►
Description
Numenta Research Meeting, April 20, 2020. In this meeting, Jeff suggests that grid cell encoding of large location spaces can’t happen just by superimposing multiple grid cell modules. Suggests a temporal memory like SDR encoding of location.
B
I'm
struggling
with
some
issues
about
our
columns,
implement
grid
cells
and
how
the
many
common
hypothesis
work
ties
into
that
and
how
grid
cells
work
at
all,
so
I
just
want
to
run
some
thoughts
and
ideas
by
you.
Guys
and
I
must
I
want
you
to
challenge
me
who
told
me
what
I'm
thinking
wrong
here.
First
of
all,
I'll
start
off.
When
you
see
my
screen,
we
don't
know
yes.
B
So,
the
more
again
since
we
just
Jimmy,
we
really
have
no
idea
how
I'll
create
a
metric
space.
We
proposed
in
our
paper
in
sisal
image
here
and
says:
hey
we
had
a
bunch
of
grid
cell
module
and
say
we'd
work
together,
and
that
was
really
cool
had
all
these
wonderful
properties,
but
it
seems,
like
that's
evidence
for
that.
It's
just
lacking.
B
This
requires
too
many
modules
requires
a
whole
bunch
of
them
and
they
all
have
to
be
sort
of
near
each
other,
and
it
also
requires
them
to
be
the
same
scale
and
as
far
as
I
can
tell
the
evidence
is:
there's
no
evidence
for
this.
You
know
we're
not
seeing
this,
at
least
in
the
entorhinal
cortex
we're
not
seeing
this
kind
of
like
oh
yeah
I
got
dozens
of
these
good
sub
modules
that
are
I
can
pool
from
and
pick
cells
of,
each
and
figure
out.
B
C
I
mean
I,
don't
know
what
you
mean:
I
example
with
this
video
online
and
then
people
who
in
the
community
heard
the
quote
you
just
said:
I'd-
have
to
qualify
it
with
all
sorts
of
things
like
the
idea
that
six
or
seven
modules
isn't
enough.
Like
I,
don't
objects
because
I
know
what
you
mean,
but
I
know
other
people
would
object.
Okay,.
B
Well,
I'm
not
going
to
worry
about
what
to
other
people.
I
mean
you
know.
We
need
a
distributed
representation
and
I'm,
not
aware
of
any
evidence
that
even
like
the
ntarama
cortex
there
are
it
even
multiple
groups
that
even
to
ritual
modules
per
scale,
there
might
be
seven
total
something
like
that,
but
they
vary
greatly
in
scale
in
our
world
that
wouldn't
work
and
and
it
just
it
just
doesn't
I
just
don't
I,
don't
know,
I,
don't
believe
that
and
those.
B
B
Could
be,
but
it's
got
a
working
ranch.
Let's
get
around
the
world
pretty
well
right,
so
I,
just
I,
you
know,
there's
a
whole
bunch
of
pieces
of
evidence.
You
Marcus
may
remember
some
of
them,
but,
like
you,
don't
see
a
lot
of
projections
going
across
the
scales
on
the
grid
cell
models
and-
and
you
know
in
the
spaces-
are
large.
We
will
go.
Look
at
the
tank
image
again
in
a
moment.
If
you
go
sprawl,
it's
a
circuit
right
now
go
back
to
the
tank
image.
B
This
is
one
grid
cell
module
right,
the
one
the
outline
in
pink
and
that's
pretty
big.
That's
like
the
size.
You
know
it's
like
360
or
250
microns
and
when
I
talked
the
tank
he
said
this
was
at
a
different
scale.
This
one
I
believe
he
said
different
scale,
and
so
it's
not
like
by
wanting
to
have
ten
of
these
guys
and
pack
them
in
there.
We
would
be
talking
something.
B
A
B
Guess
all
of
our
work
requires
everything
all
my
thinking
right
now
requires
that
you
have
a
very
large
potential
metric
space
and
that
individual
objects
have
their
own
spaces
and
they're
fairly
high
resolution.
So
you
need
a
large
amount.
You
need
to
have
a
lot
of
modules
and
they
otherwise,
otherwise
you
don't
have
you
need
this
in
your
space.
So
what
I
guess
something
would
I
don't
believe
anymore
is
that
this
is
how
the
brain
creates
a
proper
metric
space,
where
it's
not,
you
know,
I
think
we'd
all
agree.
B
The
individual
grid
cell
module
is
insufficient
to
representing
the
space
of
something
it's
just
the
cells
repeat
every
once
a
while,
and
it's
just
not
enough
representation
you
just
limited
to.
If
there's
20
cells
in
your
you
know,
in
your
face
quadrant,
something
like
that,
the
20
of
them
positions,
you
can
that's
it
we
get
20.
You
know
you
need
something
with
high
a
numbers.
We
need
something
that
says:
yes,
I
can
represent
large
spaces
and
many
of
them
uniquely.
