►
From YouTube: DevoWorm (2020, Meeting 39): Reinforcement Learning, General Meta-Feature Models, Neurodevelopment
Description
Task Review Board, discussion/presentation on SARSA reinforcement learning and general meta-feature models, and papers from the reading queue (development of brains, cell types, and meaning). Attendees: Susan Crawford-Young, Krishna Katyal, Shruti Rajvanshsingh, Jesse Parent, Bradly Alicea.
A
Hi,
welcome
to
the
meeting,
doesn't
look
like
anyone's
here
yet
so
why
don't
we
get
started
for
this
week?
So,
if
you're
watching
on
youtube,
welcome-
and
I
hope
that
you
find
this
information
useful
this
week-
we're
going
to
go
through
a
couple-
things:
we're
going
to
go
through
some
open
tasks,
we're
going
to
kind
of
review
things
as
they
stand
and
what
we
want
to
do
for
the
rest
of
the
year
and
then
the
calendar
year.
That
is
and
then
probably
go
over
some
papers
and
some
things
that
we
have
in
the
queue.
A
Sorry,
let's
start
off
talking
about
the
major
tasks
I
will
present
my
entire
screen.
A
And
so
this
is
the
passport.
Actually,
this
is
the
dashboard.
This
is
the
task
board
for
our
group
meetings
task
board,
and
so
this
has
all
the
things
that
we
usually
that
we've
been
addressing
throughout
the
year
in
our
group
meetings,
things
that
are
sort
of
loose
ends,
or
things
maybe
are
done.
B
C
A
I'm
going
currently
going
through
the
issues
board,
so
why
don't
I
go
through
the
issues
board
and
I'm
recording
for
youtube.
So
and
then
we
could
talk
about
something
afterwards.
Oh
there's
krishna
too.
D
A
I'm
fine,
how
are
you
all
right
we're
just
about
to
go
through
the
issues
board
here?
Okay
and
then
you
know
we
can,
then
we
can
move
on.
Maybe
did
you
have
anything
to
present
today
or.
E
Oh
yeah
I'll
I'll
present
half
of
the
research
paper.
D
F
E
First,
you
can
go
on
and
I'll
start
in
there.
A
Yeah
yeah
and
so.
E
C
A
Task
board-
and
we
haven't
reviewed
this
in
a
couple
weeks,
but
I
think
probably
we
can
at
least
make
some
headway
on
it
here.
So
why
don't
we
look
at
what's
finished
here?
That's
always
a
good
thing
to
do
to
see
where
we
are
to
make
appreciate
where
we
are.
A
So
we
have
the
present
collective
pattern,
generators
idea,
so
that
actually
was
a
neural
match
presentation.
So
this
is
done,
but
we
still
want
to
do
things
to
follow
up
and
I'm
not
sure
what
those
things
are.
Yet
I
don't
know
if
I
think
jesse
saw
the
presentation.
I
don't
know
if
krishna.
C
A
Anyways
yeah
so
well.
We
want
to
follow
up
on
this,
but
I'm
not
really
sure
how
we
would
do
that.
Yet
we
got
to
make
some
issues
for
that,
and
then
we
of
course
have
this
this
an
associated
paper
with
the
basil
area
and
psychophysics,
and
all
that
coming
up
early
next
year.
I
have
we've
already
put
the
proposal
in
so
that'll.
Be
something
to
follow
up
on
hacktoberfest
is
that
kind
of?
A
I
guess
fizzled
out.
It
was
just
that
we
had
a
number
of
people
who
participated
early
and
then
we
left
we
didn't
get
any
more
people
participating,
but
I
think
we
got
about
maybe
10
people
to
participate,
which
is
actually
pretty
good.
No
one
won
any
of
the
things
that
were
offered.
A
So
you
know,
if
you
make
more
than
five
commits,
you
get
a
t-shirt,
but
you
know
it's
it's!
Okay,
it's
just
you
know.
I've
never
really
done
it
before.
We
tried
to
do
it
like
in
2019,
but
nothing
ever
happened.
So
we
didn't
get
any
contributors.
So
you
know,
so
that's
that's
something
that
we
can
revisit
next
october.
Maybe
I'd
do
it
again,
just
simply
because
it's
not
that
much!
A
You
know
you
might
have
to
do
a
little
bit
of
advertising
for
the
event
for
the
month,
but
you
can
get
potentially
get
people
to
come
into
the
platform
that
way
gsoc
final
presentations.
That's
really
old,
create
videos
of
diffusion,
limited
aggregation,
so
I've
actually
created
a
couple
examples
of
this
and
put
them
on
the
youtube
page.
If
you
go
to
the
youtube
page
and
look
just
running
some
examples
in
that
logo.
A
I
like
the
idea
of
making
little
tutorial
videos,
so
what
I
did
was.
I
just
ran
a
simulation
on
my
computer.
I
recorded
the
video
and
then
put
it
up
on
the
youtube
channel
as
a
sort
of
a
supplement
to
some
of
the
conversations
we've
had
in
the
group
here.
A
So
I
and
I
kind
of
like
the
idea
of
making
little
tutorials,
I
have
some
tutorials
of
other
things
up
there.
That
would
be
good
to
check
out
as
well
if
you're
interested
yeah.
This
is
the
narrow
match,
talk
again
that
goes
back
to
the
top
1657
model
of
motile
cells
in
the
embryo,
not
sure
what
that
was,
this
lecture
on
pca
unit,
umap
and
t-sne.
That
was
several
weeks
ago.
A
So,
if
you're,
you
know,
I
think
it's
like
three
or
four
weeks
ago,
if
you're
interested
in
that
you
can
go
check
that
video
out
or
if
you
want
to
follow
up
on
it.
That's
perfectly
fine.
Let's
see
revisions
to
the
bachelorette
morphology
paper,
that's
actually
done,
and
all
these
are
done.
So
I
think
we
have.
A
I
think
we
have
a
lot
of
yeah.
We've
done
a
lot
of
stuff.
So,
let's
see
in
progress,
number
44
krishna
paper
review,
so
that
was
the
thing
that
krishna
is
going
to
present
on
today.
A
So
that's
going
to
be,
I
guess
in
finished,
maybe
well.
We
still
have
to
review
the
paper,
though
so
it's
probably
not,
but
I'll
look
forward
to
that.
This
is
the
issue.
Actually
I
linked
it
to
hacktoberfest
tutorials
for
youtube.
So
that's
what
I
just
mentioned:
number
55..
So
if
you're
interested
in
making
tutorials
for
youtube
on
anything,
you
can
think
of
from
like
simulations
to
maybe
some
other
thing.
A
Bibliographies
is
like
putting
it
in
something
like
endnote
and
having
a
list
of
references,
but
actually
phones,
some
interesting
templates
for
and
just
as
an
aside
here
and
something
called
periodicals
for
annotated
bibliographies.
A
So
if
you've
never
heard
of
annotated,
bibliography
you'll
see
that
it's
it's
basically,
you
know
a
list
of
references
and
then
there's
like
a
description
underneath
the
reference
and
then
from
those
descriptions.
You
can
sort
them
out
so
like
this
one
is
the
life
of
behavior.
A
The
authors
are
here:
this
is
the
date
of
publication
doi
and
then
this
is
actually
when
it's
been
posted
october,
14th
2019.
in
this
paper,
the
authors
remind
us
that
neuroscience
are
objective
studies
living
stuff.
It's
they
go
on
about
the
three
essential
principles
of
behavior,
and
so
you
can
have
like
a
reference,
it's
identifier
and
then
this
blurb,
underneath
that
tells
you
what
it
is
and
then
you
can
have
you
know
you
can
link
those
different
entries
in
different
ways
based
on
the
description,
so
this
is
actually
paramecium
neuroscience.
