►
From YouTube: Grid Cells for Conceptual Spaces (NRM Feb 19, 2020)
Description
Numenta Research Meeting - Feb 19, 2020
Discussion at: https://discourse.numenta.org/t/numenta-research-meeting-feb-25-2020/7238
Paper: https://www.cell.com/neuron/fulltext/S0896-6273(16)30707-3
A
B
A
Was
surprised
about
not
confused
about
what
I
mentioned?
It
may
be
obvious
to
somebody
else.
I
know
Marcus
might
others
when
the
question
was,
why
does
this
63
preference
show
up
in
there?
For
my
sake,
where's
it
being
generated
from
and
exactly
what
they
proposed
here
was
I.
Think
we,
previous
discussions
is
that
oh
there's
always
a
conjunctive
sells.
We
know
about
conduct
ourselves
and
the
conductor
cells
represent
both
greediness
and
direction
of
travel.
All
right
I
got
the
turn
with.
A
All
perfectly
lined
up
these
60
degree
angles
and
therefore
what
you're
going
to
really
do
it
in
the
nineteen
between.
So
it's
not
like
you
know
so,
and
therefore
I
view
if
the
activity
of
those
conjunctive
cells,
if
you
measure
them
they've,
been
more
active
on
60
degree,
points
the
directions
and
others
which
makes
a
lot
of
sense.
A
So
so
this
so
to
be
I'm
trying
to
tune
through
together
they
bring
this
up.
They're
saying:
oh
there's
these
conductor
self
observe
direction
and
know
about
those,
but
I
thought
they'd
be
more
appreciated
the
60-degree
angle.
They
never
mentioned
the
fact
that
they
were
derived
from
head
Direction
cells,
so
so
I
just.
E
D
A
A
A
A
D
Paper
was
2016.
3
is
why
there's
no
question
okay,
but
in
2019
people
are
still
saying.
Is
this
happen?
And
and
when
you
talk
to
the
people
who
are
setting
this
question,
they
tell
you
that
that
60
degree
thing
about
conjunctive
cells.
Most
labs
have
tried
to
replicate
it.
In
fact,
oh
the
original
I
think
star
Bellini
paper
for
2007,
where
they
found
that
that
phenomenon
of
the
60
degree
distribution.
A
A
A
A
A
The
point
of
the
matter
is,
is
an
assumption
outside
of
the
hybrid
model
or
outside
of
the
oscillatory
plus
scan
mode,
there's
an
assumption
that
there
is
some
sort
of
directionality
imposed
on
the
whole
network
and
so
I
think
you
could
accept
the
hybrid
models.
Okay,
hyper
model
requires
that
direction.
A
E
E
C
A
A
C
A
A
A
A
A
A
D
D
Having
like
to
essentially
flippers
and
like
a
button,
essentially
yeah
and
I'm,
just
inferring,
this
is
how
it
worked,
but
like
they
basically
decide
how
to
move
these
letters
up
and
down,
one
is
like
the
net
velocity.
One
is
like
a
leg
velocity
and
by
moving
these
up
and
down
and
then
pressing
the
button
that
causes
them
even.
D
A
A
A
Unified
computational
principle
in
the
brain
for
encoding,
physical
and
conceptual
spaces,
as
we
doing
that
can
be
studied
non-invasively
in
humans
being
with
fMRI,
raises
the
host
of
tantalizing
research
questions
and
so
seems.
Like
some
stuff,
you
were
just
asking
students
I
like
how
our
grid
cells,
dynamically,
constructed
or
co-opted
to
encode,
new
concepts
and
I
guess
we're
staying
is
like.
Oh,
we
have
this
two
dimensional
grid
cell
thing
and
somehow
they're
managing
the
two.
You
know:
dimensional
spaces
versus
constructed
and
I
need
to
learn.
A
A
E
So
there
was
this
interesting
paper
looked
at
like
sort
of
the
evolutionary
origin
of
mental
models,
and
one
thing
they
use
this
idea
from
special
relativity
of
look
the
light
cone
and
the
idea
is
that,
like
the
sort
of
space
of
mental
models-
and
in
this
case
you
know
this
Facebook
conceptual
spaces,
it's
sort
of
limited
by
the
sort
of
range
of
behaviors
of
the
of
the
organism
in
the
sense
of
like
this
is
the
sort
of
horizon
that
which
we
can
like
usefully
plan,
and
so.
E
To
it
right
because
at
some
point
yes,
you
would
like
to
say:
okay,
you
know
stop
now,
but
you
know,
even
if
you're
like
projecting
infinitely
far
into
the
past
of
the
future.
You're,
basically
doing
this
within
some
sort
of
constrained
space,
you're
sort
of
repeating,
let's
say
naive,
you're
repeating,
is
sort
of
the
same
set
of
instructions
instead
of
just
like
branching
out
into
like
an
infinite
number
of
no.
