describing
sort
of
this
complex
embryo
or
this
complex
embryo,
more
complex,
embryogenesis
you'll,
see
in
what
we
call
regulatory
or
regulative
development.
You'll.
Have
you
have
you
know
these
different
stages,
That,
You,
Don't,
See
and
see
elegans,
See
elegans
is
just
kind
of
you
know
the
cells
differentiate.
B
They
form
this,
this
set
of
organs
and
tissues,
and
then
the
the
egg
hatches
and
you
get
this-
you
get
these
different,
comparable
stages
that
allow
you
to
have
a
little
bit
more
differentiation
and
new
cells
that
are
sort
of
useful
in
the
local
environment,
and
then
you
get
the
Old
Stage
in
mammalian
embryos.
It's
far
more
complex,
and
so
this
is
these
are
some
of
the
most
recent
four
areas
into
this
and
there's
been
some
criticism
of
this
online.
B
You
look
at
RNA
expression,
you'll,
get
gene
expression
and
you
look
at
like
the
marker,
the
marker
rnas,
that
need
to
be
there
and
that's
fine,
but
that
doesn't
mean
that
those
cells
are
identical
to
what
you
might
find
in
an
embryo.
So,
for
example,
with
our
induced
pluripotent
cells,
we
think
that
they
behave
like
pluripotent
cells
or
like
stem
cells,
but
they
aren't
necessarily
those
kind
of
cells.
B
A
lot
of
times
induce
neural
cells
are
like
that,
where
you
can
create
a
neuron
that
has
some
sort
of
action
potential,
but
it
isn't
like
any
kind
of
neuron
you
see
in
nature
or
in
the
human
body
or
any
other
organism,
so
these
cells.
You
know
these
are
kind
of
like
facsimiles,
where
we
try
to
induce
the
right
factors
and
we
maybe
get
something
functionally
similar
to
what
we
want,
but
it
isn't
that
thing,
and
so
this
is
what
the
Martinez
Arias
talks
about
in
this
post
that
he
had
on
on.
B
It
was
another
paper
that
I'm
not
talking
about
here,
but
this
is
a
very
similar
problem,
so
beside
from
all
those
identification
issues
and
identity
issues
to
put
things
in
perspective,
a
human
embryo
at
stage
14,
which
is
the
stage
of
an
alien
development
there's
like
Mouse
14,
but
human
embryos
also
have
a
stage
14.,
it's
not
completely
analogous,
but
basically
it's
at
this
stage
of
development,
a
human
embryo
at
stage
14
is
a
complex
structure
shown
below
with
yolk
Sac
attached
to
the
epiglass
that
will
form
the
embryo
and
both
covered
by
two
membranes,
the
chorion
and
the
trophoblast.
B
So
you
see
here
that
there's
this
amnion
amniotic
cavity,
the
epiblast
is
here
the
hypoblast
is
here
and
then
you
have
this
chorionic
cavity.
B
C
B
Have
to
be
well
timed,
so
if
they're
out
of
sequence,
you
don't
get
so.
This
is
14
days
post
fertilization,
you
don't
get
a
viable
embryo,
and
so
there
are
a
lot
of
things
that
can
go
wrong
here,
and
so
this
is.
This
is
what
you
see
in
an
implantation.
You
know
implanted
embryo
and
you
know
its
Associated
function,
so
it's
definitely
we're.
Definitely
not
there.
Yet
we
can
get
differentiated
tissues.
We
can
get
other
structures,
but
we
can't
necessarily
this
isn't
like
cloning
humans.
B
It's
my
takeaway
message
from
this,
but
it's
interesting
stuff,
and
especially
in
the
light
of
a
lot
of
stuff
we've
talked
about
with
differentiation,
and
you
know
other
types
of
systems
like
organoids
and
things
like
that.
B
So
that's
all
I'm
going
to
say
about
that.
I
see
Gio
Hans
here
welcome.
B
E
Yeah
yeah
I'm
a
little
bit
late
today,
because
I
was
busy
as
something
something
else
and
just
got
home
a
few
minutes
ago.
I
think.
B
Yeah,
so
you
had
mentioned
that
you
wanted
to
work
on
a
paper
and
this
would
be
on
I
guess,
craft
neural
networks
and
the
topological
data
analysis,
area
and
so
himanshu
is
here,
of
course,
is
doing
the
summer
of
code.
For
that,
so
did
you
want
to
talk
about
that?
A
little
bit.
E
Yes,
Yes
actually
I
want
to
share
one
idea
regarding
the
topological
data
analysis
and
actually
the
persistent
homology
specifically
and
a
graphing
Network,
and
how
our
policy
can
be
applied
to
the
cell
tracking
and
for
now,
I
choose
tracking
as
a
downstream
task,
because
I
think
I'm
a
little
familiar
with
sale,
tracking
and
it
says,
there's
many
deep
Learning
Works
have
been
successfully
applied
to
cell
tracking,
but
maybe
there
are
some
other
other
choice
choices
but
I'm,
not
really
sure,
but
I
wanna
share
the
idea
here.
E
Right
couldn't
see
my
screen
share,
yes,
I
share
what
paper
and
actually
discovered
the
graph
new
natural
for
still
tracking
in
microscopy
video
secretary
I
think
we
have
discussed
this
paper,
maybe
the
last
year
and
2022,
and
maybe
you'll
still
remember
that,
and
this
paper
is
actually
the
starting
point
of
my
idea
and
I
for
now,
I
wanna
for
tonight,
I
wanna
briefly
introduce
some
papers.