B
B
A
B
D
B
D
There's
something
that
I
haven't
been
clear
with
yet
after
reading
this
papers,
the
I
understand
how
you,
like
men,
I
just
not
how
you
like
your
unique
location
code
by
like
finding
overlapping
prints
or
presentations.
But
in
these
modules,
are
they
like,
in
a
in
a
column?
Are
they
in
like
a
larger
area?
I'll
show.
D
B
D
B
In
a
column,
three
yeah-
this
is
all,
would
be
within
a
comp
because
in
our
in
our
theory,
in
the
thousand
brains,
Theory
each
column
is
able
to
represent
two
complete
sensory
motor
system
and
is
able
to
represent
complete
odds.
So
that
requires
in
that
requires
that
it
be
able
to
have
a
really
large
metric
space,
for
each
object
is
understanding
well,.
D
D
B
And
that's
that's
the
problem.
It's
not
like
I
can
represent
a
space.
I
need
precision.
You
know
when
I
make
a
prediction
about
my
coffee
cup
in
my
finger
it
just
really
precise
to
where
I
am
on
the
coffee
cup.
It's
not
super
precise,
but
it
it's
not
like.
Generally,
you
know
I'm
within
some
half
of
a
cup.
B
No
and
I
don't
be
able
to
do
this
uniquely
for
many
many
objects,
not
just
one
object,
many
other,
so
we
assumed
that
our
papers-
I,
don't
remember
the
actual
numbers
reason,
but
we
assumed
that
we
had
quite
a
few
of
these
mark
from
and
that's
how
we
got
to
be
large
spaces
so
I
just
so
just
I.
Don't
believe
this
is
the
story
right,
I,
don't
think
the
story's
writing
so.
D
I
would
think
that
you
could
get
around
this
problem
to
some
extent
by
the
fact
that
there's
topology
generally
in,
for
example,
visual
cortex,
sensory
cortex,
motor
cortex,
etc
and
the
the
cell
intrinsically
knows,
or
the
column
already
knows,
basically,
which
part
of
the
world
did.
Its
input
is
coming
from.
But
sorry
I
think.
A
B
Know
I
believe
I'm
very
confident
this
individual
column
in
visual
cortex,
if
it's
like
looking
at
the
on
its
own,
could
recognize
objects
and
it
doesn't
need
to
know
where
it
is
relative
to
other
columns.
It's
like
if
you
were
looking
at
the
world
through
this
narrow
straw,
yeah
and
and
only
that's
the
only
thing
you
can
see,
you
can
still
develop
models
of
objects.
You
just
have
to
move
the
straw
around
and-
and
you
can
do
that
and
that's
what
individual
columns
are
doing.
They
they
collaborate
later.
B
They
can
vote
right
and
you
can
vote
to
reach
a
consensus
but
quickly,
but
on
their
own
they're
capable
of
learning.
This
is
the
whole
thousand
brains.
Theory
they're,
capable
of
learning
complete
sensory
motor
objects
for
your
century.
We're
learning
to
complete
the
complete
dimensional
object,
so
you
have
to
think
about
a
column
on
its
own
being
able
to
do
everything
and
it's
limited
to
how
much
you
can
do,
but
it
can
do
everything
so
where
it
isn't,
the
way
represents
on
the
retina.
B
D
Columns
right
and
no
was
thinking
that
it
could
be
partially
true
in
the
sense
that,
like
there
I
mean
there
are
the
lateral
connections
are
immense
right
in
all
cortical
areas
and
assume
that
you
can
sort
of
loosen
that
the
criterion
for
how
many,
how
many
of
these
like
grids,
grid
cells,
read
some
modules
that
will
have
by
gaining
information.
You
know.
B
D
B
Harness
this
one
I
don't
want
to
go
down
that
path.
Maybe
that's
right.
Maybe
it's
wrong,
but
the
moment
I'm,
assuming
a
single
column,
I,
want
to
go
with
this
assumption.
I
think
it's
right,
single
column
can
recognize
complete
objects,
I'm
not
going
to
rely
on
the
neighboring
I
happen
to
know
a
column
does
and
and
if
it
has,
it
has
enough
cells
to
do
this.
We
just
don't
understand
the
representation
scheme
very
well
and
by
the
way,
I.
Don't
think
this
works
in
the
answer
on
a
cortex
either.
This
is
the
method
Weiser.
B
This
all
came
out
of
the
answer
on
a
cortex,
and
you
know
the
rat
knows
where
it
is
I.
Don't
think
this
problem
the
solvent
it's
not
solved
for
that
I.
Don't
think
people
understand
how
it
is.
You
can
go
from
a
grid
salma
agha,
which
is
a
very
queer
representation
of
space.
It's
got
metric
properties,
but
it's
very
poor.
It's
not
suitable,
but
in
some
how
you
go
from
that
to
having
a
very
rich
in
suitable
representation,
space
and
I.
