►
From YouTube: DevoWorm (2021, Meeting 13): Cellular Automata Rules, Neural Organoids, Wiring Neural Circuits II
Description
How to Program Cellular Automata and create rules of life, DevoWorm onboarding guide, Project board updates, Neural Organoids, wiring neural circuits II. Attendees: R Tharun Gowda, Ujjwal Singh, Jesse Parent, Richard Gordon, Susan Crawford-Young, Krishna Katyal, Bradly Alicea, Mayukh Deb, Mainak Deb, Aswath Narayana, Aayush Kumar, Shruti Raj Vansh Singh, Akshay Nair, and Muhammed Abdullah
A
B
B
A
A
A
A
A
A
A
A
A
It's
the
international
neuroinformatics
coordinating
facility
and
so
they're
enabling
a
lot
of
open
science
and
they
actually
sponsor
the
gsoc
program
that
we
participate
in
and
so
I'm
presenting
to
them.
I
think
they,
like
the
fact
that
I
tied
that
into
the
talk-
and
this
is
this
is
like
this
didn't
happen
last
year,
but
it
happens
every
year.
A
Usually
it
was
in
warsaw,
poland
in
2019,
and
I
had
a
student
who
attended
that
con
congress
and
he
actually
did
a
demo
of
his
software
at
that
event,
and
he
said
it
bustered
a
lot
of
interest
in
what
he
was
doing.
So
that's
good,
so
I
don't
know
if
they
have
a
page
for
the
assembly
yet,
okay,
this
is
it.
So
this
is
the
assembly.
A
A
A
A
Open
innovation
and
intellectual
property
data
science
and
neuroinformatics
so
yeah,
it's
it's
gonna,
be
a
a
pretty
decent
event.
I
don't
know
why
it
ends
at
the
23rd.
I
think
they
have
events
afterwards
that
are
that
go
on
for
you
know
a
couple
days
after
that,
but
anyways
that
should
be
an
interesting
event.
A
C
D
A
Dick
says
clay
is
100
meters
deep
on
mars,
which
I'm
taking
as
representative
of
haitian
earth,
which
of
course,
is
like
just
earth
of
the
distant
geologic
past,
so
they're
working
on
a
paper
for
the
origin
of
life.
So
that's
what
that's
about
and
we
have
a
sort
of
a
separate
group
of
people
for
that.
So,
if
you
want
to
join
in
on
that,
let
us
know,
let's
see,
did
anyone.
I
I
asked
this
earlier,
but
only
a
hand
a
fraction
of
you
were
here.
C
E
E
E
E
D
D
E
E
It
would
be
really
helpful
for
you
to
understand
what
different
cellular
automata
with
different
rules
look
like
yeah,
so
the
game
of
life
was
basically
created,
keeping
biology
in
mind
how
from
a
very
simple
orgasm
like
bacteria
and
the
whole
thing,
the
whole
life
on
earth
started.
It
is
what
it
is
trying
to
replicate
and
it
does
it
very
beautifully.
It
passes
the
tuning
test
and
stating
that
yes,
very
simple,
simple,
simple
library
that
this
can
imitate
the
complex
environment
around
us.
So
this
apparent
game
is
a
zero
player.
D
E
E
There
were
like
256
rules
because
there
were
eight
combinations
that
were
possible,
but
here
what
is
it
that
we
have
for
a
single
cell
since
it's
2d,
so
the
neighbors,
the
number
of
neighbors
for
each
cell
would
be
nine,
and
hence
there
will
be
512
possibilities
of
rules
and
having
so
many
rules
and
calculating
basically
based
on
them
would
be
like
really
really
tedious.
So
what
we
have?
We
have
generalized
the
root
in
the
game
of
life,
and
we
have
just
put
it
down
with
three
things
that
is,
the
death
birth
and
status.
E
Still
that
happens
because
of
two
conditions,
operation
and
loneliness
like
in
general,
like
when,
there's
too
much
of
anything
or
too
less
of
anything.
You
know
we
have
people,
not
people
die,
but
actually
they
said
here
guys
where
the
overpopulation
means
that
if
this
cell
has
more
than
four
active
members
that
the
neighbors
here
are
the
more
than
four
of
them
are
number
one.
Their
group
died
or
if
they
are
less
than
one
one
or
zero
number
of
active
labels
would
also
lead
to
the
death
of
the
cell.
E
E
And
there
won't
be
any
death
or
birth
here
that
you
can
see
there
yeah.
So
when
you
implement
these
rules,
some
very
fascinating
patterns,
others
some
are
static
and
some
are
not.
For
example,
the
top
ones
which
you
see
here
are
like
the
block
that
we
had
the
books.
These
are
the
static
ones
and
they
do
not
change
the
next.
E
One
is
the
this
one
which
keeps
on
oscillating
to
using
all
these
combinations
of
you
know
from
it
always
starts
from
a
very
single
cell,
and
then
it
replicates
itself
with
time
and
oh,
I
guess
a
picture
is
missing
here.
Nevertheless,
so
yeah,
what
happens
is
when
you
put
all
of
this
together
in
the
form
of
a
bigger
picture
on
a
grid
on
a
2d
grid.