I
won't
go
into
their
details,
but
I
would
like
to
describe
my
motivations.
E
So,
first
of
all,
so
first
of
all,
this
paper
I
won't
introduce
the
model
design
of
this
paper,
but
I
would
like
to
introduce
some
related
words.
That's
introduced
by
this
paper,
for
example,
cell
tracking,
is
a
is
a
very
preference.
It's
a
very
important
task,
and
essentially
in
this
test,
you
want
to
build
some
sell
tragedies
which
connects
cells
across
adjacent
frames.
For
example,
you
have
a
microscopy
video
and
you
want
to
sample
some
friends
a
friend
by
frame
and
for
each
frame.
E
You
have
a
lot
of
cells,
but
you
only
have
the
so
the
pixels
or
the
image
patches
of
the
cells
and
at
first
you
want
to
use
some
cell
segmentation
model,
such
as
some
CL
models,
resnets
BGG
or
something
else
to
segments
the
patches
of
these
cells,
so
that
you
have
one
patch
for
each
cells
and
each
patch
will
contain
some
pixels
and
with
with
this
patches,
you
can
use
some
CM
models
to
generate
the
embedding
of
each
cells
based
on
the
segmented
patches
and
and
then
we
go
to
the
sales
connection
task,
I'm,
sorry
that
I
didn't
yeah,
as
you
saw
the
pipeline
yeah
yeah.
E
Yes,
and
this
and
for
now
you
haven't
imposed,
you
have
a
video
and
you
sample
some
friends
and
then
for
each
cell
you
would
have
a
patch
will
have
a
patch
of
pixels.
Then,
in
the
feature
extraction
stage,
let
me
mean
in
a
feature
extraction
stage.
You
will
have
some
kind
of
CM
models
to
like
map
each
patch
into
the
embedding
of
the
cell,
and
there
are
a
lot
of
same
models
you
can
choose.
E
But
it
is
not
very
important
in
in
my
presentation,
because
this
is
not
my
focus
and
so
that's
you
will
have
an
initial
embedding
like
this
is
the
CNN
on
some
other
complicated
modules
and
then,
after
that,
actually
you
have
a.
You
will
have
a
point:
Cloud
temporal
Point
Cloud
I
mean,
for
example,
for
each
friend
you
have
a
point
Sprout
and
actually
this
point
call
Will
Change
along
the
time.
E
Yes,
so
you
have
a
way
to
convert
these
Tempo
points
Cloud
into
a
directed
graph
and
I'm,
not
sure
if
the
issue,
if
in
this
paper
there
is
there
is
any
any
picture,
can
illustrate
my
idea
over
the
moment-
maybe
not
so
so,
let's
go
back
to
the
same
picture
yeah.
So
it's
very
easy.
It's
a
very
intuitive
way
to
build
a
directed
graph.
For
example,
for
a
friend
you
have
a
lot
of
cells.
E
You
can
use
something
like
can
block
to
connect
the
sales
data
which
are
very
close
to
each
other,
and
also
you
can
connect
cells
between
adjacents
friends.
E
If
in,
if
you
think
these
cells,
they
are
actually
very
close
to
each
other
in
a
spatial
Dimension
yeah,
so
that
means
you
have
a
way
to
build
a
directed
graph,
and
next
you
want
to
connect
cells
to
build
the
cell
trajectories.
This
is
what
we
do
in
the
cell
tracking
task,
and
it
can
be
done
by
by
a
very
how
can
I
say
very
classical
class
in
the
graphene
Network,
which
is
the
link
prediction.
E
That
means,
when
you
build
a
graph,
you
want
to
predict
which
each
are
which
each
are
are
the
real
age.
They
are
they're
existing.
You
want
to
predict
this
stuff,
and
if
there
are
some
issues
which
you
predict
are
not
existing,
you
will
remove
this.
Each
and
the
remain
edges
are
the
edges.
They
connect
cells
across
frames.
So
that's
you
can
build
a
sales
address
because
it's
easy.
They
connect
sales
Friends
by
friends,
friends
by
printing,
something
like
this
and
by
now
you
have
and
do
you
have
any
questions
like.
A
E
No
okay!
Okay!
Actually
you
can
interrupt
me
anytime
during
this
presentation:
okay,
yeah
yeah.
Actually
you
will
find
that
in
this
paper
the
authors
propose
a
way
to
construct
a
graph
structure
or
we
can
say
a
natural
structure
on
the
microscope.
Video
based
on
the
microscope,
video
and
actually
there
are
many
other
papers,
also
adopt
this
idea,
like
we
look
at
the
cell
development
process
from
the
perspective
of
of
a
network
of
graph
of
points
cloud
or
something
else.
E
So
when
you
have
a
point
spread
or
will
you
have
a
natural
structure,
you
can
investigate
the
geometry
topology
and
something
else
from
this
points
Cloud
this
one
very
classical
story
in
the
machine
learning
and
deep
learning
like
when,
when
you
have
a
data
sets.
Actually
you
have
a
set.
You
have
a
set
of
points,
but
many
papers
would
assume
that
there
is
an
underlying
manifold
underlying
the
this.
E
This
point
Cloud
underlying
these
data
sets
and
you
need
to
adopt
some
waste
some
matters
to
like
to
explore
the
underlying
the
geometry
of
this
underlying
manifold
and
you'll.
Be
this.
