►
Description
Previously, Jeff Hawkins has discussed the possibility that reference frame rotations might occur locally in a cortical column. In this research meeting, he proposes an alternate possibility that movement and sensed features are translated in the thalamus. Using vision as an example, Jeff gives an overview of the thalamus and discusses the role and mechanism that thalamo cortical cells might play.
Jeff Hawkins on Reference Frame Transformation in the Thalamus: https://youtu.be/eQv-MjnTodM
- - - - -
Numenta is leading the new era of machine intelligence. Our deep experience in theoretical neuroscience research has led to tremendous discoveries on how the brain works. We have developed a framework called the Thousand Brains Theory of Intelligence that will be fundamental to advancing the state of artificial intelligence and machine learning. By applying this theory to existing deep learning systems, we are addressing today’s bottlenecks while enabling tomorrow’s applications.
Subscribe to our News Digest for the latest news about neuroscience and artificial intelligence:
https://tinyurl.com/NumentaNewsDigest
Subscribe to our Newsletter for the latest Numenta updates:
https://tinyurl.com/NumentaNewsletter
Our Social Media:
https://twitter.com/Numenta
https://www.facebook.com/OfficialNumenta
https://www.linkedin.com/company/numenta
Our Open Source Resources:
https://github.com/numenta
https://discourse.numenta.org/
Our Website:
https://numenta.com/
A
This
is
a
just
another:
it's
a
cartoon
version
of
the
thousand
brains.
Theory,
it's
focusing
on
a
single
column
and
just
to
remind
ourselves
the
basic
idea.
The
basic
idea
is,
you
have
some
sensory
input
coming
from
a
sensory
patch,
whether
it's
just
tip
your
finger
or
an
eye
or
some
part
of
retina.
There's
a
movement
command,
that's
coming
into
the
column,
so
a
quarter
single
column
gets
some
sensory
input.
A
The
problem
with
this
is
there
we
go
is
that
there
is
an
or
a
problem
of
orientation
and
there's
two
things
regarding
orientation
both
on
the
movement
vector
and
on
the
sensory
features.
So
here's
showing
a
finger
touching
the
coffee
cup
and-
and
we
have
a
problem
because
when
you
move
the
sensor,
you're
moving
it
in
some
particular
pattern
like
in
this
case,
the
two
arrows
represent
an
extension
of
the
finger
and
a
movement
to
the
side
of
the
finger
and,
of
course,
how
that
moves
on
the
object.
A
You're
sensing
depends
on
the
orientation
of
the
hand
and
the
finger
to
the
object.
So
an
extension
of
the
finger
on
the
left
picture
means
it
moves.
The
finger
goes
from
that
location
towards
the
upper
rib
lip
and
an
extension
on
the
middle
one
means
it's
going
to
the
left
and
the
extension
and
the
third
picture.
It's
going
some
diagonal
position,
so
we
can't
you
can't
just
take
the
movement
sensor
center
of
data
and
say:
oh,
where
am
I
going
to
be
how
to
update
my
things?
A
B
A
So
this
is
a
problem.
This
is
a
problem,
you
know.
Basically,
you
have
to
take
movement
and
sensation
in
one
reference
frame
and
and
change
it
via
orientation
to
make
it
relative
to
another
reference
room.
That's
basically,
what's
going
on
here
and
I'll,
show
you
a
I
thought,
maybe
fun.
Although
personal
dramatic
example
of
this
in
the
next
slides,
I
hope
this
isn't
too
personal.
A
This
is
I'm
lying
in
bed
in
the
morning
when
I
wake
up.
This
is
what
I
see
and
I
was
in
bed
the
other
day,
and
I
just
noticed
how
dramatic
this
was.
So
I
was
on
my
bed.
I'm
lying
down.
My
head
is
propped
up
on
a
pillow
and
I
look
out
the
window
and
I
can
see
some
water
and
I
usually
look
to
see
what
kind
of
bows
are
out
there.
A
A
So
this
is
the
actual
input
coming
into
my
head
into
my
eyes
and
it's
not
vertical,
because
my
head's
propped
up
slightly,
and
I
noticed
that
when
I
scan
the
horizon,
what
I'm
actually
doing
is
I'm
moving
my
eyes
mostly
up
and
down
at
this
very
steep
angle.
Also,
look
at
the
railing
outside
you
can
see.