What
you
can
see
is
a.
E
D
E
E
D
E
G
H
E
By
a
cellular
automata
where
they
take
the
image
and
convert
it,
and
you
know
using
the
image
instead
of
new,
just
a
neural
network,
what
did
it
take?
Is
a
cellular
automata
to
get
the
for
semantic
image
segmentation
and
apparently
it
shows
very
good
results.
They
have
compared
the
neural
network
with
the
cellular.
E
E
E
D
E
E
And
we
handle
it
really
well
and
we
can
explore
the
data
for
further
analysis
as
well.
It
was
basically
developed
by
ann
carpenter
she's
from
broad
institute,
and
it
is
something
that
is
used
very
very
very
now.
Nowadays,
it's
in
trend
right
now
for
bio
images.
So
some
of
the
features
where
this
this
software
was
something
that
caught
my
eye
was
because
from
a
series
of
image
processing
modules-
and
we
already
have
some
predefined
because
since
it's
for.
E
Some
pipelines
for
for
image
segmentation
for
for
labeling,
for,
in
fact
not
just
for
that
for
nuclei-
and
I
am
not
very
good
at
all
this
biology
stuff
but
yeah
for
literally
a
different,
different
aspects
of
phenotypes.
I
guess
they
are
called
for
different
different
aspects.
They
have
different
different
modules,
smaller
modules,
which
we
can
work
on.
Also
we
can
adjust
the
settings
according
to
our
need.
We
can
label
it
according
to
if
you
want
the
brightness
to
be
more
to
less
separating.
E
D
E
E
So
I
guess
this
this
tool
is
something
that
can
work
really
well
for
us,
especially
for
biologists,
who
is
particularly
made
for
that,
and
I
guess
I
will
be.
I
will
be
working
more
on
it
and
exploring
what
I
I
just
discovered
yesterday
and
I
did
not
have
a
lot
of
time
to
go
through
at
all,
but
I
hope
this
will
be
something
that
will
help
us
a
lot
in
our
projects.
E
H
E
G
B
G
H
G
E
H
A
D
E
A
A
E
E
E
B
E
E
A
A
D
A
E
E
A
A
This
is
something
that
we're
going
to
be
I'm
going
to
be
presenting
it
on
behalf
of
the
group's
activities
on
behalf
of
divalern
and
a
couple
other
things
and
we'll
go
through
the
slides
next
week,
so
I
haven't
quite
gotten
them
together
yet
and
then
we'll
you
know
this
will
be
on
the
week
of
the
19th.
I
believe
so
this
is,
you
know,
look
into
it
and
see
if
you
want
to
attend
so
okay.
So
let
me
close
that
out.
A
A
I
wanted
to
make
this
very
general
so
that
it
contains
a
lot
of
different
things
that
we
have,
so
you
can
see
there's
an
index
up
at
the
top
and
it's
you
know
maybe
becoming
a
little
too
long,
but
I'm
trying
to
that's
why
I
built
this
index
so
that
we
have
you
know
if
people
want
to
go
to
different
places
in
the
you
know
different
different
topics.
They
can
go
there.
So
you
know
one
of
the
things
that
I
found
in
past
years
and
I
haven't
had
a
really
good
way
to
deal
with.
A
Is
people
want
to
become
involved
in
the
group
and
they
don't
really
have
a
background
in
biology
where
they
want
to
find
some
data
sets
to
work
on
where
they
want.
You
know
in
this
case
they
want
to
know
about
the
projects,
so
this
onboarding
guide
I'm
trying
to
build
a
lot
of
that
into
it,
so
that
we
have
it.
Just
that
one
link,
I
can
send
someone
a
link
and
they
can
read
through
it.
A
A
A
A
A
The
resources
we
have
now
are
basically,
like
you
know,
different
basic
introductions
to
the
biology
of
these
systems,
so
for
diet
for
the
bacilluria.
I
have
a
couple
wings
and
diatoms
thomas
harbic's
resource
on
diatoms
and
a
scientific
american
article.
So
it's
it
just
tells
people
about
what
diatoms
are
c.
Elegans
has
some
this.
A
A
You
know
some
image
sample
images
of
c
elegans,
microscopy
images
and
then
some
basic
biology
and
microscopy
and
then
for
axolotl
a
couple
links
as
to
you
know
what
it
is
and
how
it's
important.
So,
if
you're
doing
a
proposal,
I
would
suggest
you
know
going
through
some
of
these
links-
and
I
know
a
couple
of
you
have
reviewed
your
proposals
and
you
didn't
really
put
a
lot
of
emphasis
on
the
problem
that
you're
trying
to
solve.
So
maybe
a
fourth
part
of
structuring
the
proposal
is.
A
Do
a
very
brief
introduction
to
the
thing
you're
trying
to
study
so,
if
you're
applying
to
project
one,
for
example,
you
would
talk
about
c
elegans
and
how
it's
important-
and
you
know
why
you
know
we
might
want
to
study
it.
I
mean
maybe
a
couple
sentences,
not
not
a
very,
not
not
a
whole
page
of
content
just
enough
so
that
we
can
establish.