This
geometry
and
this
topology
will
be
very
helpful
for
the
downstream
task,
so
the
story
also
applies
to
here,
and
so
the
point
is
the
starting
point
is:
if
we
haven't
I
have
a
master
or
we
can
say
we
have
a
perspective
to
build
the
whole
things,
the
wholesale
development
process
as
a
network.
How
can
we
explore
the
topology
and
geometry
of
this
process?
E
And
actually
another
question
is:
is
it
really
reasonable
to
like
to
explore
the
topology,
because
what
I
mean
is
if
we
want
to
explore
the
topology
of
these
data
sets
we
we
need
to
have
some
like
some
Crews
or
some
evidence
to
show
that
these
data
says
they
have
some
significance
or
important
topological
features,
if
not
to
explore
the
topology
of
this
data.
Sciences
is
not
reasonable.
I
I,
in
my
opinion,
I
think
so
here
are
some
papers.
E
Actually
some
of
these
cure
are
some
papers,
and
some
of
them
have
been
covered
in
our
meetings
before
and
there's
one
rather
newspaper
I
think
self-differentiation
processes,
spatial
Network
and
actually
I've.
I
have
read
this
paper
in
the
in
the
last
year
in
Json,
2022
and.
E
Also
have
proposed
has
utilized
many
network
analysis
techniques
such
as
such
as
sorry,
okay
means
cluster
analysis
and
such
as
click
analysis
and
Community
detection
and
modularity.
This
techniques
to
analyze
the
natural
structure-
and
there
are
a
lot
of
results
and
I-
am
not
able
to
cover
them
today,
because
I
don't
want
to
take
up
too
much
time
and
but
the
thing
is
all
these
results
will
convey
some
interesting
conclusions.
E
For
example,
the
connectivity
is
a
very
important
property
in
cell
development
process
in
the
network
structure,
within
the
cell
development
process
and
for
example,
if
you
look
at
the
conclusion
you
find
we
can
see
that
cells
in
ciardigans
embryos
only
partially
segregates
by
linear,
lineage,
sub
lineage
and
cell
states
are
cool,
located,
Main
Service,
so-called
weight
connectors
in
the
network
representation.
So
in
my
understanding,
I'm,
not
sure
if
I
understand
it
correctly.
E
In
my
understanding
that
means
ins
in
a
certain
sub
lineage,
those
cells
will
be
connect
to
each
other
very
close
very
closely,
but
between
the
sub
lineage,
these
cells.
The
connections
between
these
the
cells
across
different
sub
lineages
are
very
like
very
sparse.
So
that
means
we,
the
information
of
connectivity.
E
B
I
think
you
have
that
about
right
and,
like
the
the
things
some
of
the
things
you
mentioned
in
this
paper
were
like
you
know,
you
have
these
spatial
networks,
so
the
networks
are
usually
Network
Theory.
You
know
they're
doing
networks
on
they
create
networks
from
data,
so
it
could
be
like
social
media
networks
or
it
could
be
like
communication
networks
and
the
networks
aren't
necessarily
spatially
defined.
So
you
just
say
there
are
two
things
that
are
associated.
B
It
could
be
like
that
they're
correlated
or
it
could
be
that
they're
just
kind
of
like
in
proximity
in
time.
You
know
so
they're
they
don't
have
a
distinct
spatial
location
in
the
embryo
networks.
They
have
a
distinct
spatial
location,
they
have
an
identity,
and
so
you
can
make
you
know
you
can
basically
use
that
as
a
Criterion
for
connectivity.
Are
they
close
together
or
are
they
farther
apart?
B
That
are
close
together
will
cluster
into
modules
and
those
that
don't
you
know,
won't
and
then
usually
those
correspond
to
these
sub
lineages
in
the
C
elegans
lineage
tree,
and
so
this
is
useful
because
in
biology
the
cells
are
signaling
one
another.
You
know
they're
sending
out
chemical
signals
from
each
cell
and
they're
communicating
it's.
It's
called
paracrine,
signaling
and
so
a
lot
of
times.
That's
important
for
development.
It's
more
important
for
those
sort
of
regular,
regular
regulative
embryos
that
we
talked
about
earlier,
but
in
C
elegans.
B
You
still
have
that
kind
of
signaling,
and
so
these
kind
of
networks
will
tell
you
maybe
something
about
that
kind
of
signaling.
It
could
tell
you
about,
like
you
know
what
positions
they
need
to
be
in
to
differentiate,
or
something
like
that.
So
you
know
building
a
network
from
like
microscopy
data,
especially
if
you're,
using
like
a
graph
neural
network
approach,
where
you
know
they're
different
methodologies
for
creating
the
graph
neural
network
message
passing
or
some
of
these
other
methods.
B
I
think
you
said,
I,
don't
know
what
you're
thinking
about
for
sort
of
a
Criterion
I
mean
we
could
even
work
out
a
biologically
specific
Criterion
for
like
building
a
graph
neural
network
and
just
building
the
embedding
from
the
data,
because
we
have
like,
with
with
the
microscopy
images
see
this
was
got.
This
was
taken
from
like
cell
tracking,
where
you
have
like
positions
that
are
average.
B
B
Or
the
spatial
location
relative
to
cell
divisions,
they
still
haven't
worked
that
out
very
well
with
this
method,
but,
like
you
know,
with
with
just
taking
it
from
microscopy
data
and
building
embeddings,
you
can
kind
of
get
these
graph
topologies.
That
are,
you
know,
maybe
change
with
time
and
they
change
their
shape
or
something,
and
so.
B
Of
where
this
might
be
useful
and
kind
of
giving
that
description
of
what
cells
sort
of
cluster
together,
do
they
lead
to
the
formation
of
tissues?