B
Don't
think
you
can
do
it
with
a
lot
of
good
some
modules
and
so
in
a
rat,
even
though
it's
a
rat
has
a
pretty
good
sense
of
where
it
is
in
the
world
and
and
and
let's
go
on
so
to
me
right
now,
unless
I
don't
see
a
way
to
read
at
the
moment,
sting
with
a
single
column,
I
don't
see
the
way
the
system
works,
Marcus,
perhaps,
and
so
I'm
gonna
go
on.
I
wanted
this.
B
B
This
is
what
this
is:
his
image
of
a
grid
so
module
in
Toronto
quartet,
and
so
this
this,
and
this
is
one
grid,
so
module
and
lose
another
one
of
the
yellow
is
another
one
and,
as
I
said,
these
are
changing
in
scale,
and
it
is
not
clear
to
me
whether
there
are
other
regional
modules
in
the
rats
in
Tirana
cortex
that
are
at
the
same
scale
as
this
one.
I,
don't
know
that
I
asked
Hank
that
he
didn't
know,
I
guess
other
people,
they
don't
seem
to
know.
B
So
all
he
knows
that
that
this
thing
of
this
size
rep,
is
one
good
cell
module
and
it's
reasonably.
Why
do
you
think
about
a
rat
for
a
mentor
for
rats
in
Tirana?
Quick
tricks
is
not
that
big
and
we
know
it
has
a
bunch
of
these
at
different
scales
laid
out
linearly
roughly
so
it's
it's
it's.
It
felt
like.
Oh
there's
gonna
be
a
hundred
of
these
flying
around
and
that's
not
possible.
Could
the
Matua
tree
for
maybe
then
even
for
is
not
enough
for
me.
It
might
so
what
was
what
was
unusual
about?
B
This
was
what
thing
did
a
couple
things
you
jump
right
out
of
it
it.
This
is
the
first
time
that
I'd
seen
an
image
which
showed
where
the
actual
grid
cells
are
relative
to
the
grid
cell
module.
Now,
what's
the
purpose
of
this
paper
in
some
sense,
and
so
he
showed
that
there's
actually
in
this
particular
grid
cell
module,
there's
four
sort
of
mini
grid
cell
modules.
Would
you
called
phase
quadrants
in
each
of
these
sick
I
mean
there's
six
of
them
screaming?
B
C
B
What's
good,
that
sort
of
mcdonnen
see
that's
good,
you
know
you
wouldn't
want
and
and
and
then
there's
another
really
really
interesting
thing
that
came
out
of
this
paper,
which
was
that
he
showed
that
in
a
reliable
sense,
sometimes
one
of
these
or
a
couple
of
these
cells
would
not
become
active
when
your
expected
and
it
was
reliable,
meaning
in
the
context
of
a
specific
location
in
an
environment.
Let's
say
this
cell
here
we
just
would
not
become
active
when
you'd
expected
you
and
the
other
ones
would
in
a
different
location.
B
Another
two
cells
would
become
wouldn't
be
active,
so
there
was
like
an
additional
sort
of
coding
scheme
that
went
on
top
of
this
and
it
wasn't
noise.
It
was
very
critically.
It
was
just
very
precise
to
where
the
animal
was.
There
was
another
sort
of
encoding
scheme
that
were
laid
on
top
of
this,
so
you
could,
you
could
say,
by
weeding
out
which
of
the
red
cells
were
active,
that
not
only
are
in
the
red
area
in
the
world,
but
there's
an
additional
coding
like
these
four
cells,
are
active
in
each
yourselves
for
an
actor.
B
So
that
gives
you
a
way
of
scaling
or
multiplying
the
number
of
things
you
could
represent
here.
So
in
the
and
I
wrote
that
up
here,
I
said
thinking,
others
have
shown
their
multiple
phase,
quadrants
in
a
module
and
any
midgel
cell
in
the
quadrant
may
not
be
active,
and
this
is
a
coding
larger
space.
So
in
this
case
you
could
have
up
to
six
shoes
and
different
encodings
of
well.
You
know
assuming
there's.
B
This
for
this
reactor
at
a
time-
something
like
that
so
I
said:
oh,
what
a
community
cell
that
could
not
active
in
6
choose
3,
but
that's
still
pretty
small,
and
then,
if
you
rely
on
that
mechanism,
then
you
maybe
have
fewer
cells
that
are
active,
then
am
I
already
have
like
3
shells
interacted
that
represents
the
location
of
this
animal.
That's
just
not
enough
to
do
anything.
You
know
you
can't
rely
on
that,
so
this
is
I
kind
of
walk.
Try
saying
these
are.
B
This
is
a
really
big
clue
that
this
this
cell
Margo
is
not
just
a
bunch
of
6l
to
redundancy.
There's
some
other
coding
scheme
going
on
there
there's
something
else,
there's
something
else.
It
doesn't
explain
that
why
those
sometimes
our
cells
are
active,
so
I
wrote
it
down.
It
seems
to
be
an
important
clue
and
type
of
trying
to
figure
out
what
that
mean.
So
this
is
not.
This
is
not
a
proposal
how
a
metric
spaces
are
encoded.
It's
just.