A
You
know
that,
then
it
gives
you
the
advantage
of
kind
of
getting
your
feet
wet
with
some
of
these
model
systems.
I
think
that's,
that's
a
good
way
to
approach
that,
and
so,
if
you
need
more
references
or
if
you
have
ideas
for
references,
we
can
add
to
this.
Let
me
know-
and
it's
just
you
know
just
a
way
to
get
your
feet
wet.
We
have
other
resources
on
some
of
our
other.
A
A
A
The
open
word
movement
database,
which
is
something
that
is
run
by
a
different
group
in
open
worm,
but
they
do
c
elegans
movement,
and
so
we've
actually
used
that
in
years
past,
because
for
some
projects,
because
it's
actually
a
very
nice,
it's
a
nice
data
set,
it's
annotated
and
it's
it's
adult
c
elegans
moving
around
in
different
ways.
So
they've
taken
like
images
of
different
movements
that
c
elegans
makes
c.
Elegans
makes
a
series
of
stereotypical
movements
and
those
movements
then
are
classified.
A
You
can
classify
them
using
machine
learning.
You
can
understand
more
about
c
elegans
movement
patterns,
so
those
are
things
that
you
might
draw
from,
especially
if
you're
doing
like
machine
learning,
it
could
be
a
benchmark
or
it
could
be
something
you
know
you
might
do
something
interesting
with
a
movement
with
some
of
the
movement
data.
A
I
think
mayuk
did
some
work
with
this
and
a
couple
other
people
have,
you
know,
been
interested
in
it
in
years
past.
So
that's
something
to
keep
in
mind
and
then
I
have
a
a
description
of
some
of
the
types
of
biological
analysis
that
we
might
do
with
some
of
the
data.
So
once
we
have
the
data,
what
do
we
do,
and
this
is
something
maybe
it
needs
to
be
flushed
out
a
little
bit
more.
A
A
So
you
know
the
idea
is
that
we
do.
You
know
we're
interested
in
doing
a
lot
of
this
machine
learning
stuff
and
this
analysis
of
images.
But
then
you
know
we're
making.
We
have
to
make
connection
on
the
other
end,
which
is
to
do
really
interesting
types
of
analysis
with
the
data
that
we
extract,
and
so
that's
how
we
build
a
paper.
It's
basically
that
two
fold
strategy
we're
gonna,
extract
a
bunch
of
things
out
of
images
and
then
we're
gonna
put
them
into
this.
A
These
types
of
analysis
and
then
we're
gonna
show
the
world
what
that
what
we
can
do
with
it?
It's
basically
the
idea.
So
you
know
if
you
want
to
be,
if
you
want
to
get
involved
in
that
that
you
know
that's
a
pathway
to
a
paper,
it's
a
pathway
to
something
very
you
know
potentially
very
interesting,
so
yeah.
So,
let's
you
know,
if
you
have
suggestions
for
this
onboarding
guide,
please
make
them
available.
A
C
A
A
A
I
think
I
need
we
need
to
have
like
probably
three
or
four
times
that
length
for
this
event,
so
we're
gonna,
I'm
gonna,
add
some
slides
in
and
I'll
pass
it
around
to
my
oak
and
maybe
some
other
people
and
we
can
put
together
a
really
nice
presentation,
and
I
think
this
will
be
a
nice
showcase
to
the
incf
people
to
show
them
that
we're,
inter
you
know,
we're
really
building
upon
what
they're
sponsoring
for
us.
So
this
is
the
divalern
repository.
A
A
Repository
and
that
was
we
had
shruti,
we
had
josh,
we
had
a
couple.
Other
people
contribute.
So
this
is
you
know
this
is
a
nice
way
to
like.
You
know
again,
get
your
work
out
into
the
world
and
I
don't
know
if
sherdy's
interested
in
pushing
her
cellular
automata,
so
the
divalern
platform
isn't
just
the
diva
learned
software.
We
have
other
types
of
things
going
on,
and
so
we
have
the
diva
organisms.
A
I
guess
maybe
I
want
to
check
before
I
suggest
this,
but
I
was
wondering
if
surety
was
interested
in
pushing
some
of
her
cellular
automata
stuff
to
a
repo
that
I
will
make
here
and
we
can
put
it
into
this
as
a
repo,
for
you
know,
other
types
of
models
that
are
outside
of
sort
of
the
standard
machine
learning
type
of
model,
this
general
biological
model
is
something
krishna
proposed,
and
I
made
that
separate.
So
that's
you
know
that's
there,
but
I
don't
want
to
put
it
there.
A
E
A
So
I
think
that's
then
I
wanted
to
go
over,
maybe
some
of
our
issues
that
we
have
on
our
board.
Let's
see
oops
projects,
your
tasks.
So
again
we
have.
This
is,
I
know,
jesse's
been
helping
to
organize
this,
but
there's
still
a
lot
of
stuff
in
here.
That's
and
we
need
to
update
it
a
little
bit
more
because
things
are
always
moving.
So
we
don't
know
we
don't
know
where
things
stand.
We
have
a
lot
of
different.
A
A
Let
me
put
a
link
to
it
here,
so
this
is
a
github
project
board
and
you
know
we
have
our
different
issues
here,
and
these
are
github
issues.