A
So
this
is
familiar
because
we've
been
talking
about
bacillaria
and
neuroscience,
but
it
did
the
same
thing
here
where
they
have
these
descriptions
of
the
of
the
reference
you
know
and
kind
of
like
some
background,
and
so
that
might
be
an
interesting
thing
to
build
for
some
set
of
topics.
A
I
did
this
when
I
was
an
undergraduate,
I
built
a
an
annotated,
bibliography
and
they're
quite
useful.
I
think
that's
something
we
could
maybe
do
as
well.
So
that's
that's
part
of
this
bibliography
thing.
We
have
a
couple
issues
outstanding
on
that,
but
that's
something
we
can
update
later
create
an
outline
for
action
items
on
periodicity
paper.
So
last
week
I
think
susan
was
in
the
meeting.
A
We
talked
about
the
periodicity
paper,
which
I
don't
I
haven't
really
made-
that
much
progress
on
it,
but
it's
in
the
is
where
we
have
a
a
google
docs
draft
of
this,
and
maybe
next
week
I'll
go
over
it
a
little
bit
more
talking
about
like
where
we
are
I'll,
try
to
make
some
more
progress
on
it
and
try
to
figure
out
where
we
are
and
how
people
can
make
contributions
to
it.
I
know
we've
been
talking
about
it
earlier.
A
A
I
think
this
is
something
that
usual
was
working
on,
but
he's
been
not
feeling
well
the
last
couple
weeks,
so
he
hasn't
been
coming
around
the
meetings,
but
this
is
something
we
want
to
keep
on
the
sort
of
the
mid-range
burner,
because
we,
you
know
it's
an
interesting
area.
We
have
the
data
and
we're
it'll
be
a
nice
addition
to
what
we
have
in
the
group.
A
Let's
see,
oh
we've
got
a
new
member
here,
hello.
How
are
you.
E
G
E
Yeah
she's,
I
know
she's
for
my
college.
Okay,.
A
So
you're
from
krishna's
college
at
the
same
time,
okay,
well
welcome.
Thank.
A
F
Actually,
yeah,
that's
what
I
I'm
new
to
open
source,
so
I
don't
know
much
about
it.
I
was
very
interested
in
when
I
saw
krishna
status
and
I
asked
about
your
organization
and
I
thought
maybe
I
can
join
in
and
learn
something.
A
So
we
were
talking
about
this
thing
called
a
task
board,
and
so
this
is
where
we
put
a
lot
of
our
tasks
and
we
use
github
here
for
for
our
task
board.
We
generate
tasks
in
github
and
then
we
revisit
them.
We
put
labels
on
them
perhaps,
and
then
we
review
them
every
once
in
a
while
in
our
meetings,
maybe
every
couple
weeks
so
this
time
we're
going
through.
So
we
go
from
like
right
to
left.
A
We
go
through
finished
in
progress
and
then
to
do,
and
then
you
know
we,
it
kind
of
helps
us
think
about
like
what
we're
doing
in
our
meetings.
So
if
we
do
something
a
lot
of
things,
we
do
in
our
meetings,
they
just
kind
of
maybe
would
otherwise
fall
through
the
cracks
and
we're
trying
to
you
know
keep
on
top
of
a
lot
of
different
things.
So
that's
one
thing
about
like
an
open
source
organization
is
that
we
like
to
review
these
progress
or
task
boards,
and
so
this
is
something
that's
publicly
available.
A
So
people
can
view
it.
You
know
they
don't
have
to
register
for
anything
to
look
at
it,
but
there
are
a
bunch
of
tasks
on
here
that
anyone
can
participate
in.
A
I
don't
know
if
I
have
a
link
to
the
open
paper,
but
basically
you
know
people
can
contribute
to
these
issues
as
they
wish.
If
they
have
something
to
add-
and
that's
the
idea
here,
so
we
have
a
bunch
of
ongoing
projects.
One
of
those
is
complexity,
measures
number
three,
so
complexity
measures.
I
don't
think
we've
kept
up
on
that
too
much.
We've
talked
a
little
bit
about
different
things,
but
we
haven't
really
gotten
any
momentum.
E
A
That's
something
that
I
think
jesse
was
involved
with
thick
was
involved
with
so
we'll
we'll,
probably
update
on
that
soon.
Stop
sharing
a
stream
by
the
way.
Oh,
okay,
let
me
reshare
it.
Maybe
sometimes
it
doesn't
work
all
right.
A
Okay,
so
this
is,
I.
A
All
right
yeah,
so
the
complexity
measures
is
open.
We've
had
prior
conversations
about
different
things,
like
you
know,
different
network
measures
and
different.
We
have
this
collection
of
materials,
so
recapping
we've
talked
about
trees
and
metabolic
scaling
and
information
measures.
A
C
A
C
Return
to
that
once
at
more
free
space
to
do
it,
yeah.
A
A
Yeah
so,
and
then
this
bestselleria
non-rule
cognition
paper.
So
this
is
the
thing
that
is
linked
to
a
lot
of
these
issues,
with
the
presentation
at
neuromatch
and
so.
A
Oh
yeah,
you
can
join
it,
we're
still
working
on
it.
It's
it's
sort
of
in
the
state
where
I
think
we
have
the
good
a
good
main
idea,
but
we
need
to
work
on
some
of
the
fleshing
out
the
details
of
like
the
measures
and
and
how
they
apply
and-
and
it's
still
not
clear
whether
we
can
actually
do
a
formal
analysis,
but
we
can
at
least
make
propose
some
measures,
and
things
like
that,
so
I
think
we'll
probably
follow
up
with
what
was
in
the
presentation
for
at
neuromatch.
A
And
so
that's
that's
something
we
can
work
on
the
next
couple
weeks.
There's
a
deadline
on
that
of
early
next
year.
So
I
I
submitted
a
proposal
and
it
was
accepted.
So
I
I
think
it's
like
march
or
something
so.
We've
got
a
little
bit
of
time
on
that,
but
yeah
it's
it's
still
very
much
not
done.
A
And
then
we
have
some
more
issues
and
hold
on
axolotl
data
analysis,
so
this
is
susan's
data
set
that
she's
opened
up
for
us
and
we're
thinking
about
ways
to
analyze
it
and,
I
know
again,
usually
had
some
ideas,
but
he's
been
out
of
action
for
a
bit,
so
we
might
revisit
that
soon,
but
we've
still
got
that
on
on
top
of
top
of
our
minds,
so
krishna
proposed
last
summer,
a
general
biological
ml
model,
and
so
that's
also
something
that
is
one
of
those
big
ideas
that
you
know
we
we
have
to
like
kind
of
get
our
hands
around.
A
I
know
this
little
like
you
know.
This
is
a
little
bit
of
a
reach
for
like
a
quick
project,
but
have
you
thought
well
krishna
you've
got
that
paper
that
you're
working
on
so
maybe
that
is
tied
into
this
as
well
right,
yeah.
Okay,
then
there's
this
crate
embryo
model
in
blender
number
38..
So
that's
again,
something
I
think
lujan
was
working
on.
A
I
don't
know
what
the
status
of
that
is,
but
we'll
find
out.
Maybe
in
a
couple
week
next
couple
weeks
he
was
working
on
a
3d
model
of
embryos,
so,
like
blender,
is
a
3d
modeling
program
that
you
can
render
3d
objects.
So
it
was
taking
data
from
imaging
data
and
turning
it
in
you
know:
extruding
the
objects
and
turning
them
into
3d
objects
so
and
we're
still
trying
to
get
something
for
the
container
for
the
docker
container
for
openwork.
C
A
We'll
have
to
see
how
it
develops
over
time
to
see
if
that's
suitable
recruit
people
as
diva
learned
contributors,
so
that
was
tied
to
the
hacktoberfest
initiative,
and
I
don't
I
mean
I
don't
know
if
we
need
to
do
a
post-mortem
on
the
hacktoberfest
stuff
but,
like
I
said
we
had
about
five
or
maybe
about
10
people
participate.