D
B
But
I
you
know
it
was
my
recollection
was
that
it
was.
It
was
very
specific.
It
was
like
if
the
animal
recognized
environment.
This
is
what
happened
again.
It's
it
was
part
of
the
encoding
scheme.
You
know
just
like
if
the
animal
good
put
you
put
the
animal
back
to
the
same
environment,
the
next
day
in
recognizes
that
environment,
the
same
shelter,
the
actor
with
the
same
locations
in.
B
It
didn't
look
like
a
working
memory
things.
It
would
look
like
an
encoding
scheme
of
some
sort.
Okay,
they
just
made
an
observation.
They
said
look,
this
should
be
self,
don't
always
fire
and
here's
and
they
had
a
bunch
of
data
about
it.
I
could
read
it
again.
They
have
a
lot
of
data
on
that,
so
but
I
think
they
were
trying
to
show
that
it's
reliable,
it's
consistent
and
it's
location
dependent.
That's
my
recollection,
but
I
haven't
read
the
paper
yeah.
C
And
there's
a
there's:
there
was
this:
there
were
two
papers
that
had
a
results
like
this.
This
one
was
with
Mike's
using
calcium,
imaging
running
on
a
1d
track.
That's
where
all
this
comes
from
there.
Then
there
was
another
one
from
Jeffrey
Dorie
dyrdek
men
then
did
it
with
rats
running
freely
and
2d
environments
and
they
the
papers
even
called
grid
cells,
encode
local
position
and
for.
B
Maybe
you
pointed
out
to
me
Marquis,
but
I
think
those
other
papers
were
earlier
yeah
that
was
already
known.
I
only
observed
it
here
for
the
first
time
that
was
you
right.
There
was
papers
just
on
this
topic
kind
of
stuck
in
my
head,
like
yeah,
that's
a
reliable
result,
there's
no!
That's
multiple
researchers
like
okay,
so
now
I'm
stuck
like
here's
somebody
interesting
data
and
the
method.
We
propose
how
this
works.
B
I,
don't
believe
anymore,
and
now
first
question
is,
and
maybe
that
maybe
Marcus
you
because
of
this-
are
there
any
other
proposed
mechanisms
by
how
large
space
is
represented
with
so
much?
Oh,
there
are
other
people
saying.
Oh,
this
is
how
I
do
it.
You
know
this
one
get
away,
get
away
from
the
greenness.
You
know
cuz
good
cells.
Everyone
agrees
we're
not
very
good
on
the
individual
cell
that
repeats
it's
not
very
good,
so
other
other
proposals,
nope.
C
B
B
B
B
I
know,
but
we
were
talking
I
think
Marcus
was
talking
about.
We
discussed
some
of
the
stuff
before
I
know
you
but
yeah
anyway,
but
yes,
Lauren
talked
about
some
of
this
okay
I.
Maybe
I
could
look
for
that,
but
I
knew
we
talked
about
how
we
talked
about
how
the
whole
data
precisely
the
help
recession
thing
could
be
done
by
the
theta
oscillation
oscillatory
model,
but
I
don't
remember
reading
something.
Is
this
how
you
actually
encode
you
know,
take
a
single
grid
cell
model
this?
B
How
codes
where
you
are
that
could
work,
but
it
seems
in
the
end,
if
you're
going
to
encode
a
location,
you
have
to
have
a
bunch
of
cells
that
are
firing
at
the
same
time
and
that
somehow
uniquely
representation
is
based.
So
even
if
you're
doing
face
some
sort
of
facial
scarring
thing
you'd
have
to
you
still
need,
it
would
be
Majan
take
this
module
here
from
tank
and
you
still
need
a
bunch
of
cells
firing
at
the
same
time
to
represent
the
location.
So
you
still.
C
We
have
this
need
for
this
huge
number
of
unique
location
codes,
because
we're
kind
of
running
with
the
vision
of
you
have
a
unique
location,
representation
for
every
location
of
like
every
location
and
every
coffee
cup
or
whatever.
But
there's
there
are,
of
course,
other
ways
to
do
this,
where
you
might
actually
reuse
the
grid
cells
and
have
another
populations
repartee
representing
object,
identity
or
something
like
that
visually.
B
Neocortex
lecture
of
the
world
and
that's
how
we
did
it
right.
We
had,
we
didn't
know
about
grid
cells.
We
didn't
know
about
this.
I
give
unique
encodings,
so
we
assumed
that
this
location
space
was
going
to
reuse
everything
and
therefore
we
had
to
rely
on
the
object
cell
layer
to
do
that,
and
so
we
said:
oh
look,
we
have
its
objects,
all
they
say
free
and
we
have
you
know
and
and
then
we
had
a
lot
of
issues
with
that
one
I
begin
with
all
the
issues.
B
C
B
B
Okay,
I,
don't
know
my
grid,
there's
a
spectrum.
You
have
a
cellular
mechanisms
using
the
object.
Yes
I
suppose
you
could
anyway,
if
you're
not
going
to
have
a
very
good
metric
space,
then
you're
actually
going
to
have
to
rely
on
some
other
thing.