These
aren't
necessarily
related
to
code
related
tasks,
they're
just
like
issues
that
we
have
open.
So
we
have
the
evolution
conference
presentations,
for
example.
A
This
is
requires
submission
of
a
an
abstract
by
one
of
us,
or
you
know
the
revision
of
an
abstract,
and
so
once
that's
done,
then
we
move
that
to
say
like
finished,
we
have
other
issues
like
update,
divorm
ml
lectures,
that's
not
something
necessarily
that
will
get
done
in
one
shot.
You
know,
it'll
be
like
a
gradual
thing,
so
that
will
stay
in
action
items
until
it's
finished
and
then
you
know
we
have
our
divor
bibliography.
A
A
A
A
I
don't
have
any
chance
to
work
on
it,
since
I
showed
it
to
you
so
then
we
have
things
like
getting
data
from
different
places
like
susan's,
ball,
microscope
and
she's.
Obviously
working
on
that,
as
she
said
before
in
the
meeting
when
we
have
other
things
that
are
so,
we
have
like
this
category
of
things
that
are
off
the
radar
too.
So
if
you're,
if
there's
something,
we've
talked
about
a
lot
of
things
in
the
meeting
and
some
things
kind
of
fall
off
the
radar,
some
things
get
revived.
A
A
You
know
in
a
slack
message
or
email
once
I'm
interested
in
this
issue.
Where
are
we
with
it,
and
I
can
kind
of
tell
you
yeah,
so
we
have
all
sorts
of
even
tutorials
for
youtube
if
you're
interested
in
doing
a
tutorial
on
some
topic-
and
you
know,
recording
a
short
video
of
yourself
presenting
on
it,
we
could
put
that
up
on
our
youtube
channel
that
could
work.
A
So
there
are
a
lot
of
ways
to
contribute
here
and
we
don't
have.
You
know
it's
still.
It's
purposefully
busy
because
there's
a
lot
of
stuff
going
on,
but
so
that's
our
project
board
and
I
know
I
keep
I
sometimes
I
go
through
the
issues
one
by
one,
but
sometimes
that's
not
particularly
effective,
because
it's
kind
of
you
know
oftentimes.
We
just
have
things
that
yeah
we're
following
up
on
that:
okay,
so
they
have
something
in
the
chat.
A
Bradley,
surety
and
krishna
are
working
with
me
on
the
fish
ladder
toy
model
for
a
thermodynamically
at
equilibrium.
Origin
of
life
in
a
liquid
world
could
be
added
okay.
So
this
is
the
fish
ladder
toy
model,
thermodynamically
equilibrium,
origin
of
life
in
a
leopard
world.
Do
you
have
any
smaller
shells
like
you
shouldn't
earlier
to
dick
yeah?
Let
me
add
that
to
the
list:
let's
see,
what
is
it?
What
is,
let's
see,
how
would
I
summarize
it?
The
fish
letter
toy
model
of
life
on
a
leopard
world.
A
So
this
is
a
project
issue,
so
this
is
something
that
I
don't
know
if
I
can
who
I
can
assign
this
to
who's
in
the
list
of
contributors.
So
if
you're
like
a
member
of
this,
it's
it
brings
you
up.
I
don't
know
if
I
can
issue
it
to
people
here.
Oh
assignees,
yeah
yeah.
I
don't
have
any
the
people
on
the
list,
but
that's
okay.
If
you
want
to
join
this
board,
you
can
become
a
contributor.
A
A
A
Is
maybe
this
is
actually
going
to
be
like
something
we'll
get
a
lot
of
data
on
at
this
point,
so
I
made
two
projects
here,
based
on
the
issues
in
the
chat.
Jessie
says:
origin
of
life
projects
are
interesting.
Do
you
need
our
email,
ids
yeah?
If
you
well,
all
you
need
to
do
to
become
a
contri
or
a
collaborator
on
the
project
board
is
a
send
me
a
note
with
an
email
I
can
invite
you
via
email
and
then
you'll
receive
an
email,
invite
and
you
you
become
an
you,
become
a.
D
A
I
guess
of
that
repository,
so
you
can
be
assigned
to
tasks
and
you
I
think
you
can
make
your
own
pushes
to
the
repository
so
that'd,
be
a
good
that'll,
be
a
good
way
to
get
people
involved
in
some
things
here.
So
I
wanted
to
move
on
here
to
our
papers
for
the
weeks
and
we
finally
have
some
some
time
here
to
papers
and
I'm
gonna
go
probably
a
little
bit
over
the
top
of
the
hour.
A
A
So
you
can
build
a
model
of
a
cell
as
just
a
sphere
or
a
sphere
with
like
maybe
a
couple
of
parts
like
a
nucleus,
or
you
know,
maybe
some
other
types
of
things
in
it.
So
you
know
if
you're
doing
modeling.
You
know
that,
like
it's
a
matter,
it's
a
matter
of
what
level
of
abstraction
you
want
to
pick
to
model,
so
you
can
model
a
black
box.
Just
a
you
know,
a
shape
that
looks
like
a
cell
and
you
might
be
interested
in
how
it
moves
in
its
environment.