A
Maybe
a
couple
people
participated
multiple
times,
but
not
enough
to
get
any
sort
of
prize
from
github.
So
you
know
the
idea
was.
If
you
did
five,
if
you
did
five
pull
requests,
you'd
get
a
t-shirt.
A
I
think
we
got
three
was
the
top
top
number
of
pull
requests,
and
so
you
know
that
people
didn't
get
t-shirts,
but
I
liked
the
interaction
on
diva
learn.
It
was
sort
of
all
bunched
up
at
the
beginning
of
the
month,
but
you
know
I
think
people
forget
about
oktoberfest
after
a
while,
I
didn't
really
advertise
it
after
maybe
about
the
sixth
of
the
month.
So
you
know
just
kind.
A
A
C
Repo
owner-
but
I
think
even
I
I
think
it's
that
you
can
do
a
bunch
of
more
requests,
but
they
don't
have
to
all.
A
C
The
same
organization,
so
maybe
they
did
get
t-shirts
just
not.
E
Yeah,
a
total
of
four
a
total
of
five
requests
are
necessary.
You
can
even
create
a
pull
request
on
your
own
repository.
You
just
have
to
label
it.
Give
it
a
list.
Oh.
A
A
A
Too,
like
I
don't
know,
if
anyone
committed
in
other
organizations
and
then
ended
up
getting
their
prize,
but
I
like
the
idea
of
hacktoberfest,
you
know
kind
of
being
invested
in
it
a
little
bit.
I
don't
know
what
we
could
give
on
our
end
in
terms
of
like
prizes.
A
You
know
that
would
just
stimulate
people
to
to
contribute
more,
like
you
know,
some
sort
of
like
recognition
or
I
don't
know
if
we
have
any
swag
that
we
could
give
them,
but
that
that's
also
something
we
could
do
outside
of
like
what
github's
doing
like
you
know,
if
you
make
maybe
like
two
commits
or
a
really
good
commit,
you
know
you
get
some
prize,
so
I
mean
that's.
A
I
think
I
did
recognize
everyone
who
made
a
commit
like
in
the
newsletter,
so
I
put
their
names
in
the
newsletter
and
but
yeah.
I
think
if
we
do
this
next
year,
it'll
be,
I
think
you
know
it'll
be
interesting
to
see
if
we
can
get
more
people
like
you
know
how
to
promote
it
now,
a
little
bit
so
yeah
we'll
look
into
that
next
year,
but
I
think
it
went
pretty
decently.
A
A
So
then,
the
there's,
this
embryo
visualization
task
number
six,
that's
for
the
open
arm,
docker!
So
that's
what
I
was
talking
about
when
I
talked
about
the
embryo
model
and
blender
that's
connected
so
open
worm
foundation.
Has
this
docker
container,
which
is.
C
C
A
A
There
are
all
these
different
applications
for
each
project
and
dvoram
does
not
have
an
application,
and
the
reason
for
that
is
because
we
really
haven't
been
focused
on,
like
you
know,
just
you
know
easy
to
execute
simulation.
We've
done
a
lot
of
data
analysis.
We've
done
a
lot
of
machine
learning,
we've
done
a
lot
of
theory
and
other
things,
but
not
really
anything
as
straightforward
as
what
they.
Some
of
the
other
projects
have
done.
A
So
we've
been
thinking
about
how
to
build
a
docker
showcase,
app
or
something,
and
so
this
embryo
model
might
be
one
example.
But
it's
got
to
be
something
lightweight.
A
You
know
it
can't
be
something
really
because
we
already
have
a
lot
of
apps
in
that
docker
container.
That
are,
you
know,
generate
a
lot
of
output,
data
or
high
memory
applications,
so
we're
still
searching
for
that
thing
to
put
in
there,
but
I
thought
I'd
bring
that
because
we
keep
mentioning
it
and
there's
a
lot
of
interest,
but
then
it's
like
it's
hard
to
really
know
what
to
actually
put
in
there
and
that's
no
one's
fault.
It's
just
kind
of
it's
a
hard
problem.
A
So
we
have
a
couple
of
these
open
papers
which
are
basically
papers
that
we
have
that
anyone
can
contribute
to.
I
think
I
need
to
make
that
more
prominent
on
the
website,
because
I've
talked
about
this
in
meetings
and
you
know
probably
isn't
clear
to
people
where
to
go
for
that.
I
don't
know
if
I
have
a
link
to
it
in
this
board,
but
let
me
see
yeah,
I
don't
think
so.
Well,
anyways
we
have
a
spot
on
the
github
where
you
can
put.
I
think.
A
Actually
it's
in
group
meetings,
I'm
not
mistaken,
maybe
not
well
anyways!
Well,
let's
see
yeah
group
papers,
so
we
have
these
papers
here
in
the
readme
of
group
meetings
and
they're
a
couple
of
open
papers.
So
some
of
these
one
of
the
ones
I
was
talking
about
was
this
bacillary
non-cognition
sub-project
and
that
has
two
open
papers,
and
so
they
can
either
be
google
docs
or
a
github
repo.
So
this
this
paper
say
like
emergent,
behavior
in
c
elegans
and
other
organisms.
A
A
You
can
just
edit
it
in
the
document,
so
here,
if
you
wanted
to
edit
this
document,
you'd
hit
this
pen
and
you
it
would
you'd
have
to
create
a
it'll,
create
a
github
will
create
a
fork
for
you
and
then
you
would
just
issue
a
pull
request
on
that
fork
and
then
push
it
up
to
the
main
directory,
and
so
this
is
this
open
papers
idea.
A
A
Than
get
out,
but
so
that's
that's.
If
people
are
interested
in
participating
in
open
papers,
that's
a
good
way
to
go
so
then,
finally,
we're.
I
think
we
have,
let's
see
any
more
issues.
Oh
the
create
a
theory
layer
for
diva
learn
so
the
diva
learn
organization.
A
I've
created
a
couple
new
repo
or
a
couple
new
repositories
there.
One
of
them
is
called
theory
building.
So
if
we
have
this
is
the
idea
of
creating
a
theory
layer
for
diva
learn.
So
we
have
the
machine
learning
resources.
We
have
the
data
science
resources
and
we
also
have
this
theory
building
layer
which
is
sort
of
very
in
very
early
days.
I
need
to
sketch
it
out
more
and
and
put
some
conditions
on
what
we're
talking
about
there,
so
that
people
don't
come
in
and
they're
just
kind
of
lost.
A
There's
some,
maybe
we'll
take
some
of
the
ideas
that
we've
been
working
on
some
of
these
other
projects
and
make
them
sort
of
into
principles
for
contributing.
So
you
know
you
know
we
have
these
data,
this
data
analysis
that
we've
done
on
a
lot
of
these
things.
A
A
A
A
We
also
have
the
community
board,
which
is
basically
the
project
board
for
divalern,
and
then
we
have
finally,
this
theory
building
repository,
which
is
basically
the
theory
behind
a
lot
of
the
things
that
we're
analyzing,
and
so
that's
the
way
we're
going
to
have
this
under
education.
I
wanted
to
bring
this
up.
I've
been
working
on
this
journal
of
open
source
science
or
open
source,
whatever
it
is
so
anyways.
This
is
a
publication
that
builds
upon
last
summer's
gsoc
projects,
so
I've
gotten
it
actually
into
the
form
that
they
wanted.
A
I
think
pretty
close,
I
think,
is
myoku
created
this
little
template,
and
so
that's
what
this
is
here.
It's
a
little
table
that
has
like
the
title
and
the
authors,
so
it's
my
oak
usual
myself
and
the
summary
of
this.
So
if
you're
interested
in
looking
at
this
and
reviewing,
maybe
you
can
provide
some
comments.
If
you
wish
there's
our
chat
window
here,
it
is
so.