Some
other
object,
ID
or
pose
ID
or
something
I
mean.
A
D
B
B
Was
more
than
that
anyway,
so
I'm
I'm,
trying
to
come
up
with
scheme
where
you
could
have
you
know
so,
some
of
the
general
properties
we
like,
which
you're
just
like
large
spaces,
a
reason
to
differentiate
able
I
mean
we
achieve
that
in
our
temporal
memory.
The
temporal
memory
allows
you
to
represent
gazillion
things,
because
we
use
this
multiple
cell
from
any
column,
and
until
you
have
a
simple
representation,
which
is
the
many
common
representation,
then
you
have
this
really
really
large
representation,
which
is
the
individual
cells
mini-com.
D
B
B
What's
the
right
word
forth,
there's
no
origin,
and
so
it's
just
this
big
maps
of
space
that
you
can
that
that's
what
this
method
achieve.
The
method
we
described
in
the
frameworks
paper-
and
you
know
some
of
the
other
papers
this.
This
basically
says
the
space
represented
by
a
bunch
of
bits
or
modules,
is
super
super
large,
but
there's
no
origin,
there's
no
center!
There's!
No!
Just
don't
you
know
it's
just
you
start
anywhere
and
you
can
go
anywhere,
but
you
know
the
beauty
of
this.
Is
that
you
could
you
path
integration
from
any
point?
B
You
have
this
almost
infinite
art
space.
They
pick
a
point
and
pass
that
integration
works
and,
and
so
everything
locked
everything
thereby
would
be
all
together
and
they'd
all
be
separated
from
each
other
in
this
huge
space,
so
that
that
represented
some
interesting
challenges,
but
it's
it's
definitely
different
than
you
think
about
it.
In
terms
of
you
know,
engineering
spaces,
where
you
have
an
origin,
there
is
no
origin
and
the
whole
displacement
cell
concept
was
based
on
that
idea.
There's
no
order
to
which
do
these
things,
okay,
so
this
is
good.
B
B
Okay,
now
I'm
gonna,
run
long
run
by
an
idea
and
you
can
recruit
try
to
understand
hopefully
and
critical,
be
critical
of
it.
So
one
thing
the
next
thing
I
did
I
said
well:
okay,
I'm,
really
I'm
really
fun
about
the
grid
cell
thing.
I,
just
don't
see
how
that
this
works
and
there's
some
mysteries
about
it.
Thank
papers
got
mysteries.
It
just
did
something
seems
wrong
here,
so
I
kind
of
my
back.
Is
it?
B
How
would
I
redesign
the
system
if
I
didn't
think
about
the
anatomical,
those
anatomical
and
trainings,
but
I
thought
about
other
anatomical
constraints?
So
I
went
back
to
our
weeny
column,
idea
and
I
said
okay
and
this
by
the
way
it's
got
problems
so
I'm
not
going
to
advocate
this,
but
I'm
gonna
run
this
by
you
anyway.
She
gets
you
to
put
on
I
showed
myself
so
in
this,
this
sentence
here
could
have
Minnie.
B
Imagine
a
mini-com
and
I
made
the
argument
of
a
week
or
so
ago
that
in
many
kind
of
hypotheses,
every
mini
column
has
every
she'll
time.
So
therefore
there
has
to
be
at
least
one
type
of
itself
in
many
columns
there
has
to
be
grits
on
the
make
on
because
every
day,
only
just
every
cell
be
this
remain
calm,
so
I
was
like
yeah
and
and
of
course
see
if
I,
you
know,
there's
about
120
excitatory
cells
in
a
column-
and
you
know
it's
unlikely.
I
would
just
have
one
good
solid
one
displacement.
B
You
know
one
motor
of
it
because
they're
just
there's
more
cells
and
I,
don't
have
a
hundred
twenty
cell
types
in
the
mini
column,
I
might
have
a
dozen,
and
so
that
might
imply
they
mean
I,
have
not
ten
cells
of
each
type,
roughly
paying
waving
other
thing
so
I
said
well,
whatever
mini-com
had
ten
grid
cells
and
they
kind
of
act
similar
to
our
temporal
memory.
That
is
under
unknown
context.
B
B
I'm
making
this
up
there
online
and
it
would
be
like
it'd,
be
a
1d
grid
cell
module
and
if
I
had
400
mini
columns
in
my
column,
I'd
have
400
one
big
Richard
Mitchell
modules
and
that
his
each
one
having
ten
ten
you
know
ten
steps
before
it
repeats
and
that's
a
huge
space,
that's
monster
space
I
could
represent
everything.
I've
been
needed
to
that
would
work
really
well.
B
The
problem
with
that
is
that
then
I
would
expect
to
see
if
that
was
the
case
and
I
and
I
go
back
to
this
picture
and
I.
Imagine
that
each
of
these
column,
you're,
a
member
of
many
column
in
Iraq,
is
like
20
microns,
and
so
this
is
to
be
1/6
of
this.