A
You
know
the
nucleus
for
whatever
reason,
but
what
they're
doing
is
they're
trying
to
build
what
they
call
a
realistic
model
of
the
cell
or
high
fidelity
model
of
the
cell,
and
so
in
this
case,
what
they're
looking
at
is
they're
constructing
the
whole
cell
3d
model.
So
they
have
this
pipeline
where
they
do
this
protein
structures,
reconstruction.
A
They
gather
and
validate
structural
data
for
each
protein
structures
are
retrieved
from
structural
databases
are
generated
via
homology
modeling.
The
final
recipe
includes
482
protein
monomers
and
201
protein
complexes
and
they're.
Actually,
you
know
they're
not
working
on
a
human
cell
they're
working
on
this
mycoplasma
genitalium
bacteria,
so
this
is
a
very
simple
cell
as
compared
to
something
like
a
human
cell
or
any
sort
of
eukaryotic
cell.
So
it's
a
you
know:
it's
not
the
I
mean
even
with
like
something
that's
a
fairly
simple
cell
was
at
least
as
we
think
of
it.
A
A
So
now,
they're
looking
at
modeling
nucleic
acids
these
this
lattice-based
method,
they're
using
you
know,
they're
modeling,
the
genome
of
the
bacteria,
which
is
much
simpler
than
ours,
but
still
requires
a
lot
of
work
to
model
it,
and
then
they
simulate
the
3d
organization
of
the
cell.
So
now
with
this
with
the
other
two
parts,
then
they
put
it
into
this.
A
A
So
this
is
actually
then
they
say.
Well,
you
can
explore
the
cellular
mesoscale,
so
they're,
looking
at
using
you
know
different
color
palettes
to
look
at
different
features
in
the
cell.
So
when
you
look
at
this
thing
here,
it
looks
like
you
can't
make
any
sense
of
it,
and
indeed
you
can't,
if
you're
just
looking
at
it,
but
one
of
the
things
they
do
here
is
they
use
color
coding
to
pull
out
different
features.
A
So
this
is,
for
example,
looking
at
copy
number
variation.
This
is
looking
at
structural
confidence
and
then
this
is
looking
at
the
source
of
structural
data
and
they
have
a
legend
underneath
here
to
tell
us
where
you
know
what
these
mean.
So
you
know
that's
that's
the
way
you
end
up.
You
know
you
end
up
building
this
huge
model,
and
then
you
have
to
figure
out
a
strategy
of
how
to
navigate
all
those
data.
You
know
another
way
to
navigate.
A
So,
if
you're
interested
in
very
specific
genes-
or
you
know
other
types
of
things
in
the
model,
you
can
find
them.
So
this
is
a
nice
I,
this
is
a
poster
by
here
are
the
authors
here.
This
was
a
poster.
I
can't
remember
where
I
got
this
from.
It
was
off
the
web,
but
it's
sponsored
by
scripps
research
and
I
will
put
the
link
in
the
chat.
A
A
A
They're
talking
about
cortical
organoids,
matching
key
early
post
needle
transitions,
so
the
abstract
is
human.
Stem
cell
derived
models
provide
the
promise
of
accelerating
our
understanding
of
brain
disorders,
so
they're
modeling
the
cortex
in
the
brain
in
mammals,
and
this
is
the
area
of
the
brain
that
we
often
associate
with
cognition
and
other
types
of
behaviors
in
invertebrates,
but
not
knowing
whether
they
possess
the
ability
to
mature
beyond
mid
to
weight.
Fetal
stages
potentially
limits
their
utility,
so
they
have
these.
A
So
we
leveraged
a
direct
differentiation
protocol
to
comprehensively
assess
maturation
in
vitro
based
on
genome-wide
study
analysis.
It
means
they
did
it
in
a
culture
dish
instead
of
like
in
an
organism
based
on
genome-wide
analysis
of
the
epigenetic
clock
and
transcriptomics,
which
is
how
genes
are
expressed
and
like
epigenetic
effects
of
that
as
well
as
rna
editing,
we
observed
that
the
three-dimensional
human
cortical
organoids
reach
post-natal
stages
between
250
and
300
days,
a
timeline
paralleling
in
vivo
development.
A
We
demonstrate
the
presence
of
several
known
developmental
milestones,
including
switches
in
the
histone
d-cellotase
complex
and
the
nmda
receptor
subunits,
which
we
confirm
at
the
protein
and
physiological
levels.
These
results
suggest
that
important
components
of
an
intrinsic
and
vivo
developmental
program
persist
in
vitro
meaning
that
there's
this
program
that
is
set
in
in,
like
you
know,
in
a
in
an
embryo
that
you
can
replicate
in
an
in
vitro
model.
If
you
have
it
set
up
right.
A
We
further
map,
neurodevelopmental
and
neurodegenerative
disease
risk
genes
onto
in
vitro
gene
expression,
trajectories
and
provide
a
resource
and
web
tool,
and
it's
this
gco
or
gene
expression,
cortical
organoids
site
to
guide
disease
modeling.
So
this
is
pretty
heavy-duty
molecular
biology
and
disease
science
or
disease
biology.