This
is
a
very
short
main.
A
We
have
this
organization,
we
have
this
need
to
to
take
data
and
quantify
it
and
explore
it
a
bit
with
you
know,
machine
learning,
tools,
deep
learning
tools,
and
then
we
have
this.
This
is
our.
These
are
some
schematics
showing
the
divolent
platform
the
deform
program
and
then,
where
you
know
how
to
execute
this,
and
then
we
have
some
references.
A
So
this
is
I'm
going
to
add
a
couple
more
things
here.
I
need
to
add
a
section
on
well.
I
think
I
already
did
actually
add
this
section
of
meta
features,
but
there
are
a
couple
more
things
I
need
to
add
to
it,
but
we're
making
progress
on
this,
and
I
thought
I'd
bring
this
up
because
I've
been
putting
this
off
for
a
long
time
so
yeah.
A
So
that's
all
I
have
to
talk
about
about
the
world
of
our
github
and
updates.
Now
krishna,
you
had
a
presentation
you
wanted
to
make
on
your.
A
E
E
So
it's
on
reinforcement
learning.
If
I
have
to
explain
it
in
simple
words,
you
know
it
generally
works
on
reward
policy
rewarding
punishments.
E
So
here
the
the
main
focus
of
the
agent
is
to
maximize
its
points,
for
example,
to
a
person
is
playing
chess,
and
if
we
have
that,
if
you
or
you
know
hit
the
pawn
you'll
get
that
much
of
points,
if
you
hit
the
rook,
you
will
get
that
much.
If
you
go
for
the
night,
you
will
get
that
much.
So,
at
the
end,
the
person
has
to
maximize
its
points.
So
same
goes
with
the
reinforcement
learning,
so
you
can
in
simple
words
you
can
say
it's
a
method
for
a
markov
decision.
E
Process
yeah,
for
example,
if
we
have
to
do
a
certain
number
of
tasks
and
we
have
like
one
two,
three,
four
five
six
here:
it's
it's
a
micro
markov
reward
process
or
a
robot.
So
if
it
does
a
task
fight
it,
it
gets
a
reward
of
minus
one,
but
when
it
shut
down
it
doesn't
get
any
reward,
whereas
if
it
does
task
four,
it
gets
a
positive
reward.
So
you
can
visualize
it
that
task.
One
two
and
three
are
crashing
to
a
wall
and
task.
E
Four
is
going
successful
to
its.
You
know,
end
point
and
the
refuel
is
you
know
just
the
way
when
it
you
know,
fills
it
if
visit
petrol,
so,
for
example,
if
yeah,
if
we
have
that
red
dot,
if
starting
point
and
we
have
to
reach
to
the
star
and
all
these
purple
blocks-
are
having
minus
one
reward.
So
what
we
have
to
do
that
we
have
to
reach
to
the
star
with
the
shortest
distance
possible.
E
So
here
we
can
have
a
number
of
ways
to
go.
We
can
follow
this
thing.
We
can
follow
this
route
or
we
can
follow
this
route.
So
after
all
the
routes,
our
agent
learns
that
whenever
it
goes
through
the
hills,
the
rewarded
skating
is
the
least
one.
So
it
try
it
to.
You
know
maximize
its
chance,
not
to
visit
the
hill.
So
there
are
two
types
of
reinforcement,
learning
thing
one
is
on
policy
and
one
is
off
policy.
E
If
you
are
going
on
off
policy,
if
we
have
three
groups
route
number
one,
two
and
three-
and
when
we
go
on
the
route
three,
we
get
a
negative
reward.
So
we
come
to
know
that
our
way
is
on
root
number
one
or
two,
and
we,
when
we
go
on
route
number
two,
we
again
get
a
negative
reward.
So
we
come
to
know
that
our
root
is
one.
E
So
here
we
are,
you
can
say
we
are
excluding
the
choices
and
you
know
we
are
deciding
the
root
we
have
to
take
by
the
other
roots,
but
in
case
of
on
policy
that
is
sarsa
policy.
We,
you
know,
go
to
route
number
one
and
then
we
see
is
it
feasible?
E
So
so
this
much
work
has
done
it's
regarding
you
know
I've
done
about
the
abstract.
I
have
explained
what
a
markov
reward
process
is
I've
given
a
few
examples,
and
you
know
it's
rough
about
five
pages
and
I
have
to
add,
you
can
say
bibliography
to
add
references
and
you
know
done
a
comparison.
That's
not
completed
yet.
So
this
is
the
first
paper.
A
Do
you
have
anyone?
Anyone
have
any
comments
before
I.
A
A
E
Yeah,
it
would
be
much
more
of
a
literature
survey,
literary
survey
it
wouldn't.
I
can
do
the
benchmark
scores
also,
but
I'm
I'm
thinking
to
present
it
into
you
know
much
more
of
a
theoretical
way.
If
I
go
into
the
benchmark
also,
it
would
be,
you
know,
become
a
lengthy
one,
so
it
would
take
even
more
times
it's
being
incomplete
from
you
can
say
past
one
month
and
if
I
have
to
add
the
benchmarks,
it
will
take
another
one
and
a
half
month.
E
A
Yeah,
this
idea
of
of
policy
and
on
policy.
So
this
is
like
this
almost
sounds
like
off.
Policy
is
like
exploration,
yeah,.
E
So
we
also
have
a
thing:
it
is
versus
exploitation,
for
example,
if
we
are
having
10
roots
and
from
you
know,
first
three
roots,
I
come
to
know
that
the
first
root
is
the
better
one.
So
there
is
a
dilemma
of
you
know
exploiting
the
other
routes
or
you
know
you
can
say
going
with
your
existing
knowledge.
E
So
if
I'm
able
to
explain
it
hello,
yeah
yeah,
it's
you
know
between
exploitation
and
exploration
that
either
I
want
to.
You
know,
exploit
all
the
possible
routes
or
with
my
existing
knowledge
or
with
my
exp,
express
existing
experience.
I
want
to
stick
to
that
and
make
it
make
the
reward
as
much
as
I
can,
because
you
know
if
I
go
on
exploring
there
are
chances
of
getting
even
a
better
path
or
even
getting
a
worse
one.
E
Because
it's
a
dilemma,
kind
of
thing
that
you
want
to
you
know
get
out
of
your
comfort
zone
or
sometimes
getting
out
of
comfort
zone.
Can
you
know
not
be
much
useful
so
here
I'll
explain
that
in
the
latest
section
of
the
paper
and
now
I
have
to
present
the
second
thing
that
general
biological
cnns?
Okay,
I
just
got.
E
B
B
E
E
Is
a
custom-based
convolution
neural
network
for
biological
data,
as
as
we
people
know
that
most
of
the
cnns
that
we
are
using
are
trained
on
imagenet
dataset,
it's
a
data
set
that
comprises
of
vehicles,
traffic
lights,
cup
cup,
plates
human
beings,
cats
and
dogs.
So
all
the
you
know,
state
of
the
art
cnns,
for
example
the
resnet,
the
lx
net.
All
of
them
are
trained.,
but
then
they
are
retrained
on
the
area
of
interest
data
set,
for
example.
E
If
I
want
to
classify
benign
and
malignant
cancer
first
I'll
use
my
resume.
That
has
been
pre-trained
on
the
image
net,
that
data
set
that
contains
cat
dogs
cup
plates
human
beings
and
vehicles
and
then.
E
It
on
my
own
data,
but
I
guess
this
thing
can
be
improved
a
little
bit.
It's
a
tedious
task.
You
know
you
can
say
a
big
project,
but
if
we
take
all
the
biological
data
and
we
brain
pre-trained
our
neck
on
the
biological
data,
instead
of
that
card,
dog,
animal,
human
beings
and
vehicles-
and
then
we
retrained
it
whenever
a
person
wants
it
to
do.