B
B
We
don't
know
that,
but
these
are
kinds
of
things
you
could
do
with
these
kind
of
embryo
Networks,
but
I
like
you
know,
just
I
think
we
need
to
kind
of
say
we
could
probably
implement
something
similar
to
this,
but,
of
course,
with
graphical
networks,
it
gives
us
the
ability
to
generate
these
embeddings
and
use
them
to
you
know,
to
sort
of
say
something
about
the
embryo,
instead
of
just
like
cells
and
segmenting
cells.
So
we
get
this
extra
layer
of
structure.
On
top
of
that,.
E
Yeah
yeah,
actually,
yes,
I
agree
with
you.
Okay
I
think
like.
If
we
look
at
this
paper,
just
like
what
you
said,
maybe
we
don't
if
we
sometimes
we
explore
some
kind
of
topology,
but
maybe
it
would
be
very
difficult
to
say
like
to
clearly
understand
what
those
typology
means
and
what
those
laboratory,
how
does
Prodigy
works
and
how
does
corporate
you
with
effects
other
modules
or
other
functions,
but
sometimes
like
for
machine
learning
different
models?
E
This
topology
we
can
try
to
like
impose
it's
detective
technology
into
the
deep
learning
models
and
see
if
this
inputs
is
helpful
for
deep
learning
models
to
do
some
Downstream
tasks
such
as
the
cell
tracking
and
if
the
performance,
if
finally
performers
say
yes,
it
helps,
then
we
can
see
that
this
topology
is
helpful,
although
maybe,
although
it's
still
very
very
difficult
to
explain
how
this
topology
helped
yeah.
B
Yeah
I
think
it's
it
yeah,
it's
sort
of
an
open
question.
We
don't
know
like
why
it
would
be
important
for
a
certain
topology
over
another.
We
just
I
mean
in
network
science.
A
lot
of
the
state
of
the
art
is
like
well.
Actually
people
have
done
a
lot
of
things
with
some
things
that
topological
data
analysis,
but
it's
it's
very
early
days
like
if
I
showed
you
like.
You
know
they
have
these.
These
things
called
hairball
networks
and
it's
sort
of
like
this,
except
it's
there's
no
structure
at
all.
B
And
you
know
people
have
identified
things
like
that,
but
and-
and
you
know
clearly,
you
can
like
visually
identify
things
like
modules
or
like
clusters
right,
but
knowing
what
those
mean
and
kind
of
getting
at
the
structure
of
those
in
their
relationships.
That's
so
they've
tried
to
do
that
with
Community
detection,
where
they
kind
of
try
to
detect
these
subgroups
or
these
communities
within
the
network
hairball
or
within
the
network
topology.
B
But
that's
like
a
very
hard
computational
problem
we
have
to
do
like
you
have
to
you
know
basically
bootstrap
the
get
where
you
have
to
do:
jackknifing
or
bootstrapping,
or
some
technique
like
that
to.
B
Like
compared
to
deep
learning,
it's
not
that
computationally
difficult,
but
you
know
you
have
these
networks
that
have
like
a
lot
of
nodes
and
a
lot
of
connections.
A
lot
of
links
so
you're
dealing
with
a
lot
of
you
know
it's
it's
I,
don't
know.
Maybe
there's
a
shortcut
using
deep
learning.
I,
don't
think
anyone's
looked
at
that,
but
you
know
interpreting
that
topology
I
agree
is
a
very
important
thing
to
sort
of
get
our
hands
around,
but
there
is
really
isn't
a
lot
of
state
of
the
earth
that
we
can
draw
from.
D
E
All
right,
okay,
let
me
go
ahead
and
yeah,
because
I
think
it's
I,
don't
wanna!
Okay,
if
I
take
up
too
much
time,
actually
yeah
and
actually
in
this
paper,
the
connectivity
is
actually
I.
Think
at
the
end
of
my
princess,
I
would
introduce
the
persistent
homology,
which
is
to
detect
multiple
dimensions
of
topology
and
zero
dimension.
E
E
So
actually,
in
this
paper,
I
like
to
say,
the
connectivity
introducing
this
paper
is
exactly
the
Zero
Dimensional
topological
signature
detected
by
the
persistent
homology,
yes,
but
so
apart
from
the
connectivity
we're
having
some
other
papers
and
these
papers,
there
are
also
introduced
in
our
group
meeting
in
the
table
we're
meeting
before,
for
example,
in
this
paper,
the
author
saw
that
I
think
for
some,
maybe
some
organizations
also
for
some
multicellular
organizations
of
diverse
systems.
E
They
absorb
some
kind
of
like
Robo
rotation
of
those
collectivity
cell
migration,
yeah
yeah.
So
this
is
one
kind
of
like
high
dimensional
topological
signature
in
the
context
of
persistent
homology
we
would
like
to.
We
would
like
to
say
this
kind
of
rotation.
E
It
is
a
circle
or
you
can
say
it
is
a
set
of
circles
and
circles
in
the
in
the
context
of
persistent
homology
is
the
one-dimensional,
topological
signature?
Actually,
yes,
and
if
you
go
to
this
is
another
paper,
and
it
also
introduced
in
the
group
meeting
before
and
I
think
it's
favorites
to
wanna
to
say
that
the
creativity
corrective
corrective
cell
migration
is
a
very
important
process
and
the
substration
interface
of
this
process
are
typically
curved
like
for
example.