So
each
of
these
cells
would
being
like
in
a
different
mini
column.
B
Cortex
I
would
have
10
cells
here,
not
one
out
of
10,
and
then
the
question
I
had
was
well
what,
if
that
was
really
happening,
what
would
happen
to
tank
imaging
that
they're
doing
some
sort
of
imaging
technique
and
and
they're
looking
down
on
this
surface,
the
correct
action
to
try
and
see
which
cells
are
active?
Well,
I
had
actually
10
cells
in
each
of
those
locations.
They
would
be
within
about
20
microns
of
each
other
laterally,
and
they
pretty
much
would
be
right
on
top
of
one
another.
B
So
then
I
said
to
myself:
how
do
we
know
that
that,
when
tank
shows
this
image,
how
does
he
know
that
this
is
one
cell?
Maybe
this
is
Tim
cells
here
and
he's
picking
out
anyone
else
and
what's
the
depth
of
his
imaging
technique,
if
it's,
if
it's
more
than
if
it's
less,
if
it's
more
than
20
microns,
he
would
pick
up
every
cell
on
that
and
he.
A
B
D
B
D
A
B
D
Maybe
were
maybe
they
didn't
communicate
this
to
imagine
this
is
your
imaging
plane
and
there's
like
cells
coming
up
right
over
somebody,
so
there's
a
cell
body
over
here.
Sorry,
it's
bad
there's
somebody
over
here
yeah
and
then
there's
a
cell
body
over
here
and
you
see
it
in
a
different
cross
section
in
depth.
The
noise
in
the
in
the
depth
dimension
is
very
small,
but
you
will
see
the
cell
because
it
sells
three-dimensional,
you
will
see
a
cross-section
of
it
and
that's
sufficient
yeah
yeah,
it
works
very
fun.
D
B
Something
we
should
be
able
to
look
in
the
paper.
Maybe
I'll
go
look
at
it
tomorrow,
because
I
was
a
man
I'm
imagining.
If
I
was
a
researcher
doing
this,
then
I
could
set
the
depth
of
the
thing.
I
want
to
set
the
depth
that
I
make
sure
that
I
get
a
good
representation
of
cells
in
all
these
space
projects.
I,
don't
want
to
be
like
you
know,
I
don't
want
to
mr.
cell,
because
it's
10
microns
too
high
or
another
one's
a
little
bit
lower.
It's.
B
Depth
of
plan
then,
and
if
the
cell
bodies
10
microns,
then
I'd
only
be
able
to
detect
cells
that
are,
you
know,
you
know
9
microns
above
9,
my
current
role,
our
type
of
thing
and
my
point,
my
point
is
in
the
actual
experiment:
they
were
not
trying
to
isolate
individual
cells
as
much
as
make
sure
that
they
were
able
to
see
shells
and
all
these
face
projects
at
once
now
it.
This
is
a
tissue
here,
and
this
is
going
over
a
fair
amount
of
distances
here.
B
So,
if
I'm
going
across
360
microns,
what
does
the
chance
I'm
going
to
set
this
thing
to
even
find
microns
of
death,
because
the
chances
of
me
having
a
properly
intersection
all
these
cells
and
some
layers
is
pretty
nil,
because
the
actual
tissue
is
moving
around
it's
on
so
I'm
curious
I
mean
I
would
be
surprised,
I'm
gonna
go
look
it
up.
It
should
be
in
this
paper
what
the
DEP
she
said
it
to,
because
he'd
come
to
try
to
get
it.
B
D
B
B
D
You
get
there's
every
microscope,
looks
at
a
plane
right
yeah,
so
the
difference
between
a
photon
microscope
in
a
normal
fluorescence
microscope
is
that
the
normal
fluorescence
microscope
gets
signal,
gets
light
from
like
different
parts
of
the
Z
plane.
Yeah.
Definitely,
but
the
interference
creates
a
blurry
image
because
they're
not
focused
so
the
two-photon
thing
you
see
more
or
less
only
on
that,
like
Z
plane
what
you
I
think
what
you
would
want.
You
would
need
to
move
your
move,
your
objective,
such
that
it
samples
from
different
depths
and
then
collect
every
frame
so.
B
That's
right,
I'm
gonna
read
their
paper
and
see
what
they
do,
because
that's
critical
in
this
case,
you
know.
What
are
we
looking
at
here
is
really
interesting
question.
If
he,
if
he
moves
the
depth
up
and
down
the
plane
up
and
down
until
he
finds
a
cell,
then
how
much
did
he
have
to
move
it
and,
and
then
is
he
perhaps
sampling?
B
One
of
you
know
I'm
talking
about
something
that
might
be
let's
say
maximum
over
20
microns
deep.
Well,
you
might
have
10
cells
in
that
or
something
like
that
and
and
he
might
have
to
move
up
and
down
20
microns
just
to
get
the
right
style
after
here
until
you
find
a
cell,
that's
actually
fitting
that
requirement.
I
need
the
impression
he
was
imaging.
