►
From YouTube: Recognizing objects without needing to guess their distance | Numenta Research Meeting
Description
Live Numenta Research Meeting. Adding scale-invariance to our model.
A
B
A
B
Yeah,
so
this
is
a
continuation.
It's
a
much
more
polished
version
of
what
I
talked
about
two
weeks
ago,
so
I
guess
I'll
just
give
like
a
quick
table
of
contents
of
what
I'm,
what
I'm
about
to
show
like
just
I'll
scroll
through
a
few
figures
here,
I
also
point
to
the
whiteboard
a
little
bit
I'll
start
by
pointing
at
the
whiteboard
and
just
introduce
this
notion.
I
just
came
up
with
the
word
yesterday
to
describe
what
I'm
talking
about
I
call
I
call
it
a
partial
pose
that
might
be
a
placement
or
an
injury.
B
There
there's
a
good
chance.
It
feels
like
a
feels
like
a
really
nice
idea
to
keep
crime.
Guessing
people
have
thought
of
it
before
and
it
has
a
name
but
I
just
haven't
found
it,
and
so
the
general
idea
is
I.
Think
a
lot
of
processing
and
the
visual
system
does
not
occur
in
this
form,
where
you're
seeing
objects
at
like
particular
locations
like
this
cube,
that's
nearby
and
the
cylinder
that's
far
away.
A
lot
of
the
processing
can
happen
in
this
world
where
like
where?
B
What
you're
representing
is
an
object
and
an
orientation
at
a
certain
position
of
my
retina
in
a
certain
size
of
my
retina,
which
is
which
is
different
from
actually
representing
the
depth
of
it,
because
I
keep
the
short
distance
yeah,
which
keep
keeps
ambiguous,
how
far
away
the
optics
are
and
how
large
they
are.
And
I
I
want
to
argue
here
that
that
it
seems
really
likely
to
me
that
the
visual
system
does
a
lot
of
its
object.
C
B
Okay,
so
like
in
this,
just
read
the
sentence
and
this
write-up
bike
Center
model
so
that
it
can
recognize
spatial
arrangements
of
components
without
needing
to
solve
the
ambiguity
between
between
whether
the
some
subject
is
nearby
and
small
or
distant
and
large.
Then
they
follow
it
up
by
saying
the
model
does
solve
the
ambiguity.
It
does
figure
out
distances,
but
it
happens
in
the
separate
complementary
process
now
just
to
show
like
the
I'm
going
to
stroll
through
the
figures,
without
going
through
them
in
detail,
but
just
to
show
what
this
is.
B
I
talked
about,
the
subpopulations
of
layer,
6
in
the
cortical
column
and
I
have
the
same
back
and
forth
between
layer,
foreign
layer
6
that
we
have
currently
today,
but
I
divide
up
the
representations
and
6
differently.
I
also
show
where
the
limit
scaling
fits
into
this
I.
Having
this
notion
of
how
large
the
object
is
on
the
retina
and
passing
that
down
to
the
thalamus,
so
that
I
can.
C
B
And
as
I
go
through
this
I
kind
of
do
this
left
and
right
thing
where
I
show
in
the
cortical
column
or
in
neural
populations,
how
this
all
connects
together
and
then
analytically,
thickly
I'm,
defining
what
these
are
and
and
I
and
in
the
writer
right
immediately
after
this
walk
through
these
variables,
I
say
like
this
orientation
of
the
sensor
relative
to
the
object.
How
do
neurons
represent
it?
This
I'm
tell
about
these
variables.
B
Just
a
breeze
through
this
to
show
the
rest
of
it.
I
then
have
the
second
section
where
I
talk
about
layer,
five
and
layer
six
or
where
layer
five
is
defined
as
our
displace
in
that
cells
layer,
which
we,
which
may
be
layer
five
this.
What
possibly
there.
Five
and
I
talk
about
what
it
represents
and
how
it
essentially
represents
the
same
thing
that
it
represents
today,
with
this
extra
notion
of
relative
size,
I
had
this
extra
variable
of
relative
size,
and
so
that's
the
part.