A
But
the
lesson
here
is
that
you
can
build
these
three-dimensional
cellular
ensembles
that
are,
you,
know,
self-organizing,
and
that
form
these
sort
of
in
vitro
structures
that
resemble
tissue
development.
So
they're,
not
you
know,
they're
not
replicating
development
as
it
might
be
replicated
in
the
in
the
body
in
the
in
the
in
the
embryo.
I
should
say,
but
it's,
but
it's
something
that
we
can
learn
lessons
from
so
yeah,
so
they
what
they
do.
A
Is
they
embed
these
cells
into
a
matrix
and
they
have
the
matrix
as
nutrients
and
things
like
that
and
they're
able
to
grow
not
just
with
a
large
number
of
cells,
but
with
shape
they're
able
to
assign
a
shape
and
they're
able
to
assign
some
geometry,
and
this
is
how
they're
able
to
build
these
organoids
so
they're.
Actually
the
first
group
really
to
sort
of
give
a
systematic,
unbiased,
functional
analysis
to
demonstrate
the
maturation
of
these
different.
A
You
know
the
the
structure
as
it
goes
through
these
different
stages,
which
are
parallel
to
what
you
see
in
an
embryo
really
yeah
a
question:
was
there
any
indication
that
they're
doing
any
time
lapse
where
anybody
has
organized?
Let's
see,
I
don't
know
what
they
have
in
this
paper
in
terms
of
the
figures.
Let's
see
what
they
have
in
the
figures.
A
B
A
A
A
Right
so
I
mean
yeah
and
I
don't
know
how
these
like
the
organoids
are
grown
in
these
controlled
environments.
I
don't
really
know
what
the
the
details
are
with
respect
to
how
they
like
change
out
the
nutrients
in
the
medium
or
the
matrix.
How
that's,
how
that's
done,
how
the
imaging
can
be
done?
You
know
if,
if
it's
something
that
you
can
actually
do
like
you
know
time
lapse
on
or
if
it's
something
that
you
can.
A
A
Looking
at
specific
things
being
expressed
in
the
cells
and
so
for
those
of
you
who
aren't
you
don't
have
a
lot
of
background
in
biology,
that's
sort
of
a
you
know:
they'll
do
this
for
different
markers,
and
so
sometimes
you
have
to
do
massive
numbers
of
experiments
to
get
like
some
of
these
maps
that
they
have
where
they
have.
You
know
three
or
four
hundred
genes
that
you
can
select
in
a
model
that
they
have
in
an
interface.
A
lot
of
that
is
like
the
serial
experiments
that
they
have
to
perform
to
get
this
information.
A
A
Is
a
lot
harder
than
you'd
like
to
have
it,
and
so
they
just
kind
of
describe
this.
These
different
modules,
which
are
different
parts
of
the
organoid,
and
they
have
you
know
so
they
look
at
the
differentiation
day
and
they're.
Looking
at
these
different
markers,
which
are
these
different
colors.
So
these
different
color
functions
represent
the
distribution
of
these
different
things.
So,
like
the
black
line
are
immune
response
genes,
the
white
cyan
is
extracellular
matrix,
so
these
are
just
different
genes.
A
That
code
for
different
things
in
this
in
the
embryo,
and
then
these
functions
are
the
distribution
of
these.
So
these
dots
are
different
colors
and
then
these
lines
are
the
sort
of
the
average
of
these
different
dots
over
time,
so
you're
sampling
at
different
times
of
differentiation,
and
you
can
see
that
these,
like
in
different
parts
like
the
glio
modules,
the
neuronal
modules
they're,
just
comparing
all
of
those
in
that
category,
so
they're
using
something
called
go
terms
which
are
just
something
that
is
generated
for
a
certain
gene
by
external
groups.
A
They
look
at
like
how
the
gene
is
expressed
and
they
give
it
a
term
and
then
that's
supposed
to
tell
you
what
it
does
and
it's
an
imperfect
method,
but
they
can
basically
then
look
at
gene
expression,
sort
it
by
these
categories
and
then
give
you
these
sort
of
trends
over
time.
So
you
know
dark
green
there's,
not
a
lot
of
variation
here.
There's
this
purple
that
goes
up
early,
that's
the
axon
guidance,
neuronal
migration
and
gtpase
activity,
so
that
happens
rather
early,
and
then
you
have
things
like
the
pink.
A
A
A
This
is
where
the
nervous
system
is
put
together,
and
you
can
see
that
there's
a
spike
at
about
100
to
300
days,
where
there's
a
lot
of
activity
going
on
and
it
trails
off
slightly
towards
the
end
of
their
time
series.
So
you
know
it
sort
of
mimics
human
development,
but
there
are,
of
course,
going
to
be
differences
in
these
two
and
that's
you
know.
A
The
idea
is
to
get
a
a
good
sense
of
what's
going
on
in
development,
so
we
can
maybe
study
diseases,
maybe
like
a
disease
state
versus
a
normal
state
and
by
disease
state
I
mean,
if
you
have
like
a
gene
for
some
disease
versus
a
gene,
not
a
gene
for
some
disease.
We
can
look
at
the
differences,
so
this
is.
There
are
a
lot
more
images
in
this
paper.