E
If
you
look
at
this
this
picture
this
substrates
there
are
it's
like
a
semi
severe,
maybe,
and
this
substrates,
the
curved
and
actually
in
persistent
homology.
We
say
if
we
just
look
at
the
homology
theory,
homology
Theory,
homology
tour.
This
theory
is
not
able
to
detect
it's
not
able
to
differentiate
two
spaces
which
have
different
conditions
actually,
but
the
persistent
homology
based
on
this
point,
spouse
there
are
some
papers,
show
that
resistant
homology
is
able
to
detect
the
curvature.
E
So
something
like
this,
they
are
higher
dimensional
topological
signatures
such
as
the
two
or
three-dimensional
yes-
and
this
is
another
paper,
the
growing
topology
of
the
theaters
connected,
and
this
paper
is
interesting
because
it
is
actually
using
the
persistent
homology
to
analyze
the
topology
on
the
department
of
the
arrogance
connection
and
if
you
look
at
there
are
some
results
and
I
want
to
show
one
approach.
And
if
you
look
at
this
figure
one
many
persistent
cavities
emerge
in
the
developing
neural
connection
of
C
elegans.
E
You
can
see
that
there
are
some
names
like
barcode,
baby
number
or
any
density
and
actually
barcode
and
many
numbers
they
are
names
from
the
homologous
theory
is
faster
in
the
consists
of
homology
and
if
I
I
I
won't
like
introduce
the
definitions
of
this
Nest,
but
I
would
say
the
very
numbers.
So
that's
if
any
numbers
is
very
big.
E
That
means
there
are
a
lot
of
High
dimensional
Cycles,
such
as
the
the
cycle
or
such
as
the
the,
how
you
say,
The
Voice
or
such
as
the
connected
components,
which
are
a
lot
of
these
topological
signatures
in
this
space.
You
can
see
that
the
belly
number
is
very
big
like
60,
80
or
100.
That
means
this
took.
There
are
a
lot
of
a
high
dimensional
signatures
in
this
space,
and
if
you
look
at
this
barcode,
you
will
see
there
are
lots
of
Lights,
which
are
very
long
if
it
slides
alone.
E
That
means
there
are
lots
of
signatures,
topological
signature,
that
they
bone
it's
a
very
early
stage
and
they
die
it's
a
very
late
stage.
So,
like
this
signature,
this
these
Cycles,
they
live
for
very
long
period.
Maybe
it
is
a
if
it
is
very
difficult
to
understand
what
does
it
mean
for
the
signature
they
can
live
for
a
long
period
or
for
a
long
time,
but
in
the
context
of
position
of
homology.
E
That
means
there
are
some
important
topological
signatures,
so
I
think
I
won't
introduce
the
position
of
commology
today,
because
we
we've
taken
very
long
time.
But
what
I
want
to
say
is
what
I
want
to
say
is
we
can
see
that
there
are
lots
of
theological
signatures
during
the
say,
development
process,
and
so
then,
but
in,
but
actually
in
what
I?
What
I
want
to
say
is
actually
in
those
previous
works
of
cell
tracking
tasks,
including
those
classical
works
and
including
those
deep
Learning
Works.
E
Those
methods
are
essentially
based
on
the
like.
The
visual
features,
the
cells
like
they
think
these
two
cells
they
are
visual.
They
are
visually
similar
to
each
other,
like
their
visual
traits.
They
are
very
similar
to
each
other,
so
they
connect
these
two
cells.
So
these
are
the
is
the
essential
ideas
behind
those
previous
words.
E
I'm,
not
sure
my
statement
is
is
correct,
because
I
still
need
to
refer
to
a
lot
of
other
papers,
but
for
now
for
me,
in
my
opinion,
it
seems
that
there
are
few
words
you
considering
the
topology
to
logical
information
within
the
development
process
to
conduct
the
cell
tracking
task.
E
So
I
would
like
to
bring
I
like
to
bring
the
consistent
homology
this
tour
to
analyze
the
topological
information
within
the
cell
development
process
and
use
this
information
to
enhance
the
cell
connection
model
or
cell
tracking
model,
such
as
the
graphene
Network
that
I
introduced
before
and
actually
there's
a
very
simple
demo
pipeline
for
this.
For
this
idea.
First
of
all,
you
have
is
this
the
homology
and
let's
just
call
it
pH
for
Simplicity,
and
you
have
pH
and
assuming
that's.
E
Ph
is
a
model
and
you
impose
the
the
points
Cloud
you
have
into
the
ph
and
pH
generates
some
some
results,
some
outputs,
which
is
exactly
the
Precision
diagrams,
and
then
you
use
some
methods
to
vectorize
the
position
diagram.
So
that's
the
position
diagrams
I,
converted
into
some
embedding
so
some
vectors
right.
So
actually
in
this
process.
There
are
many
problems,
for
example,
this
conversion
or
this
mapping
is
not
different.
It
cannot
be
it's
not
differentiable.
E
That
means
it
cannot
be
like
automatically
differentiable
by
those
python
tensorflow
or
some
other
other
Frameworks,
but
it
is
minor
progress
and
we
would
solve
this
in
the
future,
but
for
now
after
that's
the
resistant
diagrams
I
converted
into
the
the
embeddings,
that
means
you
have
been
embedding,
which
is
the
topological
summary
of
your
data
sets
like
if
like,
for
example,
if
your
your
data
set,
there
is
a
global
rotation
in
the
cell
development
process.
E
Then
this
embedding
can
summarize
this
Global
rotation,
because
the
global
rotation
is
a
set
of
circles.