For
example,
when
a
person
can
use
resnet
any
in
any
way,
but
it
is
still
pre-trained
on
a
broader,
broader
spectrum.
E
Impulse
importance
that
that
is
just
the
learning
white
weights
and
devices,
and
so
why
retrain
data?
Because
if
you
want
to
obtain
something
from
scratch,
it
would
take
a
lot
of
time
and
it
would
consume
a
lot
of
computation
power.
It
would
not
be
feasible
with
time
and
with
economy.
So
what
we
do
instead
of
that
learnable
weights,
that
I
discussed
in
the
previous
slides
that
they
are
not
initialized
from
zero.
They
have
some
random
value
that
was
learned
during
its
initial
training
on
adobe
cards,
human
and
vehicles.
E
So
once
if
we
are
not
starting
it
from
zero,
they
tend
to,
you
know,
perform
better
and
they
tend
to.
You
know,
learn
fast.
For
example,
if
a
person
knows
how
to
you
know,
drive
a
car,
so
he'll
be
able
to
drive
a
bus
or
some
big
bikers
in
a
better
way.
But
if
we
are
teaching
a
person
to
you
know,
drive
a
16,
tired
bus
or
a
big
cherry
truck,
so
he's
more
prone
to
crash.
E
So
here
what
we
are
doing,
we
are,
you
know,
not
initializing
the
weight
from
zeros.
Our
model
knows
something
about
images,
but
not
particularly
the
type
of
image
that
we
are
going
to
feel,
but
if
we
narrow
this
approach
just
for
biological
data-
and
we
have
plenty
of
it,
for
example,
if
yeah,
for
example,
if
we
want
to
you,
know,
predict
cancer
and
our
training
data
set,
had
cancer
images,
plant
images,
zygote
images,
microbial
images,
the
probability
or
you
can
say
the
the
like-
the
what
you
can
say
there.
E
All
these
type
of
images
with
will
share
a
common
thing.
To
some
extent,
a
cancer
cell
would
be
distinct
to
a
vehicle,
then
to
a
plant
cell.
So
even
though
they
are
wearing
a
plant
cell
and
animal
cell
varies
a
lot,
but
at
least
they
don't
vary
as
much
as
a
car
and
then
a
cell.
E
So
what
are
the
benefits
of
a
pre-trained
model?
So
the
it's?
You
know
very
simple:
to
incorporate
once
we
code
it,
and
you
know
we
put
all
the
weights
and
biases
and
you
know
create
a
model.
Then
any
person
with
not
even
a
much
knowledge
of
deep
learning
just
has
to
import
the
model.
You
can
say,
give
his
training
cut
and
get
the
output,
so
even
a
person
with
little
bit
computation
knowledge
can
perform
it.
E
Other
second
thing
is
that
there
is,
as
mentioned
earlier,
the
results
are,
you
know
they
they
are
quick
to
get
and
since
there
is
not
much
label
data.
E
So
if
I
have
pre-trained
my
model,
1000
images,
then
it
it
can,
it
would
be
able
to
perform
even
on
100
to
150
images
when
retrained,
and
the
second
thing
is
versatility,
for
example,
if
I
am
taking
the
broader
sense
of
biological
data
that
includes
plant
animal
and
all
all
the
things
from
microbes
to
you
know
a
cancer
cell,
an
embryo
to
everything
possible,
whatever
level
data
that
we
can
find.
E
So
we
will
have
a
lot
of
versatility
that
even
a
water,
any
person
with
plants
can
use
it,
and
you
can
say,
for
example,
in
a
person
working
on
and
when
that
you
can
also
use
it.
E
So
we
will
have
a
lot
of
versatility,
even
though
that
versatility
is
less
as
compared
to
a
normal,
resonant
yeah,
but
if,
but
it
would
be,
you
can
say
focused
on
biological
data
and
it
would
be
versatile
on
every
biology
domain,
so
yeah,
first
of
all,
it
would
be
also
have
a
few
of
hurdles.
First
is
computation
power.
If
we
want
to
train
thousands
of
images,
we
have
to
you
know
stream
them
from
scratch.
E
First,
we
have
to
collect
all
of
the
biological
data
that
we
can
publicly
get,
and
since
biological
data,
especially
concerning
with
human
beings,
is
difficult
to
get
because
of
the
privacy
concerns,
but
we
do
have
a
few
things
like
you
know,
radical
logic,
association
of
northern
america.
E
It
would
take
a
few
weeks-
or
you
can
say
one
two
months
but
yeah
we
can
do
it.
We
have
a
lot
of
things
like
kaggle
and
google
open
data
sets,
but
yeah
we
have
to
collect
it.
Then
we
would
be
needing
gpus
for
training
it.
That
would
be
also,
you
can
say
it
is
difficult.
As
a
free
google,
collaboratory,
google
crap
that
we
get
has
you
know
limitations,
it
can
be
only
done
for
a
certain
hours
in
a
month
and
its
gpus
are
like
kt.
E
E
So
that
would
be
another
limitation,
but
once
we
have
successfully
applied
the
pre-trained
model,
I
guess
we
would
be
first
one
first
open
source
organization
that
would
have
you
know
just
focus
the
biological
data
in
case
of
cnns.
We
do
have
unit.
That
is
a
segmentation
type
of
model.
E
That
is
using
biomedical
analysis,
but
we
don't
have
a
journal
biological
model
that
would
be
used
in
biomedical
to
and
normal
animal
to
a
plant
data,
and
that
is
what
I
earlier
explained
that
image
data
exclude
these
type
of
things
and
our
biological
data
is
this.
So
I
guess
there
is
not
even
0.1
percent
of
similarity
between
them
and
that's
just
the
thing
that
I
earlier
mentioned.
A
Yeah,
I
think
it's
good
to
have
a
refresher,
I'm
glad
you've
been
thinking
about
it.
Does
anyone
have
any
comments
or
questions
about
it?.
C
E
A
Okay,
shirdi
or
susan,
do
you
have
anything
else
to
add
any
questions
or
comments.
G
There
should
be
a
way
of
getting
some
medical
data
from
data
sets
like.
Sometimes
you
can
get
open
source
ones,
I'm
not
sure
about
it.
But
I
know
there
are
medical
data.
E
E
Plant
related
data,
but
that
would
be
also
you
know
in
abundance.
If
you
are,
if
we
try
hard
and
spend
some
time
on
collecting
it,
and
we
would
be-
I
guess
the
first
one,
I'm
not
sure
that
someone
is
is
working
on
specialized
bio
models.
People
are
working
on
biomedical
models,
for
you
know
disease
prediction,
but
on
in
general,
no
one
is
working
on
bio
data.
As
far
as
my
knowledge
is
concerned,.
G
I'm
hoping
to
generate
some
more
data,
more
biological
data
using
optical
coherence,
tomography
and
apparently
you
can
use
it
for
histology
as
well.
So
there
might
be.
G
Be
another
type
of
data
set.
E
Yeah
that
that
would
be
great-
and
that
would
be
you
know,
rare
form
of
data
that
is,
you
know,
can
really
give
a
versatility
to
a
model.
E
A
Yeah
yeah
yeah.
I
think
that's
great.
I
like
the
focus
on
pre-trained
models,
so
you
know
this
last
summer
we
did
a
pre-trained
model,
but
I
think
it's
a
very
rich
field,
I'm
not
sure
you
know,
I
think
it's
absolutely
necessary
for
like,
if
you're
dealing
with
a
lot
of
the
biological
phenomena
that
we're
interested
in
this
group.
A
So
you
know
you
can
do
I've
seen
actually
people
built
built,
pre-trained
models
for
different
features
of
a
single
cell,
so
there's
some
pre-trained
models
that
exist
for
like
segmenting,
like
mitochondria
out
of
cells
and
things
like
that
or
different
organelles,
and
then
they're,
of
course,
pre-trained
models
for
cells.