A
A
If
you're,
you
know
touching
objects,
you're
constantly
doing
this,
and
you
even
do
it
in
audition,
your
your.
Your
perception
again
is
in
the
reference
frame
of
the
of
the
object,
you're
perceiving
not
relative
to
your
sensors
or
your
ears.
So
this
is
a,
I
argue,
a
ubiquitous
problem.
It's
certainly
ubiquitous
at
some
point
in
the
in
the
in
the
century.
Processing.
A
A
So
vanilla
yeah,
so
it's
just
an
amazing
observation
that
this
is
going
on
all
the
time
they're
just
not
aware
of
it
mostly
so
this
has
to
occur
someplace
in
the
neural
tissue
I
have
previously
in
previous
research
meetings.
I
have.
I
have
gone
through
some
ideas
about
how
this
could
be
occurring
in
the
cortical
column.
So
I
said
I
explore
the
reference
chain
trans.
A
I
call
this
the
reference
ring,
translation
and
and
how
it
could
occur
within
a
quarter
of
a
column,
and
the
basic
idea
here
is
that
you
had
an
additional
layer
of
cells
that
are
representing
orientation.
Those
would
be
equivalent
to
the
head
direction
cells,
but
this
orientation
in
general
and
I
went
through
a
series
of
mechanisms
which
I'm
not
going
to
outline
here.
I
have
the
slides
for
them.
A
I
can
show
them
again,
but
the
basic
idea
is
that
you
would
take
that
information
locally
in
the
column
and
try
to
figure
out
how
you
do
these
translations
and
I
came
up
with
several
ways
of
thinking
about
this.
Remember
I
was
thinking
about
slabs
and
introducing
slabs,
none
of
them
really
very
satisfactory,
and
I
never
felt
like
I'm
working
well
enough.
A
So
today
I'm
exploring
a
different
idea,
the
different
ideas,
the
following,
that
this
translation
is
occurring
in
the
thalamus,
and
so
we
know
that
everywhere
in
the
cortex,
every
cortical
column
information
that
it's
a
feed-forward
information
goes
through
the
thalamus
even
from
other
cortical
areas.
It
goes
through
the
thumbs
and
it
goes
to
these
things
called
relay
cells.
A
And
so
here
you
see
from
my
that,
whether
the
tip
of
my
finger
touching
the
cup-
I
would
have
sense
movement,
a
sensation
from
what
I'm
feeling
and
I'd
have
movement
data
and
it's
going
to
go
through
the
thalamus.
There's,
a
cell
layer
cells
in
the
column
which
we'll
call
layer
six
a
but
it's
just
there's
a
bunch
of
cells
in
them
in
the
in
the
carticle
column,
that
project
back
to
the
thalamus,
it's
a
very
well-known
projection,
and
that
the
idea
here
is
that
the
information
comes
into
the
thomas.
A
A
Yeah,
if
that
one's
a
little
harder
to
argue
here,
it's
absolutely
certain
that
you
have
to
do
this
for
your
primary
sensory
data,
especially
for
vision
actually,
because
in
vision
as
I'll
talk
in
a
moment,
there
is
nothing
else
right.
It
goes
right
from
the
retina
to
the
thalamus
to
the
cortex
touch
doesn't
do
that.
The
touch
has
a
couple
of
intermediate
stops
along
the
way
before
it
gets
to
the
thalamus
and
so
does
audition
I
could.
I
could
show
that,
or
or
maybe
I
could
and
then.
B
B
A
So
it's
like
this
per
click
is
the
old
visual
system
and,
and
then
the
spherical
collisions
does
project
to
the
thumb
as
well.
Again,
I
think
that
is
the
efforts
motor
copy
where
here
the
movement
is
actually
sensory
data.
It's
the
flow
pattern.
Okay,
I'll,
come
back
in
a
second
well,
I
I
could
show
them
both.
I
I
I
don't
know
I
mean
there's
two
issues
here:
what
do
I?
What
do
we
talk
to
you
about
with
carmen
on
wednesday?
A
I
believe-
and
one
of
my
questions
is-
I
believe
this
is
happening
in
every
column.
I
can't
prove
that-
and
I
think
maybe
marcus
asked
me
that
question
once
he
said.
How
do
you
know-
or
maybe
you
asked
me
super
tight-
someone
said
someone
said:
are
you
certain
it's
happening
in
every
column?