It's
a
set
of
one-dimensional
topological
features
which
can
be
detected
by
the
persistent
homology
yeah.
That's
my
points,
so
that
means
you
use
Precision
hormone
to
detect
those
global
topology.
E
So
after
that's
when
you
have
the
topological
embedding
apart
from
this
topological,
embedding
you
have
another
embedding,
which
is
the
visual
embedding
generated
by
cell
segmentation
model.
Maybe
we
maybe
we
want
to
say:
okay,
we
we
don't
need
this
visual
embedding,
but
actually
because
we
need
to
do
the
sales
segmentation
stage
and
we
will
have
a
sales
segmentation
model
like
some
CM
models,
and
if
we
have
these
models
we
don't
need
to
remove
the
outputs
of
this
model,
which
is
the
visual
imbalance.
E
Yeah.
Maybe
I
didn't
myself
career
here,
but
yeah
yeah.
So
you
know
what
I
mean
yeah
yeah,
so
yeah
when
you
have
typological
embeddings
and
you
have
visual
imbalance,
you
have
a
lot
of
ways
to
combine
them
and
you
can
use
some
something
like
contamination.
You
can
use
some
MLP
models
to
combine
them
and
there
are
lots
of
ways
in
deep
learning
scenarios,
so
it
is
not
very
challenging.
But
after
that,
will
you,
after
you
combine
these
two
embeddings
into
a
single
embedding?
E
That
means
this
inputs
is
embedding.
It
has
two
logical
information
in
yet
and
it
has
the
visual
information
of
cells.
Then
you
want
to
inputs
like
this,
embedding
into
a
cell
connection
model
such
as
a
graphing
Network,
and
you
can
take
the
same
same
process
of
the
graph,
your
natural
paper
that
I
introduced
before
to
do
the
link
prediction
to
connect
sales.
So
that's
a
that's.
That's
all!
E
That's
a
very
simple
demo,
Pipeline
and
actually,
during
this
whole
process,
we
have
a
lot
of
problems
to
solve,
for
example,
so
the
differentiability
is
the
first
problem
that
I
introduced
before
and
another
problem
is.
You
may
still
remember
that
the
topological
embedding
is
topological
summary
of
the
whole
data
sets,
but
what
will
what
we
want
to
do
is
to
connect
cells.
So
that
means
we
are
doing
work
on
the
sale
level
we
need
to.
We
need
the
information
of
each
single
cell,
but
so
that
means
we
have
a
requirements.
E
That
means
we
want
to
map
the
topological
summary
of
the
whole
data
sets
into
the
cell
level,
topological
information
yeah.
So
this
is
another
challenge
yeah,
so
actually,
for
now
there
are
lots
of
other
challenges,
but
I
will
I
won't
like
introduce
all
of
them
here.
So
that's
my
idea
for
how
to
like,
yes
use
TDA
for
the
cell
tracking,
and
so
some
other
stuff
like
for
me,
because
I'm,
a
machine
learning,
guy
I'm,
a
different
guy,
I'm
I,
know
more
about
the
Deep
learning
and
less
about
the
biology.
E
So
if
we
want
to
say
we
wanna
do
a
research
and
prepare
a
paper
and
we
want
to
Target
a
conference
a
journal
for
me
just
just
for
for
me,
I
know:
one
conference,
the
icar,
the
I
career,
which
will
happen
I
mean
no
no
happened.
I
mean
the
deadline
for
the
submission
for
this
conference
in
general,
the
late
September
or
the
early
October,
maybe
I
remember
the
conference
will
happen
in
Austria
in
next
year.
I
mean
yeah
I'm,
not
sure.
E
So
maybe
it
is
very
interesting,
so
so
so
in
my
brain
big,
because
actually
this
is
because
actually
I'm
now
pursuing
my
computer
degree,
maybe
you
don't
know
which
what
what
is
the
inferior
is
the
master
of
research
in
Hong
Kong.
Yes-
and
it
is
one
of
my
risk
research
topic
during
my
my
study
yeah.
So,
in
my
plan
of
time
timeline
I
want
to
be
posting
is
today,
is
19
June
and
maybe
in
June,
July
and
August
and
September
is
three
and
a
half
months.
E
I
would
like
to
focus
on
this
project
in
life.
If
we
can
conduct
any
words
on
this
idea,
yeah
for
now.
B
E
B
That'd
be
great
that
I
yeah
I'd
be
happy
to
supervise
it
and
make
sure
that
we
can
get
something
out
by
wherever
the
fall
be
great.
Hamanchi
had
a
couple
questions
in
the
chat.
One
was:
does
the
single
embedding
that
is
generated,
work,
a
multiple
Downstream
tasks,
or
are
you
planning
to
focus
on
a
single
task
that
can
be
then
branched
out.
E
Very
cool
question:
actually,
for
now,
what
I'm
focusing
on
is
only
the
sale
tracking
task
but
I
think
because
the
multicellular
biology
is
a
very
big
field
there.
There
can
be
many
other
tasks,
I
think
maybe
maybe
other
Downstream
tests.
We
can
focus
on
so
I
think
for
now
we
can
have
a
demo
like
we
have
a
demo.
If
our
Master
can
work
on
Cell
tracking,
we
can
consider
more
tasks
and
if
you
have
any
idea
about
other
tasks,
you
want
to
try
and
we
can
like.
E
B
B
D
Think
yeah,
so
thank.
B
E
Question
you
can
ask
me,
maybe
yeah
in
in
a
select
Channel
or
maybe
something
else,
yeah.