This
whole
thing
at
once
that
he
was,
he
would
just
set
the
focal
plane
and
the
image
so
yeah.
B
A
B
So
all
right,
this
is
an
area
of
investigation.
For
me,
what
I'm
interested
in
is
it
possible
that
that
there
are
actually
multiple
cells
at
each
of
these
locations
and
the
XY
plane
of
the
cortex,
and
only
one
of
them
is
active
but
they're
finding
that
one
and
therefore
they
say
yeah,
that's
the
cell,
okay
good,
and
this
could
explain
perhaps
why
hats
that
way.
Sometimes
you
miss
one
because
maybe
they're
just
slightly
out
a
plane
or
a
different
point
in
time.
It's
a
different
cell,
but
you
definitely.
D
B
B
D
B
B
D
D
B
I
even
asked
him
the
question
like
well,
these
are
circles
and
I
say
you
think
you
know
why
is
it
always
on
the
border?
Yes,
it's
always
in
the
corridor.
You
know
why
is
this
one
on
the
board
and
this
one's
not
on
the
border?
I
think
I
asked
him
this
sure.
This
wouldn't
see
me.
You
know
overlapping
the
blue,
ending
of
the
atom,
the
why
that
was.
B
B
Is
all
good?
This
is
exactly
one
you
want
to
get
I
want
to
understand
like
how
do
I
interpret
these
well,
maybe
there's
something
else
going
on
here:
is
it
possible
there?
Actually,
many
more
cells
are
representing
this
location.
This
phase
model
we're
only
seeing
what
and
there
are
other
ones
that
are
active
in
different
environments
at
different
places,
and
so
that
was
the
hypothesis
I
was
working
on,
but
that
that
pictures
bit
misleading.
B
B
B
What's
wrong
too
many
confident
a
form,
it's
more
good,
so
modules
each
word
so
modules,
there's
one
D
and
so
now,
I
have
enough
cells
to
sample
from
I
have
I
have
a
lot
of
whole
bunch
of
ability.
Interesting
ideas
like
you
know,
when
the
animal
first
goes
into
environment,
all
those
cells
fire.
Briefly,
just
like
the
temporal
memory,
we
know
where
we
are
so
the
inverse
of
the
temple
memories
working
from
the
bottom
up
versus
the
top
down
the
big
problem
with
this
I
would
love
this
idea.
It's
really
wonderful.
B
The
problem
with
this
would
say
that
I
would
see
active
bit
cells
in
every
single
mini
column
and
that
Clunes
not
look
like
to
be
the
case.
Yet
we
don't
see,
we
see
an
active
bit
so
here,
but
nowhere
else
in
this
phase
project.
We
should
make
it
so
here,
but
no
one
else
in
this
page
quite
and
so
the
whole
idea
is
the
only
one
you
know
the
bump
of
the
bump
map
activity
is
moving
along.
It's
not
like
we
have.
You
know
it
doesn't.
Work,
doesn't
look
like
this.
B
The
problem
and
yet
I
never
have
this
really
fundamental
question.
You
have
them.
You
have
a
set
of
cortical
column
whose,
whose
quote
the
active
many
columns
of
your
first
chapter
feels
the
miniconjou
very
much
based
on
the
sensory
input.
You
think
about
whe
one
okay.
This
is
an
orientation
cell
with
90
degrees,
75
degrees
whatever.
B
Yet
the
grid
cells
themselves,
if
they
were
underneath
that
in
the
comment
so
imagine
this
is
a
cortical
column
and
now
I'm
gonna
hit
the
grid
saw
them.
They
were
six
down
here.
Something
like
that.
These
cells,
then,
are
not
the
clearly
not
driven
to
my
sensory
data
directly.
That
is,
you
can't
say,
okay,
that
a
left
edge
here
that
this
cell
should
be
active.
B
But
it's
not
clear
to
me
how
what
would
what
would
what
would
cause
the
active
grid
cell
to
move
to
another
mini
column
if
this
gets
back
to
the
interference
model,
so
that's
right
into
pairs
model.
It's
just
really
confusing
to
me.
It's
really
hard
to
imagine
how
this
all
work.
Why
and
how?
Long
as
I
have
a
set
of
cells
at
the
bottom
of
the
column
that
have
some
pattern
of
activity
which
I
can
imagine
and
then
they
all
shift
somehow
which
I
have
trouble.
Imagine
it's
just
hard
to
imagine
what
you
know.
B
I
know
I
just
don't
know
if
you're
following
like
my
puzzlement
there,
but
it's
just
really
hard
to
imagine
how
that
would
work.
Alright,
I,
don't
want
to
cut
this
too
short,
but
I
do
have
I
did
get
some
things
out
of
this.
I
didn't
want
to
come
at
your
time,
I'm
going
to
go,
read
the
tank
paper
again
and
see
if
I
can
learn
more
from
its
method
section,
and
unless
someone
wants
to
talk
about
this
and
ask
me
questions
about
it
or
suggest
things
I'm,
not
even
following
it.