Well,
I
can't
be
certain
it's
happening
every
time
it
seems
like
it'll
be
happening,
I'm
certain
it's
happening
for
primary
century
regions,
so
it
has
to
occur
someplace
there,
but
I
think
it's
happening
everywhere.
That
would
be
the
the
the
most
copacetic.
A
Okay,
let's
keep
going
so
this
idea
has
a
lot
of
appeal
to
me
because,
partly
because
the
thalamus
looks
like
you'd
be
good
at
this
and
b,
no
one
has
any
idea
what
the
relay
cells
and
zombies
are
doing.
So
those
two
things
suggest
like
this:
it's
not
a
stupid
idea.
It's
a
reasonably
good
idea.
It
doesn't
mean
it's
right,
but
it's
reasonably
good.
A
So
here
is
a
picture
of
the
of
a
diagram
of
lgn,
which
is
the
primary
visual
nucleus
of
the
thomas.
So
it
has.
It
has
its
own
structures
inside
the
thalamus
and
just
to
remind
people
that
there
are
two
basic
movement,
pat
two-piece,
essential
data
coming
from
the
retina
to
the
cortex
one.
Is
these
magna
cellular
layer
cells
we
and
then
there's
others?
These
problems
are
the
layer,
cells
and
I've
talked
extensively
previously
that
I'm
pretty
confident.
This
is
correct
that
the
magnesium
layer
cells
are
representing
movement.
It's.
A
How
is
how
the
cortex
the
primary
way
the
cortex
learns
how
the
eyes
are
moving.
Those
cells
are
very
fast
response.
They
happen,
they
fire
very
quickly
and
they
require
to
change
changes
and
they
don't
persist.
So
if
there's
a
change
they
fire
and
but
if
it
stops
moving,
it's
like
they
represent,
say
a
dot
of
light
moving
across
the
visual
field
yeah.
A
But
if
that
dot
stops,
they
stop
firing
right
away,
so
they're
only
they're
very
fast,
and
they
represent
change
and
where
the
prophet
saw
themselves
are
the
exact
opposite,
they're
very
slow
to
respond,
so
they
wouldn't
respond
as
your
eyes
are
moving
across
the
scene.
But
when
your
eyes
stop
moving,
they
would
respond
and
they're
persistent,
so
they
might
detect
the
data
light
and
as
long
as
that
data
light
isn't
in
the
same
spot,
the
pervasive
cells
would
continue
to
fire.
A
A
So
both
these
sense
movement
sentence,
data
and
movement
data
are
coming
through
the
thalamus,
so
they
both
can
be
converted
here.
Here's
a
cartoon
drawing
of
okay,
here's
a
an
axon
from
the
retina
or
some
information
from
the
retina
there's
a
relay
cell,
the
blue
dot,
and
then
that
blue
dot
that
relay
cell
projects
to
the
cortex.
A
That's
the
basic
idea
and
they
call
these
relay
cells
because
it's
almost
as
if
a
single
spike
comes
in
on
this
axon
and
a
single
spike
comes
out
the
other
side
and
they
claim
that
there's
no
really
no
difference
between
them
and
obviously
the
thomas
is
doing
something.
But
no
one
really
knows
what
it
is.
So
they
call
these
relay
cells
which,
on
the
surface,
seems
ridiculous,
because
why
would
you
have
all
this
information
and
just
relaying
the
data?
A
But
that's
what
they
re
referred
to?
Relay
cells,
of
course,
are
complex
in
the
runs
and
according
to
the
the
various
papers,
including
in
the
in
the
post,
that
you
sent
me
this
morning,
super
tight.
They
have
maybe
about
6,
000
synapses
on
them
and
that's
represented.
These
are
the
dendritic
branches
that
represent
on
the
left
here,
the
green
arrows
again,
the
output
of
the
cell
of
the
axon
that
would
be
going
to
the
cortex,
and
so
you
have
these
sensory
bits
coming
from
the
retina.
A
A
A
So
these
relay
cells
have
what
are
called
a
burst
mode
in
a
tonic
mode
and
the
burst
mode
is
seem
to
be
when
they
are
most
effective
in
their
activity,
and
so
it's
very
unique
physiology,
and
so
the
idea
that
this
feedback
proposal,
the
feedback,
could
turn
the
cell
into
a
burst
mode.
But
more
specifically,
it
could
do
so
on
a
dendrite
by
dendrite
basis.
That
is
the
essential
element
of
the
argument.