B
So
it
looks
like
sushma
joined
us
after
Jihan
was
started
his
presentation
so
such
month
I
know
you
presented
in
on
Friday.
Did
you
want
to
add
anything
or
yeah.
D
When
we
have,
when
I
had
meeting
with
my
I,
was
not
able
to
iterate
through
the
create
a
batch
of
bounding
boxes,
so
mayuk
also
told
to
drop
the
idea
of
fine
tuning
Sam,
but
I
tried
a
little
bit
after
that
meeting,
somehow
made
out
to
take
a
batch
of
boxes
at
same
time.
Higher
here
we
can
see,
I
was
able
to
iterate
our
Prime
loader
data
loader,
so
I
was
I
tried
to
get
backstage
of
it,
and
these
are
my
maze
Dimensions.
D
This
is
the
mask
Dimension,
and
this
is
the
box
dimension.
There
are
eight
boxes
of
159,
comma
four
of
first
box.
First
box
contains
of
159
comma
four,
so
it
is
working
pretty
well
so
I,
just
I
was
going
through
the
Sam
architecture.
Image
encoder
I
was
almost
at
the
end
of
this
code.
I
just
need
to
write
the
tiny
Loop
and
try
it
on
the
loss
function.
Right
now,
I
was
working
on
this
loss
function,
which
I
caught
my
interest.
D
D
This
is
a
special
function
generally,
it
is
used
for
only
for
segmentation
things
right
now.
We
are
giving
the
model
the
bounding
box
and
it
tries
to
segment
the
image
image
in
that
bounding
box
and
then
that
which
have
been
predicted
will
be
compared
with
our
original
image
original
ground.
According
to
that
lossless
function,
this
loss,
function,
works
and
I
have
read
so
many
articles
around
this,
so
they
mentioned
like
this
is
the
only
best
method.
This
is
the
only
best
loss
function.
D
We
can
use
to
get
better
to
get
less
loss
in
the
model,
so
the
code
is
almost
ready.
I
just
need
to
write
the
training.
Loop
I
was
using
Adam
Optimizer.
Everything
was
fine
I
just
going
through
the
image.
Encoder
part,
we
are
not
changing
any
image
encoder
part,
but
I
was
changing
a
little
bit
in
the
decoder
part
as
it
is
a
90
million
parameter
model.
So
I
was
not
touching
any
prompt,
encoder
or
Vision
encode.
An
image.
D
Encoder
I
was
just
changing
the
weights
of
the
last
layer
of
decoder
layer
which
segments
the
whole
thing.
Oh
yeah,
that's
what
my
update
is.
Maybe
I
will
start
working
on
this
training
group
also
from
tomorrow.
I
have
some
resources
to
work
it
on
and
yeah,
and
another
thing
now
this
is
another
thing
I
have
got.
This
is
a
resource.
There
is
some
tutorial
on
how
to
find
your
exam,
so.
C
D
They
were
segmenting
the
they
have,
they
are
giving
annotations
for
each
file
which
I
don't
have
so
I
was
just
normally
implementing.
Without
his
annotations
file,
yeah
I
was
going.
I
was
trying
to
replicate
this
code
and
Implement
in
a
nice
way.
After
this
segment.
Anything
model
is
done,
then
I
will
move
on
to
the
download
model
and
start
working
on
it.
Actually,
I
wrote
a
little
bit
code
of
data
set.
How
to
convert
a
3DS
for
this
file
into
normal
ndra
I
was
I,
will
start
working
on
the
other
things.
D
D
He
helped
actually,
he
told
to
drop
the
idea
of
fine-tuning
yeah.
The
modal
architecture
is
kind
of
a
bit
different
and
not
able
to
We
Can't
fine
tune
it.
That
is
the
reason
it
all
to
drop
the
idea,
and
he
skipped
me
to
work
on
the
Deep
lab
D3
model,
but
I
thought
of
giving
it
a
try.
Then,
when
I,
when
I
implemented,
this
thing
I
was
able
to
get
a
data
loader,
which
is
it
terrible
by
using
Collide
function.
Actually,
the
gnm
is
what
you
have
mentioned
in
the
channel.
D
It
helped
me:
I
was
going
to
some
other
other
people
quotes
and
yeah
I
was
able
to
write
it
yeah,
maybe
I'll
start
working
on
the
training
Loop.
Also,
maybe
I
will
try
this
weekend
the
model
after
that,
after
we
get
getting
the
losses
and
the
segmentation
loss
of
the
world
I'll
start
working
on
the
development
yeah.
That's
what
my
update
is
and
I
was
thinking
to
convert
this
right.
D
That's
what
I
was
thinking
about,
because
I
don't
know
why
we
took
a
centroid
of
a
cell
also
told
me
to
start
working
on
the
development.
There
are
so
many
deprecated
libraries
it's
not
working
properly.
He
showed
me
in
the
last
meeting.
The
tests
are
failing.
Almost
the
library
is
almost
deprecated.
D
D
Normal
test
cases
for
cell
segment
or
limit
cell
membrane
segmenter
and
the
Gan
model,
but
the
tests
are
not
not
pausing
for
not
passing
for
the
lineage
population
mode.
There
are
some
deprecated
libraries,
mainly
psychic
psychic
land
Library,
which
is
not
passing
the
test
cases,
so
I
was
going
through
them.
Maybe
yeah
I'll
show
it
here
and
try
to
update
it,
but
I
couldn't
get
time
due
to
YouTube's
traveling
yeah.