A
But
then
you
have
like
broader
features,
which
we've
kind
of
touched
on
in
devo
the
diva
learn
platform.
Where
you
know
you
have
like
motion
cues
and
things
like
that
things
going
on
in
the
embryo
that
you
know
exist.
You
know
they're
sort
of
at
the
sort
of
multi-cellular
level
so
and
especially
like
something
like.
I
know,
you've
seen
basilary
in
this
group,
but
more
and
more
generally,
with
a
lot
of
multicellular
colonies.
A
You
see
like
different
features
that
are
either
emergent
or
they're.
Sort
of
you
know
collective
behaviors
that
you
want
to
capture,
and
so
those
are
you
need
to
have
like
a
pre-trained
model
for
that,
because
it's
hard
to
label
those
events
they're
just
things
that
kind
of
emerge
out
of
the
soup.
I
know
we've
talked
about.
You
know
like
chemical
reactions
in
this
group
before
as
an
example.
A
Behavior,
well,
those
sorts
of
things
are
actually
a
pretty
good
example
of
something
that
you
can't
really
label
very
easily,
but
you
can't
you
nevertheless
are
interesting
features
and
so
and
but
you
have
cells,
you
want
to
know
the
features
that
are
single
cells,
but
you
also
want
to
know
the
feature
which
is
the
collective
behavior
of
the
cells
and
that's
where
it
gets
tricky
because
you
could
probably
do
like
you
know.
You
know
you
could
probably
pick
up
features
that
represent
sort
of
collective
states
of
matter.
A
You
know
like
some
reaction,
where
you
have
mixing
or
something
like
that,
but
you
know
to
have
that
sort
of
two-level
two-layered
thing
where
you're
looking
at
cells
and
then
you're
looking
at
this
collective
thing
that
they
do.
A
A
So,
like
you
know,
are
there
features
that,
are
you
know,
common
to
all
different
types
of
cells?
Are
there
features
that
are
kind
of
like
common
to
like
what
cells
do
in
in?
You
know
collectively
moving
around
or
configuring
themselves,
because
then
you
know
you
have
like
things
that
you
see
either
medical
images
or
in
like
yeah
yeah.
E
Actually,
our
group
is
having
people
with
different
domain
experts,
so
we
can,
you
know,
leverage
that
we
have
people
with
biology
and
computation
and
it's
a
mixture
of
both.
So
we
can,
you
know,
learn
from
each
other's
domain
expert
and
you
can
say
that
make
it
state
of
the
art
you
have
that
leverage
of
that
thing.
A
Okay,
now
I
know
it's
the
top
of
the
hour,
but
I
did
want
to
get
to
the
papers
that
some
of
the
papers
that
I
had
I'll
probably
be
about
15
to
20
minutes
on
this.
So
if
you
have
to
leave,
that's
fine,
but
I'm
gonna
go
through
these
and
this
will
be
on
youtube
later.
So
thank
you
krishna
for
your
presentation
and
let
me
was
that
nothing
bad!
Oh,
okay!
That's
fine!
A
So
let's
see
I'm
gonna
present
here
my
entire
screen
so
yeah.
I
just
wanted
to
go
through
some
papers
that
I
find
interesting
here.
I
talked
about
these
two
earlier.
These
are
annotated,
bibliographies
and
if
you
missed
this
part
of
the
conversation,
it's
probably
about
10
minutes
in
you
can
watch
it
on
youtube
later.
A
But
it's
basically,
this
idea
of
building
these
annotated,
bibliographies,
which
have
like
each
entry,
is
a
paper
and
then
you
have
a
short
description
underneath
and
then
it
could
be
like
the
author
or
the
reader's
reflections,
or
it
could
be
like
just
what
it's
about,
but
then
you
can
actually
link
these
together
in
a
search
or
some
other
way,
and
so
that's
potentially
useful
for
the
group,
but
that
that's
these
two.
But
but
I
want
to
talk
about
some
of
the
other
things
that
I
found
in
the
literature
here.
A
So
this
is
a
paper
that
this
is
something
actually
that
is
relevant
to
both.
So
I
have
another
group
on
saturdays.
We
talk
about
like
a
lot
of
things
on
ai
and
developmental
ai,
and
things
like
that,
and
so
we've
been
having
these
conferences.
We
actually
did
a
presentation
at
neuromatch
on
developmental
ai,
and
I
found
this
paper
yesterday
and
so
it's
it's
more
of
a
neurobiological
paper
developmental
neurobiology,
but
I
think
it's
really
interesting
and
I'm
hoping
that
we
can
maybe
incorporate
it
into
that
work.
A
So
this
is
called
fate
and
freedom
and
developing
neocortical
circuits.
So
it's
not
like
c
elegans,
where
you
have
a
set
of
cells
that
are
connected
through
a
connectome.
This
is
these
are,
like
you
know,
human
brains
or
mammalian
brains
where
you
have
a
neocortex
and
then
that
neocortex
is
densely
connected
and
there's
you
know
so,
but
but
base
the
basic
idea
well,
we'll
go
through
the
abstract
and
I'll.
Give
you
the
basic
idea
here.
A
So
the
activity
of
neuronal
circuits
of
the
neocortex,
which
is
our
the
front
of
our
human
brain,
which
is
the
the
enlarged
area,
and
you
all
know
what
that
is.
But
I
wanted
and
of
course
this
exists
in
all
mammals,
underlies
our
ability
to
perceive
the
world
and
interact
with
our
environment
during
development.
These
circuits
emerge
from
dynamic
interactions
between
cell
intrinsic
genetically
determined
programs
and
input
activity,
dependent
signals
which
together,
shape
these
circuits
into
adulthood.
A
Now
this
sentence
here
is
interesting
because
I've
been
thinking
about
this
in
terms
of
like
you
know,
when
we
talk
about
like
we
experience
things
in
the
world
versus
like
the
sort
of
the
genetic
underpinnings
of
neurons
and
how
they
develop,
and
so
is
required
this.
I
guess,
if
you're
going
to
model
this,
this
actually
requires
two
things.
A
One
is
that
you
have
a
neuron
or
something
that
represents
a
neuron,
but
also
something
that
represents
the
genotype,
and
then
you
have
some
way
of
characterizing
environmental
information,
and
so
my
other
group
were
talking
about
these
issues
and
they're,
not
quite
gelled
yet.
But
I
think
this
is
a
good
sort
of
framework
for
thinking
about
this,
and
I'm
going
to
propose
this
in
the
in
the
saturday
group
next
week,
a
little
bit
more
formally,
but
I'm
just
previewing
it
here.
A
So
these
local
circuits
are
being
integrated
into
networks.
Cells
are
differentiating
and
he's
talking
here
about
the
you
know,
sort
of
the
way
we
can
measure
this
and
says
an
interesting
paper.
They
kind
of
go
through
this
idea
of
cortical
columns
which,
if
you're
familiar
with
mammalian
neocortex.
A
This
is
the
way
these
cells
are
organized
into
these
layers,
and
so
it's
a
very
interesting
paper
for
a
lot
of
reasons.
Kind
of
walks
through
a
lot
of
cortical
information,
flow,
direct
and
indirect
neurogenetic
pathways
during
corticogenesis,
so
cortex
has
its
own
method
of
of
neurogenesis,
so
cells
are
born
and
they
move
through
different
layers
of
the
cortex.
A
A
The
same
author,
with
along
with
some
co-authors,
also
wrote
this
paper
cross-modal
genetic
framework
for
the
development
plasticity
of
sensory
pathways,
and
so
this
is
very
similar.
It's
a
little
bit
more
like
data
driven,
I
guess
so.
This
is
again
the
neocortex
in
mammals
for
each
sensory
modality
input
follows
a
corticothalon
thalamocortical
loop,
in
which
a
first
order,
exteroceptive
thalamic
thalamic
nucleus,
sends
peripheral
input
to
the
primary
sensory
cortex,
which
projects
back
to
a
higher
order.