If
you
want
to
look
at
that
and
get
a
sense
of
what's
going
on
there,
another
paper
I'll
go
into
here.
A
A
Nevertheless,
these
models
are
poor
models
of
the
development
of
the
visual
system
because
they
posit
millions
of
sequential,
precisely
coordinated,
synaptic
updates
each
based
on
a
labeled
image,
while
ongoing
research
is
pursuing
the
use
of
unsupervised
proxies
for
labels.
Here,
we
explore
a
complementary
strategy
of
reducing
the
required
number
of
supervised,
synaptic
updates
to
produce
an
adult-like
ventral
visual
system.
A
So
what
they're
saying
is
is
that
in
machine
learning
we
use
labels
and
we
kind
of
label
objects
and
then
that's
how
we
sort
of
work
around
like
we
see
a
cup,
and
you
know
a
machine
learning
algorithm
sees
a
cup,
but
it
needs
a
category
to
latch
onto
to
say
it's
a
cup
or
not
well.
What
humans
do,
however,
is
different.
We
don't
really
necessarily
need
the
categories.
A
You
know
the
system
doesn't
break.
If
we
don't
have
the
categories,
we
can
actually
learn
the
categories
with
some
language,
but
the
language
is
emergent,
and
so
our
visual
systems
can
recognize
a
cup
versus
maybe
like
a
table
without
labels.
We
can
do
that
in
infancy
and
then
we
eventually
learn
the
labels
and
attach
them-
and
you
know,
there's
meaning
involved.
So
it's
a
very
different
system,
so
they're
exploring
the
strategy
and
so
they're.
A
Such
models
might
require
less
precise
machinery
and
energy
expenditure
to
coordinate
these
updates,
and
so
this
kind
of
approach
would
move
us
closer
to
sort
of
this
neuroscience
hypothesis
of
visual
system,
wiring
and-
and
you
know,
maybe
improve
our
ability
to
recognize
visual
thing
in
you
know:
objects
in
in
visual
in
the
visual
sense,
so
they
relative
to
the
current
leading
model
of
the
adult
ventral
stream,
which
we've
we've
characterized
this.
So
it's
not
with
something
they're
starting
off
from
scratch
on.
A
We
here
demonstrate
that
the
total
number
of
supervised
weight
updates
can
be
substantially
reduced
using
this
cop,
these
three
complementary
strategies.
So
first
we
find
that
only
two
percent
of
supervised
updates,
epics
and
images
are
needed
to
achieve
about
eighty
percent
of
the
match
to
adult
ventral
stream.
A
A
Third,
we
find
that
by
training
only
five
percent
of
the
model
synapses,
we
can
still
achieve
nearly
eighty
percent
of
the
match,
the
ventral
stream.
So
there's
a
minimal
amount
of
training
here
as
well.
The
training
comes
in
like
in
the
human
visual
system.
There's
you
know.
A
lot
of
a
lot
of
connections
are
produced
at
birth
and
then
those
connections
are
pruned
through
these
experience
dependent
situations.
A
The
world
that
you're
experiencing,
so
you
don't
need
to
train
the
model
as
much
using
this
type
of
an
approach.
So
when
these
three
strategies
are
applied
in
combination,
we
find
that
these
new
models
achieve
about
80
percent
of
a
fully
trained
model's
match
to
the
brain,
while
using
two
orders
of
magnitude,
fewer
super
supervised,
synaptic
updates.
A
A
So
in
my
other
group
we're
talking
a
lot
about
development
and
artificial
intelligence,
and
so,
but
I
thought
this
paper
would
be
interesting
to
this
group
as
well,
because
we're
interested
in
a
lot
of
issues
related
to
artificial
neural
networks.
So
we
do
this
so
they're.
Actually,
looking
at,
like
you
know
the
connections
between
neurons
here.
A
A
This
happens
across
different
species
that
have
a
prima
or
well
of
mammalian
visual
system,
and
so
anthony
zador
wrote
a
paper
recently
last
couple
years
on
innateness
in
in
sort
of
artificial
intelligence
or
deep
learning
systems,
and
he
says
that
you
know.
Maybe
we
need
to
have
an
innate
component
to
these
models
in
order
for
them
to
perform
better,
and
so
they
kind
of
go
into
that.
A
But
can
we
do?
Can
we
build
a
model
that
starts
up
kind
of
like
a
biological
model,
where
you
have
this,
this
innate
ability
to
sort
of
pick
things
out
of
the
environment?
And
then
you
know,
if
you
train
it
on
things,
it
doesn't
necessarily
need
a
lot
of
training
with
labels.
It
just
needs
to
interact
with
its
environment.
A
It
can
do
almost
as
good
as
the
machine
learning
models
that
we
have
now,
so
they
talked
about
pruning,
which
is
where
the
synapses
are
selected
within
development
to
become
more
focused
and
specialized
on
stimuli,
and
this
is
kind
of
the
approach
that
they're
using,
and
so
they
kind
of
go
through
the
literature
on
this.
If
you're
interested
in
this
topic,
I
would
read
this
paper
and
look
at
some
of
the
some
of
what
they're
doing
here
their
contributions
to
the
field.