I
believe
in
the
the
poster
that
you've
done
in
the
paper
you're
working,
yeah
yeah.
A
So
then
you
could
essentially
say
well.
This
cell
could
respond
to
one
pattern
of
sentry
bits
under
one
context.
Another
pattern
was
sent
to
bits
or
another
context.
Another
type
of
sentence,
another
context
and
although
I
haven't
worked
it
out
yet,
I
believe
one
could
come
up
with
a
way
that
this
cell
could
learn
to
represent
all
the
orientations
of
a
particular
feature
and
enable
that
orientation
properly
based
on
the
feedback
from
the
cortex.
B
Yeah
in
theory,
I
think
what
we
think
could
be
happening
is
it
could
implement
almost
any
arbitrary
transformation
there
yeah.
So
you
could
learn
a
lot
of
different
powerful
stuff
and
it's
only
limited
by
the
kind
of
the
receptive
field
like
the
how
much
convergence
there
is
that's
kind
of
limits,
how
much
yeah
it
can
do
so.
A
A
A
These
inputs
are
definitely
center
surroundings.
Everyone
says
there's
so
centers
around,
like
you
know
those
if
the
data
light
is
surrounded
by
darkness
or
a
dark
dot
dot
center
over
lightness,
that's
how
these
these
axons
and
the
retina
respond
best,
and
they
talk
about
the
output
of
these
relay
cells
is
also
being
center,
surround,
which
doesn't
make
sense
in
this
context,
right
that
now
we're
saying
no
look.
A
What
these
things
are
recognizing
the
pattern
here,
so
the
output
of
this
cell
should
be
a
more
more
sophisticated
thing
like
an
edge
or
a
line,
or
a
flow
movement
in
particular
direction.
So
there's
these
two
possibilities
here
I
asked
myself:
could
it
could
this
be
just
remapping
remapping
center
surrounds
and
so
like?
I,
I
don't
know
what
that
means.
I
can't
figure
out
how
to
get
it
to
work,
but
that's
what
that's
what
I've
heard?
The
empirical
data
is
that
they
say
yeah.
The
output
of
these
cells
is
a
center
surround
rf.
A
C
And
just
to
remind
you,
this
doesn't
contradict
you,
but
that
does
change
across
species.
What
you're
saying
is
totally
true
of
primates,
but
it's
it's
an
interesting
thing
where
rodents
this
is
oriented
and
cats,
it's
not
in
primates.
It's
interesting!
It's
it's
not
oriented,
and
it
even
gets
all
the
way
to
cortex
and
and
cortex
layer.
4C
is
also
not
oriented.
Cortex.
A
C
B
I
could
introduce
another
kind
of
twist
to
it,
there's
more
and
more
evidence
now
that
the
retina
itself
recognizes
lots
of
different
types
of
features
not
just
on
off
or
anything
for
simple
flow.
There's
all
these
different
types
of
features
that
are
recognized
in
the
in
the
retina
and
just
and
then
I
think,
I've
seen
a
paper
where
that
diversity
is
kind
of
preserved
in
the
thalamus.
So
whatever
so,
when
you're
saying
flow,
for
example,
it's
that's
very
possible
that,
because
the
retinal
you
know
ganglion
cells
that
are
that
are
detecting
flow.
B
Those
would
be
projected
here
and
now
you'd
be
doing
whatever
the
the
we
don't
like
to
call
them
relay
cells.
We
call
them
thalamocortical
cells,
yeah
they're,
not
probably
not
relay
cells,
but
those
tc
cells
can
now
do
all
of
their
stuff
on
flow
and
in
the
case
of
rodents,
I
think
they
think
the
edge
detection
might
even
happen
down
at
the
retinal
level,
okay,
and
so
that
those
those
features
are
just
passed
through
and
then
now
the
lgn
can
do
whatever
it
needs
to
do
with
with
edges
so
well.
Okay,
so.
C
A
Yeah,
although
I
have
trouble
trying
to
get
it
to
work
with
any
kind
of
center
surround
type
of
feature,
it's
like
okay
wedge
and
I
want
the
past
ten
of
surround
features
to
the
cortex.
Can
I
remap
them?
What
does
it
make
sense?
How
do
I
even
think
about
remapping
it
based
on
orientation?
I
have
trouble
doing
that.
Maybe
there's
a
way
I
can't
figure
out
how
to
do
that.
It
just
it
just
you
can't
remap
a
bit.
It's
like.