Thalamic
nucleus
that
targets
a
secondary
sensory
cortex.
A
Basically,
what
he's
talking
about
is
that
you
have
this
structure
called
the
thalamus
that
sends
projections
up
to
the
cortex,
the
thalamus,
taking
in
connections
from
other
parts
of
the
body,
and
then
it
sort
of
loops
between
the
thalamus
and
the
cortex
multiple
times
it's
sort
of
as
a
sampling
of
information,
and
so
this
is
a
this
is
another
interesting
paper.
It's
pretty
dense
and
technical,
but
I
think
if
you
are
interested
in
that
topic,
it
kind
of
opens
your
eyes
to
some
things.
A
Another
paper-
and
this
is
quite
a
bit
off
in
the
other
direction-
is
called
how
do
living
systems
create
meaning.
So
this
is
by
chris
fields.
Who's.
An
independent
researcher
does
a
lot
of
really
has
published
a
lot
of
interesting
papers
on
like
philosophy
of
science
and
other
things,
and
then
michael
levin
is
a
developmental
biologist,
computational
biologist,
and
so
this
is
quite
a
bit
far
away
from
what
I
was
just
talking
about.
A
But
it's
interesting
because
we're
also
talking
about
in
this
group
things
like
basil
area.
A
So
the
abstract
reads:
meaning
has
traditionally
been
regarded
as
a
problem
for
philosophers
and
psychologists
advances
in
cognitive
science,
since
the
early
1960s,
however,
broadened
discussions
of
meaning
or
more
technically
the
semantics,
semantics
of
perceptions,
representations
and
actions
into
biology
and
computer
science.
You
review
the
notion
of
meaning,
as
it
applies
to
living
systems,
and
argue
that
the
question
of
how
living
systems
create
meaning
unifies
the
biological
and
cognitive
sciences
across
both
organizational
temporal
scales,
so
they
kind
of
get
into
the
idea
of
meaning
in
biology,
as
opposed
to
like
in
brains.
A
It,
like
you,
know
what
we
know
that,
like
brains,
generate
agency,
and
so
people
propose
well
cells.
Do
this
too.
They
just
don't
have
brains,
and
so
then
the
question
is,
you
know:
are
we
making
like
sort
of?
Are
we
projecting
ourselves
onto
cells
or
are
we?
Is
there
actually
a
mechanism
for
this?
A
And
so,
when
you
ask
this
question,
this
brings
up
a
bunch
of
things
that
are
sort
of
related,
and
one
of
these
is
this
idea
of
meaning
any,
and
this
is
related
to
agency
in
terms
of,
like
you
know,
is
a
system
that
doesn't
have
a
brain.
Can
they
generate?
This
kind
of
you
know,
meaning
of
things
this
sort
of
semantic
layer.
So
that's
the
idea
here.
A
So,
in
contrast
to
this
focus
on
individuals
and
learning,
we
advance
in
this
paper
a
deeply
evolutionary
approach
to
meanings
and
suggests
consonant
with
the
theme
of
this
special
issue.
There's
a
special
issue
involved
that
the
construction
meetings
present
a
common
and
pressing
question
for
all
disciplines
in
the
life
sciences.
A
That
doesn't
necessarily
mean
that
doesn't
necessarily
mean
that
there's
there
has
to
be
a
meaning
system
or
a
semantic
system.
It's
just
one
of
the
things
that
organisms
do
and
maybe
even
single
cells.
Do
it's
like
a
signal
detection
thing
where
you're
looking
at
signal
versus
noise?
It's
like!
Are
you
looking
at
some
object
or
are
you
you
know?
Is
there
something
you
can
identify
like
a
chemical
signal
or
some
object
like
another
cell?
Or
is
it
just
background?
Is
it
you
know
something
that
you
don't
have
to
pay
attention
to
in
humans?
A
A
So
but
that's
one
question,
but
the
second
one,
then,
is
how
do
living
systems
switch?
Their
attentional
focus
from
one
object
to
another:
well
now
they're
using
the
language
of
attention.
So
this
is
what
they're
again.
This
is
signal.
This
is
signal
versus
noise,
you're,
rejecting
the
noise
and
accepting
the
signal.
A
Now
you
so
that's
sick,
you're,
you're.
You
assume
that
that
attentional
drive
is
focused
towards
a
signal,
hopefully,
and
so
now,
how
do
you
switch
your
attentional
focus
to
different
signals
and
then
how
to?
The
third
question
is:
how
do
living
systems
create
and
maintain
memories
of
past
events,
including
past
perceptions
and
actions?
Now
they're
proposing
a
memory
mechanism,
and
in
that
case
we
know
that
single
cells
actually
do.
C
A
Memory
mechanisms
like
epigenetic
mechanisms
or
in
one
of
the
things
that
we
proposed
in
the
bacillaria
presentation,
they're
different
extracellular
mechanisms
like
polymers
that
can
be
used
as
markers.
A
Devices
outside
of
the
cell,
so
that
the
cell,
you
know,
has
like
some
something
that
it
can
draw
upon
to
guide
its
behavior.
So
you
know
you
might
have
pheromone
trails
or
you
might
have
chemical
gradients
that
you
follow.
Those
are
all
memory
systems
in
a
sense,
then
these
questions
all
presuppose
an
answer
to
a
fourth
question:
how
do
living
systems
reference
their
perceptions,
actions
and
memories
to
themselves?
C
A
But
they
are
like
their
actual
mechanisms
in
the
cellular
world,
but
then
how
do
these
things
get
referenced
by
the
cell
like?
If,
if
there
are
things
out
in
the
environment
that
exists
like
a
chemical
trail
and
the
organism
or
cell
follows
a
chemical
trail,
then
how
does
it
reference
it
like?
You
know,
how
does
it
know
where
it
is
in
the
chemical
trail?
How
does
it
know
when
to
stop?
A
Is
it
just
a
bunch
of
signals
that
it's
just
kind
of
deterministically
you
know
responding
to,
or
is
there
some
sort
of
system
of
sort
of
meaning
and
and
reference
subreferencing
that
goes
on
in
the
or
in
the
cell
or
the
organism?
A
So
they
say
we
use
the
term
living
system
instead
of
organism
in
formulating
these
questions
to
emphasize
their
generality.
So
again,
if
we're
dealing.
C
A
Anything
from
like
complex
organisms
to
single
cells,
we
have
to
be
flexible,
and
so
this
this
paper
kind
of
goes
through
this
and
they
kind
of
consider
maybe
a
meaning
system
within
a
single
cell
or
within
a
living
system,
so
they're
very
general
about
it
and
so
they're,
making
this
link
between
like
cognition,
that
you
might
see
in
humans
or
other
non-human
animals,
and
then
single
cells
or
plants,
or
something
else
that
you
know
you
wouldn't
really
associate
with
normally
associate
with
cognition,
and
so
they
actually
get
into
some
good
examples
here,
meaning
for
assuring
e
coli,
which
is
e
coli,
which
is
a
a
bacteria.
A
It's
a
chemotaxis
is
the
mechanism
that
they
look
at
and
and
maybe
potentially
gives
meaning
to
things.
So,
like
I
talked
about
a
chemical
gradient,
that's
basically
what
this
is.
The
cell
will
be
attracted
to
certain
chemicals
or
certain
parts
of
a
chemical
gradient,
and
so
that's
they
use
to
build
meaning.
A
But
they
basically
talk
about
these.
It
could
be
reference
streams
that
you
know
they
talk
about
these
different
things
and
how
they
require
energy
and
other
space
state-space
attractors
and
get
into
some
daisy
and
expectation
modeling,
so
a
bayesian
model
only
meaningful
differences
are
detectable.
So
basically,
I
mean
this
kind
of
reminds
me
this
model
of,
like
you
know,
a
lot
of
just
basic
signal.