A
So
this
contributions
from
the
field,
they
kind
of
go
through
this
method
that
they're
using
and
they
have
their
code
on
github.
So
if
you
want
to
try
it
out,
you
can
try
it
out
there,
so
they
have
actually
they
have
so
these.
These
approaches
that
they're,
using
they
build
models
of
the
fraction
of
supervised
updates
that
retain
high
similarity
to
the
primary
visual
system.
A
They
they
call
that
brain
productivity.
They
just
want
to
be
able
to
match
up
with
the
biological
model.
We
improve
the
at
birth
synaptic
connectivity
to
achieve
reasonable
brain
productivity
with
no
training
at
all,
so
they
use
this
at
birth
connectivity,
which
is
just
kind
of
a
random,
a
dense,
random
connectivity.
A
We
propose
a
thin
critical
training
technique
which
reduces
the
number
of
train
synapses,
while
maintaining
a
high
print
brain
productivity,
so
they're
using
this
critical
period
as
a
way
to
sort
of
prune
this
network
they're
using
a.
I
don't
know
if
it's
a
critical
period
in
the
conventional
sense,
but
they're
using
this
kind
of
approach,
and
then
we
combine
these
three
techniques
to
build
models.
Two
orders
of
magnitude,
if
you
re,
supervise
synaptic,
updates
with
high
brain
predictivity
relative
to
a
fully
trained
model
and
so
they're
using
this
really
their.
A
G
A
A
Okay,
I'll
I'll
I'll,
take
care
of
that.
So,
let's
see
we'll
go
up
to
the
top
of
the
chat
here,
so
we
have
oh
yeah.
Major
tests
for
2020
needs
to
be
updated
to
2021,
but
this
is
so
jesse
skype
and
email
data
on
the
paper
or
the
data
on
the
organoids
paper
itself
is,
I
think
it's
brand
new.
I
don't
know
oh
2021
yeah
march
of
2021,
so
it's
pretty
new,
so
yeah
this
this
area
of
organoids,
it's
it's
only.
A
You
know
kind
of
an
emerging
the
last
couple
years,
and
so
it's
kind
of
the
wild
west
in
terms
of
methodology,
but
they're
kind
of
converging
on
some
specific
methodologies,
and
so
there's
a
lot
to
learn
in
this
area.
You
know
you
can
build
these
organoids
that
have
like
this.
You
know
geometry
and
this
you
know
in
brain
organ.
In
the
case
of
brain
organoids,
you
can
look
at
the
electrophysiological
activity
and
you
know
it
can
tell
you
some
things
about.
A
A
A
So
this
is
a
paper
that
decorate
with
a
collaborator
vision,
begins
a
direct
reconstruction
of
the
retinal
image
of
the
brain
season,
stores
pictures.
So
that's
something
that
you
know
and
I'm
sure
that
they
have
like
a
you
know,
they're
just
starting
on
like
one
I
don't
know,
I
don't
actually
know
what
data
sets
they're
using.
I
didn't
look
at
the
list,
but
you
know
this
is
an
area
that
people
are
starting
to
pick
up
on
where
they're.
A
Looking
at
like
biological
networks
and
they're
saying,
can
you
replicate
some
of
those
things
for
artificial
neural
networks?
Because
it's
you
know
artificial
neural
networks,
they've
built
them
in
terms
of
size,
they've
built
them
quite
large,
but
they
still
have
a
lot
of
errors
and
performance
that
biological
neural
networks
don't
seem
to
have.
Let
me
explain
what
I'm
talking.
H
Okay,
so
that's
the
standard
approach
to
compute
tomography
now
envision
actually
does
something
similar
envision.
Each
cell
in
the
visual
cortex
has
what's
called
a
receptive
field,
and
these
receptive
fields
tend
to
be
long
and
narrow,
and
so
this
paper
is
about
how
to
reconstruct
images
from
a
set
of
receptive
fields,
okay,
which
receptive
field
is
like
a
line
across
the
image
with
a
certain
width
yeah.
Well,
in
other
words,
what
you're
doing
is.
D
A
H
A
Yeah
yeah
I'll
have
to
take
a
look
at
that.
Thank
you
for
posting
that
then
we
have
jessie
said
zador
innateness
genomic
bottleneck
at
nureps,
oh
yeah.
There
was
a
thing
at
their
at
the
nurbs
conference
on
this
topic,
but
yeah
it's
like.
I
said
it's
moving
pretty
fast,
so
we
might
yeah
actually
look
at
this
more
redick's
paper
and
you
know
look
at
it
more
in
in
future
meetings.
So
and
then
christmas
is
nice
presentation.
So
thank
you
for
your
comments
in
the
chat
yeah.
A
A
Go
okay!
Well,
if
there's
nothing
else
thanks
for
attending
this
week,
I
don't
know
if
we
have
a
you
know.
Some
of
you
are
meeting
after
this
yeah
thanks
as
well,
and
we
will
be
you
know
in
in
slack
or
you
know,
via
email
and
yeah,
let's
stay
in
touch
over
the
course
of
the
week,
so
we
have
a
lot
of
things
going
on.
So
thanks
for
attending
see
you
guys
next
week,
everyone
take
care
bye.