A
I
can
remap
an
orientation
I
can.
I
can
change
the
orientation
of
something
that
has
to
be
a
vector,
but
I
can't
change
the
orientation
of
a
bit
once
I
I
can
remap
them.
I
could
definitely
put
them
through
a
multiplexer
and
remap
the
bits,
but
I
don't
see
how
to
do
that
in
any
kind
of
a
logical,
consistent
way,
and
maybe
I'm.
A
A
A
I
think
I
think
this
gets
to
the
question
why
I
wanted
to
talk
to
carmen,
because
the
issue
here
has
to
be
what
are
people
actually
recorded
from
in
the
thalamus
under
what
conditions,
because
we
may
be
working
on
data
that
comes
from
a
long
time
ago,
where
the
animals
were
anesthetized
and
as
marcus
points
out,
maybe
different
types
of
things
and
different
animals,
and
maybe
these
are
done
in
slice
preparations.
I
have
no
idea,
you
know
me,
and
so
it's
weird
none
of
this
really
adds
up.
A
I
mean
the
the
classic
view
of
these
is
center,
surround
you
know,
or
just
you
know,
some
feature
comes
in
and
some
feature
comes
out.
Those
features
have
to
be
vectors
and
they
have.
They
should
be
changing
under
different
orientations.
It
shouldn't
be
like
a
one-to-one,
so
I
don't
know
it's
just
weird,
so
so
I
mean
so
it's
theoretically,
this
fits
beautifully,
but
the
experimental
empirical
evidence
is
kind
of
wonky.
A
It
doesn't
fit.
Really
I
mean
in
the
classic
sense
I
mean
I
mean
otherwise.
You
know
you.
You
might
say
that.
Okay,
these
these
cells
in
the
retina
are
responding
to
edges,
but
why
aren't
we
seeing
that
in
the
thalamus?
Is
that
because
people
are
looking
at
primates
as
marcus
suggests
and
other
people
looking
at
rodents-
or
you
know-
I
don't
know
so-
this
leads
to.
Basically
this
leads
a
set
of
questions.
C
A
Any
particular
bit
might
have
to
be
mapped
to
any
particular,
I'm
wondering
if
there's
a
is
it?
Is
it
too
much
computational
they
ask
I
mean
you
just
said:
oh,
you
can't
do
it
too
big
an
image,
but
even
a
small
image,
because
is
it
too
much
to
ask
to
do
that
to
so,
because
you
know
any
individual
bit
could
be
mapped,
could
potentially
be
mapped
to
any
other
bit,
and
then
you
end
up
with
a
very
computationally
difficult
problem
to
solve.
It
seems
you
know,
you
know
what.
A
A
You
know
another
thing
is:
it
appears
to
me
that
we
actually
don't
do
this
in
a
purely
it's,
not
a
complete
way,
meaning
we
have
to
learn
these
orient
these
rotations
and
some
of
them
we
never
learn.
You
know
I.
I
saw
the
picture
for
me
lying
in
my
bed.
If
I
keep
rotating
my
head
until
my
head
is
now
below
the
horizontal,
I
find
it
very
difficult
to
do
the
task,
it's
sort
of
like
looking
at
a
picture
of
a
person's
head
upside
down.
A
We
just
never
see
that
and
it's
really
difficult
to
deal
with
it.
But
if
I
wrote
you
know
so,
you
can
rotate
a
certain
amount,
no
problem,
no
problem
with
the
problem
with
the
problem.
After
some
point
it
starts
getting
harder
and
also
you
just
can't
do
it
anymore.
However,
if
I'm
manipulating
an
object
like
my
pen
in
my
hand,
I
never
have
any
problems
with
it
in
any
petition
or
any
orientation.
A
So
it's
sort
of
like
yeah
with
pens,
I've
seen
them
all
different
orientations
and
therefore
I
can
write.
I
can
recognize
any
particular
orientation.
I
can
mail
it
in
any
particular
rotation
heads.
I
can't
do
that
with
the
wind,
the
view
from
my
window.
I
can
do
a
good
good
portion
of
looking
at
the
world
somewhat
horizontally.
A
A
It
also
reminds
me
of
the
inversion
glasses
that
people
could
wear
where
you
put
on
the
version
glasses.