Detection
mechanisms
so,
like
basically,
cells,
are
detecting
signals
in
the
environment
and
they're
kind
of
using
those
as
a
basis
for
like
a
little
bit
more
structure.
A
C
Yeah
I'd
like
to
I
don't:
are
you
just
gonna
go
over
the
first
couple,
the
first
one
next.
A
A
And
we
can
follow
up
on
them.
You
know
whenever,
but
I
will
put
the
link
to
the
papers
in
the
in
the
chat.
So
people
want
to
check
these
out
further.
As
usual.
A
A
So
the
reason
I
bring
this
up
is
because
the
talk
about
gastroids,
which
are
actually
a
system
that
we
should
maybe
present
on
in
the
group
in
another
another
week
where
they
talk
about
these,
these
different
little
structures
called
gastruloids
and
so.
A
A
Embryos
that
are
not
like
they
have
these
things
called
embryoids,
which
are
groups
of
cells
that
are
sort
of
grown
in
culture
and
they
sort
of
mimic
the
beginnings
of
a
neural
systems.
So
they
can
build
these
neural
embryos
which
are
like
they
take
a
bunch
of
embryonic
cells,
they're
going
to
differentiate
into
neurons,
and
so
what
they
end
up
with
at
the
end
of
their
cultures.
A
Are
these
neurons
and
sort
of
like
a
ball,
and
so
they
can
examine
sort
of
the
basis
of
these
neural
systems,
very
control
in
a
very
controlled
and
sort
of
simple
way,
and
so
this?
I
guess
this
is
an
attempt
to
do
this
with
embryos
as
well,
so
they're,
looking
at
the
emergent
emergence
of
multiple
spatial
axes
in
embryos,
and
so
the
emergence
of
multiple
axes
is
an
essential
element
in
the
establishment
of
the
mammalian
body
plan.
A
A
A
The
gene
expression
patterns
so
they've
revealed
important
aspects
of
this
process.
A
mechanistic
understanding
is
hampered
by
the
poor,
expan
experimental
accessibility
of
early
post-implantation
stages.
So
here
we
show
the.
G
A
So
they're
doing
the
same
thing
that
they're
doing
with
these
neural
embryos,
where
they're
taking
they're
just
creating
sort
of
a
very
simple
model
of
the
cells
as
they're
differentiating
the
establishment
of
the
three
major
body
axes
in
these
gastroloids
suggests
that
the
mechanisms
involved
are
interdependent.
A
Specifically,
gastroids
display
the
hallmarks
of
axial
gene
regulatory
systems,
as
exemplified
by
the
implementation
of
collinear
hox
transcriptional
patterns
along
an
extending
anterior,
posterior
axis
or
anterior
posterior
axis.
These
results
reveal
an
unanticipated
self-organizing
capacity
for
aggregated
escs
and
suggests
that
gastroids
could
be
used
as
a
complementary
system
to
study
early
developmental
events.
A
So
susan
brought
up
the
point
that
they're
not
taking
physics
into
account,
and
I
don't
yeah.
I
think
this
is
meant
to
be
like
a
sort
of
a
model
system,
but
you
know
just
a
way
to
look
at
like
like
gene
expression
and
gene
expression
changes
in
space.
But
of
course,
if
you
don't
have
physical
forces,
there
are
things
that
are
associated
with
physical
interactions
in
cells,
even
at
the
level
of
gene
expression
that
aren't
going
to
be
picked
up
here.
A
I
think
we've
talked
about
that
in
an
earlier
meeting
where
we
talked
about
like
how
cells
you
know,
depending
on
the
surface
that
they're
plated
on
or
that
they
exist
in,
will
express
genes
differentially,
especially
when
they're
differentiating
so
this.
This
is
some
of
this
work
and
it
walks
through
some
of
the
things
these
gastroids
look
like.
A
So
you
know
they
don't
really
look
like
embryos,
but
they
try
to
build
this
3d
sort
of
blob
and
get
results
similar
results
that
they
might
get
out
of
a
real
embryo
in
terms
of
cell
differentiation.
A
So
this
is,
I
think,
another
interesting
paper,
I'm
not
really
familiar
with
this
model
system,
but
I
think
it's
you.
A
That
you
know
we
get
exposed
to
here
just
a
little
bit,
and
so,
if
you're
interested
in
following
up
on
this,
we
can
as
well.
A
Embryo
we
have
any
other
questions
or
are
we
any
comments
on
any
of
those
papers.
G
Just
this
I
put
in
the
text
box
there
do
you
does.
Do
you
want
the
paper
that
I
know
of
about
optical
coherence,
tomography,
histology.
I
know
it's
off
topic,
but.
A
A
G
Method
of
checking,
histology
images
might
work
to
check
to
see
whether
the
optical
current
tomography
was
working.
I
don't
know
it
was
just
a
thought.
Yeah.
A
E
Yeah,
I
it's
just
a
thought
that
I've
read
a
few
papers
regarding
the
alzheimer's
disease,
where
you
know
the
distance
between
the
synapses.
You
know
increases
that
leads
to
random
memory
loss.
Can
we
just?
Can
we
compare
it
to
the
normal
brain?
Then
you
know
get
some
insights
into
the
new
neurocortex,
like
you
see
that
if
the
things
are
distance,
that
thing
happens
and
it's
just
a
random
knife
thought.
But
if
we,
you
know,
take
certain
brain
diseases
and
compare
whatever
it's.
E
You
know
topology
of
the
brain
at
that
time
and
compare
it
within
a
healthy
brain
and
see
that
how
things
goes
like
some
people
who
are
dyslexic
or
some
people
who
don't
have
you
know
they
have
below
the
iq
of
70,
who
don't
have
much
of
an
intelligence,
and
we
can
compare
the
results
and
see
that
what
is
the
part
of
the
brain
or
you
can
say
what
is
the
topology?
What
is
the
real
thing
that
you
know
help
us
to
you
know
summarize
the
meaning
we
can.
E
B
A
In
mri
they
actually
look
at
the
like
the
structures,
so
they
you
know,
they'll,
compare
like
alzheimer's
versus
a
normal
brain
and
then
alzheimer's
are
different.
Markers
like
the
brain
will
atrophy,
so
the
the
tissue
will
shrink,
and
then
you
can
detect
this
using
mri
fmri
to
see
where
activity
is
so
people
have
done.
You
know,
built
things
called
brain
networks
in
fmri,
so
they
build
they'll.
A
A
And
then
you
know
you
can
compare
regions
between
them
and
build
these
networks,
and
so
people
actually
have
built
brain
networks
like
this,
where
they
have
like
they're,
comparing
different
regions
of
the
brain
or
voxels,
and
then
they
can.
Actually,
you
know,
make
those
comparisons
of
topology.
So
like
this,
those
papers,
I
think,
probably
exist.
E
A
Nice
yeah
yeah,
I
think
yeah,
I
think,
there's
gonna
be
like
it'll
be
topology,
but
it'll
also
be
network
activity.
So
we'll
have
a
tapal.
You
might
have
changes
in
the
topology
like
in
a
you,
know:
alzheimer's
patient,
where
they
might
not
use
parts
of
their
brain
or
they
you
know,
use
them
differently
or
but
then
you
also
have
like
the
activity.
So
you
know
there
might
be
less
activity
between
two
regions,
but
there
might
still
be
a
connection,
and
so
it
would
be
an
interesting.
I
don't
yeah.
A
A
A
Okay,
well,
thank
you,
everyone
for
attending!
Thank
you
for
participation.
A
A
If
you
have
any
comments
on
the
papers,
you
know
you
can
put
them
in
the
slack
channel
or
whatever
and
and
we'll
maybe
talk
about
it
next
week
as
well.
So
have
a
good
week,
see
you
on
slack
yeah,
all
right,
bye,.