D
Just
let
me
show
it
once
yeah
here
it
is
mentioned,
like
scikit-learn
library
is
almost
difficulties
in
here
yeah,
you
read
some
yeah
psychic
land
yeah
is
mentioned
like
no
mod
in
remote
cyclic,
because
skeleton
has
been
changed
to
another,
but
yeah.
That's
what
the
problem
it
is.
I
will
update
this
requirements.txt
and
that's
trying
to
integrate
depend
about
which
I
have
told
last
week,
because
we
are
not.
The
libraries
are
duplicated.
It
automatically
changes
the
duplicated
language.
B
Yeah
yeah
that
sounds
great
yeah.
Then
the
question
you
had
about
centroids
or
like
how
to
find
a
centroid
or
yeah.
B
We
can't
necessarily
Define
the
nucleus
so
like
when
you
have
like,
let's
say,
a
an
image
that
like
a
fluorescent
image
where
you
have
like
this
blade
area
inside
the
cell-
that's
usually
the
fluorescent
marker,
but
that's
not
necessarily
in
the
nucleus.
It
could
be
in
the
membrane
as
well.
So.
B
Centroid
of
the
membrane
and
find
like
the
center
point
now.
That
being
said,
it
doesn't
necessarily
have
to
be
based
on
the
membrane
it
could
be
based
on,
like
you
know,
some
other
Criterion,
if
you
think
it's,
if
you
think
there's
some
other
way
to
do
it,
that's
better
yeah,
I,.
B
C
D
After
that,
upon
the
time
frame,
then
the
segmented,
so
we
will
I,
will
try
to
find
out
the
location
of
the
cell
that
we
can
use
a
centroid
rather
than
a
membrane
centroid,
that's
what
I
was
thinking
I
mean
if
you
think
this
has
a
3D
space
here,
the
cell
is
located
so
that
so
located
cells,
Android
position
in
the
XYZ
plane.
It
will
be
mentioned.
D
D
The
a
two
comma
three
comma,
four
point
or
yeah,
rather
than
a
centroid
in
a
membrane,
we
can
represent
a
segmented
cell
position
according
to
the
time
frame,
yeah
yeah.
What
my
thinking
is
I
will
try
to
implement
it.
I
will
let
you
know
by
next
week.
I
just
need
to
try
in
this
model.
Still
there
are
so
many
bugs
coming
out.
I
tried
out
to
trying
today
only
but
train
it
today.
Only
the
image
architecture
is
so
different.
D
I
I,
don't
want
to
like
see
how
this
goes
on,
but
I
will
try
this
week
completely
on
this
spot.
I
didn't
get
anywhere,
I'll
start
working
on
that
it's
taking
a
lot
of
time,
rather
than
I
expected.
You
know.
B
To
request
your
extension,
so
let
me
know-
and
you
know
yeah
I'll
just
I'll
just
do
that.
D
Now
we
have
the
ground
troops
of
The
Masks
right
for
every
cell
image,
so
from
the
ground
to
mask
I
was
extracting
the
bounding
box
and
that
bounding
box
was
given
to
Sam
from
as
a
prom.
Then
it
will
segment
the
already
it
was
pre-trained
right,
so
it
will
segment
the
required
cell
in
that
bounding
box,
and
that
is
taken
us
to
compare
with
the
loss
without
General
ground
openness.
That's
how
it
is
made.
Let
me
show
it
to
you.
D
Okay,
this
is
our
picture.
Actually,
this
input
is
referred
to
an
image
encoder.
So
here
is
the
mask,
but
we
need
to
give
the
model
not
mask
generally
what
we,
when
we
Trend
any
model.
We
generally
give
the
crown
truth
mask
and
we'll
try
to
train
it,
but
here
we
can
see
we
can
give
a
prompt
as
a
point
or
a
box
bonding
box
and
it
text
also,
but
this
version
of
thing
is
not
released,
yet
they
only
given
the
version
which
we
can
segment
to
box
or
reports.
D
So
for
the
prompt
encoder,
we
are
giving
the
bounding
box,
which
will
be
in
the
form
of
x,
max
X,
Y,
Max,
X,
Min
and
Y
minimum
of
the
box,
and
that
forms
it
for
the
Box
around
it
and
that
is
given
to
a
prompt,
encoder
and
the
image
is
given
image
encoder.
Then
we
will
get
a
mask
at
our
image
decoder.
D
What
model
is
predicting,
then
we
will
compare
it
with
our
original
mask
and
then
we
will
try
to
find
it
out
whether
we
are
will
try
to
calculate
the
loss
of
the
model.
Yeah,
that's
what
how
the
model
works.
This
is
our
texture
of
the
model,
yeah
right.
D
Yeah
we
generally
give
a
bonding
box
or
a
point
just
a
position
on
the
object
itself
that
generates
the
position.
I
mean
that
gives
that
when
we
give
to
the
model,
we'll
try
to
predict
that
whatever
or
the
surrounded
outings
are
there,
it
will
try
to
segment
it,
and
then
we
will
compare
with
the
ground
truth.
What
we
have,
then
the
losses
are
calculated.
That's
what.
C
B
C
C
Not
a
lot
of
updates.
Just
I'm
gonna
try
to
be
doing
a
little
bit
more
in
the
space
and
really
cool
to
see.
The
developments
have
been
made
here,
so
maybe
more
in
the
next
few
weeks.
But
it's
good
to
see
everybody
this
time
and
that's
all
take
care
all.
B
Right
all
right,
thanks
for
attending,
have
a
good
week.