A
A
A
B
B
Even
maybe
just
a
couple
of
small
things,
so
for
one
and
what
do
we
think
of
this
idea
that
so,
if
there
already
is
a
good
like
organization
and
it's
supported
by
you
know
some
kin
or
oh
I
mechanism
that
generates
this
grid,
there's
a
lot
of
value
to
reusing
an
existing
grid
rather
than
building
one
from
scratch
every
time
menu
right.
So
one
of
the
things
they
asked
for,
like
longitudinal
version
of
the
concept
you
never
want
to
see.
You
know
that
that
that
formation
happen,
which
obviously
doesn't
you.
A
B
A
Is
that
the
ideal
scenario
would
be
that
the
cortex
have
to
learn
the
dimensionality
of
these
cortex
just
to
learn
the
dimensionality
of
what
is
waiting
at
and
but
that
learning
is
a
longer
term
learning
than
the
objects
you're
going
to
represent
in
that
space
right.
So
if
this
section
of
cortex
learns
about
three-dimensional
visual
objects,
whatever
it
builds
around
three
national
space
is
connect,
trouble,
learning
anything
else
because
it
enough
to
start
all
over
yet
so
it's
like
it
first.
A
A
C
B
Everything
you
know
along
stir
politics
like
lines
of
you,
know,
Democratic
and
Republican.
You
know
liberal
and
conservative,
whereas
in
many
other
countries
these
things
are
a
lot
more
multi-dimensional
axis
because
they
have
five
and
eight
party
Parliament's
with
you
know
much
more
like
where,
where
the
political
fault
lines
are
along
different
dimensions,.
A
C
C
A
A
B
B
A
C
A
C
B
E
A
A
A
A
A
B
One
last
thing:
to
what
extent
do
we
think
that
I
mean
it's
known
for
navigation,
there's,
probably
six
or
seven
modules,
even
though
in
rodent
labs
are
typically
record
three
or
four
just
because
of
my
ya
grits
and
modules.
Because
of
this
spatial
thing
right,
they
presume
that
there's
one
or
two
more,
that
you
could
record
from
if
they
actually
had
arenas
that,
were
you
know,
100
meters,
wide,
there's.
A
C
C
B
That's
cool
like
predicting
for
the
people
building
these
things
yeah,
it's
a
lot
of
data
plus
the
thing
afterwards,
but
I
wonder
for
conceptual
space.
Do
we
even
think
it's
necessary
to
have
several
different
scales?
I
mean
you
like
I
know
that
Marcus
like
Miko
and
like
that
the
defeat
paper
right
argues
for
you
know.
Well,
if
you
have
an
n-dimensional
conceptual
space,
you
can
you
know,
map
it
down
to
2d
grades,
but
but
they
don't
necessarily
need
to
be
a
different
scale
right.
A
C
A
A
The
up
to
10
into
10
categories:
these
are
the
different
way
the
data
changes
in
ten
categories
and
envision
that
can
be
an
orientation
column
and
something
else
will
be
something
different,
but
it
doesn't
know
what
it's
doing.
So,
somehow
all
these
questions
don't
really
on
the
answers
like
what
does
the
data
suggest
right?
The
movement
in
the
data
suggests
that
yes,.
B
E
A
A
C
D
I
wonder
with
all
this
conceptual
space
stuff
in
studying
what
grid
cells
do
spaces.
Sometimes
it
feels
like
the
analogy
of
like
studying
the
elephant,
is
a
bunch
of
blind
people
and,
like
you're,
grabbing
the
tail
and
we're
just
studying
like
one
part
of
the
system,
and
we
can't
see
the
whole
thing
and
and
and
the
level
at
which
we're
studying
it
might
be
right
or
might
not
be.
The
a
lot
of
this
seems
like
things.
C
A
I
would
argue
you
don't
need
to
understand.
I
would
suggest.
I
would
accept
you
with
that.
Maybe
that's
a
fruitful
way
of
going
about
understanding
any
na.
So
it's
a
good
observation
but
I
think
that
says:
okay,
we
better
research
papers
on
that
falls
to
learn
about
it,
but
I,
don't
think
you're
gonna
find
an
answer.
There
I
mean
they're
gonna,
be
things
of
those
things
papers,
they're
just
completely
mess,
so
we
don't
have
to
be
happening,
I
mean
so.
A
A
B
Question
is:
how
do
you
like
the
serendipity
argument
right?
How
do
you
make
sure
that,
once
certain
observations
are
made
that
not
you
know
super
nicely
on
to
some
theoretical
argument
that
people
have
made
fifty
years
ago
without
even
knowing
about
the
brain?
Just
thinking
about
you
know
information,
and
how
do
you
make
sure
that
when
that
happens,
somebody's
there
to
say
oh.