C
I'm
not
trying
to
really
start
a
big
topic,
but
I
just
want
to
make
sure
you've
considered
I
think
you
have,
but
the
thing
you
were
just
proposing
with
1d
modules.
If
that's
what
was
actually
going
on
in
the
tank
picture
and
we're
just
seeing
one
of
the
cells
in
that
module,
those
cells
see
if
they're
1d
modules
they're
not
going
to
look
like
good
cells,
they're
not
going
to
look
like
2d
grids,
yeah,
I'm,
gonna.
Look
like
one!
Do
you
like
planes
or
in
bands?
Well,.
B
B
By
the
way
that
works
beautifully
in
terms
of
the
things,
I
need
to
have
a
high
dimensional
space
to
be
able
to
map
motor
commands
on
to
movements
of
the
grid.
Sounds
you
have
enough
cells
that
I
can
only
sample
from
and
get
a
good
representation
of
space
all
those
things
work
beautifully
in
that
regard,
it
just
doesn't
look
like
rich
sauce.
It
looks
like
something
else
on
the
wedding.
I
can
imagine
the
system
works
beautifully
and
it
rips
off
the
temporal
memory
algorithm,
which
works
pretty
well
too
and
yeah.
B
It
doesn't
look
like
that
in
the
actual
tissue,
so
that
that's
why
I
came
on
like
the
grid,
so
we're
seeing
me
some
sort
of
projection
of
what's
actually
happening
that
somehow
we
all
got
this
wrong.
We're
looking
at
grid
zones
that
there's
something
else
going
on
and
that
this
image
that
we
all
have
in
our
mind
of
what
a
good
soul
module
looks
like
it.
It's
not
really
what's
going
on,
it's
just
some
sort
of
what
you'd
see
if
you
looked
at
a
slice
of
it,
something
like
that.
You're.
B
B
So
it's
like
a
step
somehow
or
or
it
could
be
just
it's
just
an
artifact
of
the
methodology
for
how
they're,
looking
like
oh
yeah,
we're
looking
at
this
very
thin
slice,
and
so
most
of
the
good
cells
are
not
seeing
so
they're
whole
other
planes
and
grid
cells
that
are
actually
actively
doing
different
things.
We're
not
looking
at
this
one
well.
B
Well,
that's
a
little
bit
problematic
if
you
think
you
know,
I
could
be
recording
from
a
single
grid
cell
in
the
rat
running
around
an
environment
and
they
clearly
show
this
two-dimensional
tiling
for
that
cell
that
doesn't
fit
here.
You
know,
I
did
what
the
thing
I
just
proposed
doesn't
do
that
it
just
I
would
find
cells
that,
like
Marcus,
said
there
are
sort
of
one
deep,
rich
I'll
they
just
the
only
world
and
proclaim
dimensions.
B
It's
a
series
of
planes,
the
one
mentioned:
that's
what
you'd
see
she
wouldn't
see
the
thing
going
grid
cells
that
you
see
with
singlet
recordings,
I
just
think,
there's
something
wrong
here:
I,
just
it's
known,
as
adds
up
to
make
them.
Something
is
something
is
wrong.
We
don't
know
how
this
works.
I
will
be
anyone
to
house
it,
so
the
evidence
seems
to
be
contradicting
well.
E
E
One
of
the
concepts
of
the
idea
is,
is
how
do
you
get
color
constancy
in
the
human
visual
system?
Where
did
you
no
matter
what
the
illumination
is?
You
still
seem
to
recognize
the
colors
for
what
they
are,
and
the
idea
was
that,
across
the
visual
field,
there's
some
kind
of
integration
going
on
that
abstracts
the
two
phenomena
and
that
integration
of
shifting
of
contrast
as
you
move
across
the
visual
field,
somehow
comes
up
with
an
idea
of
color
constancy.
E
Well
in
that's
it's
not
quite
the
same
thing
as
location,
but
it's
still
a
multi-dimensional
concept
in
terms
of
red,
green
and
blue
and
that
combination.
That
stimulus
remains
constant,
no
matter
how
you're
hitting
with
light.
So
that's
the
only
other
phenomena.
I
know
of
that
does
integration
across
a
large
area.
A
B
I
gave
the
argument:
I
was
talking
about
in
the
consciousness
chapter.
It's
talking
about
qualia
in
different
types
of
qualia
and
I
was
arguing
that
something
like
color
constancy
could
come
out
of
something
it's
a
model.
I
was
arguing
that
it
has
to
be
poor
your
model
of
the
world,
because
color
doesn't
really
just
normal.
B
I,
don't
know
how
that
helps
me
Kevin
I'm,
not
sure
how
the
rednecks
model
helps
me
in
this
regard.
I
need
more
detailed.
You
know
I'm
looking
for
detailed
mechanisms
to
do
this
in
a
very
detailed
mechanism
about
how
the
neurons
actually
do
this
and
how
you
create
STRs
and
so
on.
That's
important.
It's
like
hey,
there's,
some
innovation
going
on.