It
was
really
really
hard
to
see
and
to
move,
but
people
wear
them
after
a
few
months,
you
get
you
get
okay
at
it,
so
somebody's
learning,
this
translation,
we
were
learning
it,
and-
and
so
I
think
I
guess
I'll
again-
I'm
just
going
to
point
out
that
it's
not
like
this
is
this
hardwired
mechanism
that
can
translate
anything
to
anything.
It's
more.
It's
going
to
be
some
sort
of
learn
system
where
it
says.
A
A
Maybe
maybe
you
can
help
me
think
I'll
play
around
with
it,
but
if
you
can
think
of
better
ways
of
thinking
about
that,
that
would
solve
one
of
my
big
problems
because
I'm
thinking
like
okay,
you
know:
why
aren't
we
seeing
these
more
complex
features,
leaving
the
lgn
leaving
them?
You
know
the
relay
cells,
so
these
are
the
questions
I
had
for
carmen.
A
So
some
of
them
you've
already
addressed
already,
I
said,
read
the
app
with
lg
and
cellular
cells
of
center
found
are.
This
is
true.
Is
there
any
contrary
evidence
you
point
out?
There
is
some:
does
anyone
characterize
these
cells
in
awake
animals
and
you
know
again
if
I
took
a
monkey
and
I
and
its
head
was
perfectly
vertical
and
I
was
doing
things
I
might
not
see
anything,
but
you
know
if
the
monkey's
sitting
there
with
his
head
tilted
to
the
side
and
then
I'm
probing
and
he's
doing
visual
tests.
A
A
A
A
A
A
A
A
A
Arrows
so
that,
anyway,
that
doesn't
make
sense,
so
I
was
questioning
that
a
little
bit
too.
You
know
it's
very
hard
for
people
to
know
all
this
stuff
and
so,
and
I've
found
over
the
years
that
sometimes
they
get
it
wrong
and
then
it
becomes
a
sort
of
dogma.
So
you
don't
want
to
assume
it's
wrong,
but
in
the
other
hand,
you
can't
100
of
everything
you
read
is
right.
A
So
that's
a
that's
complicated
so,
but
maybe
it's
more
complicated
than
that.
You
know,
but
this
this
conversion
has
to
be
happening,
and
so
you
know
at
one
point,
of
course,
the
layer,
the
layer,
six
cells
the
orientation
has
to
it
has
to
be
derived
from
sensory
data.
So
you
would
somehow
the
the
legislative
cells
have
to
learn.
A
Here's
another
interesting
number
four
here,
so
the
layer,
motor
output
cells,
every
quarter,
column
has
motor
output
cells
and
these
are
the
layer.
Five
intrinsically
bursting
cells
called
layer,
five
ib,
and
so
these
would
likely
represent
movement
relative
to
the
object
they
would
represent.
As
I
proposed
in
previous
research
meetings
in
the
in
a
particular
mini
column,
that
mini
column
would
be
representing
movement
in
a
particular
dimension
relative
to
the
in
the
object
space
and
the
layer
fives.
We
can
explain
how
they
learn
that
and
then
they
would
be
able
to
generate
behavior.
A
So
I
said
well:
where
is
that
going
to
occur?
Well,
maybe
the
layer,
5
cells
aren't
in
object
space.
Maybe
it's
somehow.
It's
already
been
converted
back
to
center
space,
but
I
I
don't
think
that'd
be
the
case,
so
you
might
argue
well
it's
okay,
that
the
cortex
is
sending
object-centric
vectors
movement
vectors.
Someone
else
will
take
care
of
it.
So
if,
if
layer,
five
cells,
we
know
that
the
projector
is
callicolous
well.
This
very
caliculus
also
has
to
do
this.
A
They
have
a
tonic
and
burst
mode
just
like
the
thalamic
relay
cells.
In
fact,
those
are
the
only
two
cells
that
I
know
of
that.
We've
looked
for
there's
some
evidence
of
similar
three
cells.
Have
this
too?
So
we've
spent
some
time
in
the
past
trying
to
say,
like
how
many
cells
have
this
intrinsic
bursting
mode.
A
So,
but
if
we
take
it
on
the
surface,
it
might
be
that
cells
that
have
this
intrinsically
bursting
or
this
bursting
mode
could
all
be
doing
conversion.
They
can
all
be
doing
multiplexing
that
it
is
possible.
In
fact,
it
is
possible
that
all
parameteral
cells
are
doing
this
possibly
doing
this
some
sort
of
this
multiplexing
I
mean
that's
a
more
radical
idea,
but
it's
it's
like
boy.
That
would
be
another
powerful
operating
principle
if
it
were
true.
B
A
We,
I
don't
remember
we
we
did
some
searching
a
while
back
we're
trying
to
find
out
do
blair
three
cells
burst
and
there
was
some
evidence
that
they
do
and
there's
some
benefits
that,
if
not
many
of
them
do
it,
and
it
was
another
one
of
these
gray
areas,
not
that
you
necessarily
need
to
have
the
bursting.
But
I'm
thinking
like
the
as
you
just
said,
you
know,
temporal
memory
requires
something
like
this.
We
don't
think
of
that
as
a
reference.
A
We
don't
think
of
it
as
a
reference
or
transform
in
that
case,
but
it
is
a
multiplexer
of
swords,
but
it's
it.
In
that
case,
it's
well.
It's
interesting
there
what
we
think
think
about
the
think
about
the
temporal
memory
cell,
the
parallel
cell.
It's
not
it
ideally
just
one
dendritic
segment,
that
is,
that
is
active
right
yeah,
but
but
we
don't
enable
one
or
disable
one
we
just
we
just
we
just
add
them
all
up
right.
A
So
in
the
temple
memory,
it's
like
okay,
any
dendritic
segment
that
generates
an
nmda
spike
it
it
there's
no
enabling
or
disabling
them
they're
just
like.
If
I
recognize
a
pattern,
then
the
cell
gets
the
polarized
here,
it's
a
little
different
here.
It's
like
I
recognize
a
pattern,
but
I
don't
want
to
send
it
to
the
soma,
or
at
least
I
don't
want
to.
I
don't
want
the
somber
to
spike
on
this.
A
Unless
I
have
the
right
context
now
we
never
talked
about
that,
but
there
could
be
an
additional
context
on
the
distal
dendrites
on
on
all
paranormal
cells.
That
would
work
like
that.
So
maybe
in
some
sense
you
could
think
of
it.
This
way,
oh
this
is
this-
is
really
interesting.
It's
coming
to
my
head
think
about
our
temple
memory.
We
have
with
nmda
spikes,
they
depolarize
itself,
but
they
don't
make
the
cell
fire.
A
That's
the
basic
idea
in
this
case
with
the
thalamic
relay
cells.
What
we're
saying
is:
oh
yeah
dendrite
recognizes
a
pattern,
but
under
the
right
context,
which
is
at
the
very
extreme
ends
of
the
dendrites,
that's
coming
back
from
the
sixth
under
the
right
context.
We
want
that
cell
to
fire.
It's
not
going
to
depolarize
itself.
We
want
to
sell
the
spike,
so
you've
got
an
additional
your
computation
here.
A
You've
got
okay,
we're
recognizing
patterns
depolarizing
the
cell,
yes,
but
we
only
want
to
sell
the
spike
if
we
have
the
context
at
the
end
there-
and
I
guess
I
was
thinking
islamic
relay
cells-
I
didn't
think
of
it
as
necessarily
having
this
depolarization
phase
of
predicting
stage
anyway.
There's
too
many
things
going
on
here,
but
the
idea
that
perhaps
in
general,
if
you
think
about
neurons,
do
there
might
be
this
okay
neurons?
Have
these
the
they
have.
A
A
You
have
an
additional
signal
which
modifies
that
even
further-
and
so
that's
that
could
be
more
of
a
general
sort
of
principle.
Interesting.
I
don't
know
well,
anyway,
going
back
to
this
problem.
We
need
to
convert
the
motor
movement
from
an
object-centric
reference
frame
or
internal
reference
frame
to
the
column,
to
to
a
different
orientation
depending
on
the
sensor
orientation
and
those
are
two
possibilities
of
how
it
might
happen.
A
It
might
occur
subcortically
or
it
might
be,
that
the
layer,
five
cells
are
doing
it
themselves
and
in
which
case,
if
it
was
the
layer,
five
in
principle,
bursting
cells,
then
you,
you
would
think
you
might.
The
lay
five
intrinsic
conversion
cells
would
need
to
have
maybe
the
orientation
signal
from
layer
six
project
to
their
distal
dendrites.
I
don't
know
if
that's
known
to
be
true
or
not
so
then
you'd
say:
oh
these
layer,
six
orientation
cells
or
whatever
I
have
orientation
cells.
They
need
to
project
to
the
thalamus.
A
They
need
to
project
to
layer
five
as
well.
So
I
don't
know
if
there's
any
evidence
for
that.
So
there's
a
lot
of
questions
here,
a
lot
it
doesn't
make
sense.
But
what
I'm?
What
I'm
excited
about
is
it's
very,
very
clear
that
this
has
to
occur
and
has
to
occur
everywhere
at
least
often
and
independently,
especially
for
touch
again
with
touch.
I
have
to
do
this
independently
for
each
part
of
my
sensory
skin,
because
they're
moving
around
independently
vision,
one
can
argue.
A
Oh,
we
could
just
do
the
orientation
transfer
once
for
the
whole
eye,
because
they're
all
moving
together.
You
know
so
that's
a
little
bit
more
complicated.
You
know
it
doesn't
have
to
be
on
a
column
by
column
basis
could
be,
it
doesn't
have
to
be
all
right.
Those
are
my
slides
for
carmen
and
my
questions
for
her.
Have
you
got
any
suggestions
beyond
what
you
guys
said
already.
C
A
B
A
C
Another
another
thing
you
might
find
interesting
that
I
can
connect
this
to
your
your
beach
pictures.
I
can't
connect
it
to
the
thalamus
very
well,
but
just
the
experimental
result
we
haven't
talked
about
in
a
long
time
involving
grid
cells.
Yes,
so
one
one
weirdness
of
grid
cells
that
we
haven't
talked
much
about
is
that
and
primates.
A
C
C
I
I
I'm
building
up
to
it
toward
a
second
part
where
it
does
actually
start
to
have
an
external
reference
frame,
but
it's
still
weird
to
me.
I
still
can't
fully
explain
it.
Maybe
let
me
say
the
second
part
and
then
any
any
more
discussion
can
can
happen.
So,
first
that
was
discovered
in
primates.
Then
they
found
in
humans
using
the
fmri
trick.
C
They
were
able
to
find
something
similar
that
two
different
lives
found
that
and
so
so
humans
to
cutting
around
so
so,
okay,
here's
the
experiment,
you're,
look
you're,
looking
at
something
that's
box
shaped
or
rectangular
shaped
you're
performing
some
tasks
like
maybe
following
a
dot
inside
of
it
there's.
There
are
two
different
studies
that
use
two
different
tasks,
and
so,
as
you're
looking
around
this
box,
you
know
they
can
find
grid
cells,
or
rather
they
find
the
fmri
hexadirectional
signal
that
indicates
there
are
probably
good
cells.
C
As
you
look
around
now,
here's
the
interesting
part,
one
of
these
studies,
I'm
I
gave
mad
props
to
josh
julian
who
who
who
chose
to
run
the
second
part
of
the
experiment.
They
took
the
rectangle
and
rotated
it
like
15
degrees,
or
something
like
that,
and
so
now
now
a
person's
looking
at
a
rectangle,
that's
been
like
turned
to
the
side
a
little
bit
and
the
hexadirectional
from
our
signal
rotated
with
it.
C
A
C
A
C
A
Yeah
I
mean
it's:
what's
interesting
is
well.
We
we've
we've
talked
in
the
past
okay,
so
you
got
your
eye,
your
retina
and
it's
at
some
distance
from
something
right,
and
the
question
is
what
what's
being
represented
by
the
grid
cells?
Is
it
representing
where
the
where
the
the
retina
is
relative
to
the
box,
or
is
it
representing
where
the
thing
I'm
looking
at?
You
know
like
the
feature
I'm
looking
at
relative
to
the
box
yeah
that
might
suggest
it's
that
the
latter
it's
like
yeah,
the
visual
cortex
is
like.
A
C
A
You
know,
then,
there's
that
paper.
Remember
that
the
tapers
from
the
chinese
lab
that
I
got
earlier
this
year,
where
they
were
claiming
they
find
grid
cells
in
place
cells
in
v1
itself
and
s1,
and
then
you
said
that
maybe
some
people
were
just
counting
that
work.
I
don't
know
that's.
I
thought
it's
fascinating.
C
A
I
I
just
just
keep
believing
that
the
end
is
all
going
to
be
very
simple
and
clear.
Once
we
get
past
all
this,
somehow
it's
going
to
be
like
oh
yeah,
it's
not
simple,
it's
not
trivial,
it's
complicated,
but
once
you
get
it,
the
whole
thing
will
fall
together
and
all
this
all
these
questions
will
be
obvious.