C
B
C
C
B
I'll
go
back
up
to
the
first
image
where
I
show
more
of
the
layer.
Six
connections
to
the
thalamus,
for
example.
It's
these
different
subpopulations,
some
of
it
involves
feeding
back
to
the
thalamus.
Some
is
more
focused
on
feeding
up
to
layer,
four
and
one
of
these
just
kind
of
lives
in
here
alone,
and
this
could
be
like
the
cortical
cortical
cells,
or
this
could
be
something
that's
more,
the
other
layer,
six
populations
that
are
not
projecting
out
of
the
deep
layers.
B
So,
for
this
is
probably
best
for
me
to
point
at
the
board,
so
I'll
just
jump
over
to
this
side,
because
it's
it's
the
same
thing
that
just
described
the
neural
populations
so
just
going
to
order
orientations.
I'm,
gonna,
I'll,
just
order
the
diagram,
so
orientation
is
orientation
of
the
sets
are
relative
to
the
object,
which
is
like
a
point
on
the
surface
of
a
sphere
and
over
the
point
of
the
ring.
I'll
come
back
to
these
labels.
Liquor
can.
D
B
B
C
B
B
I
think
part
of
the
confusion
comes
from
comes
from
when
you're
dealing
with
these
2d
projections
of
three
of
objects,
the
location
of
the
center
determines
which
projection
you're,
seeing
which
means
that,
like
there
can,
the
language
can
get
confusing
that
which,
with
the
view
orientation
like
whether
you're
seeing
like
this
I,
should
not
use
a
symmetric
optic
here
I'll.
Do
that
whether
you're
seeing
this
or
you
know
it's
seeing
it
straight
on.
You
know
it's
not
orientation,
that's
actually
location.
C
C
B
So
so
what
I
could
have
done
was
I
I
could
have
done.
X,
hat
and
II
had
to
do
like,
where,
let's
see
oh
I'll,
say
I'll
make
that
attack.
That's
the
specter
from
here
to
here
is
X
hat
suppose
the
center
is
here
and
D
as
this
polar
coordinates
would
have
been
mice.
Polar
coordinates
would
have
been.
Where
is
the
sensor
like
the
direction
from
the
object
to
the
sensor
and
the
distance?
This
X
has
confusing
yeah
yeah.
D
And
then,
in
addition,
the
I
can
also
rotate
okay
and
when
it
rotates
that
first
thing
doesn't
change
yes,
but
you
need
to
know
both
in
order
to
know
how
the
object
projects
on
to
you
right
now,
okay,
and
so
the
bottom
one
is
the
bottom
one
is
on
there,
which
is
orientation
orientation,
is
how
your
eye
might
be
rotating
around
its
axis.
Okay,
without.
D
C
B
B
B
Well,
keep
in
mind
the
orientation
is
truly
3d.
Three
degrees
of
freedom.
A
direction
is
two
degrees
of
freedom,
drug
sure
sure,
if
it's
color
already
yeah,
so
they're,
two
funded
they're,
two
different
things:
an
orientation
is
not
just
like
an
arrow
pointing
somewhere.
It
was
an
arrow
pointing
somewhere
and
a
roll
around
that
area,
and
so
orientation
is
not
the
same
thing
as
this
thing
you
pointed
in
the
direction.
C
C
D
C
C
B
C
Yeah
well
I
just
respected.
This
has
been
a
confusing
topic
from
the
beginning
because
we
talked
about
orientation,
I.
Think
what
you're
trying
to
do
here
say:
there's
two
separate
sort
of
orientation
like
components.
We
need
to
think
of
them
separately.
It's
on
interpreter
you're,
saying
so:
I
could
be
wrong.
One
is
think
about
you
know:
you're
the
object's
position
relative
to
you
and
the
other
is
your
eyeball
and
your
sensors
position
relative
to
the
object.
Is
that
not
right?
I.
B
B
C
C
C
C
B
B
B
Once
you've
inferred
that,
then
you
truly
know
the
distance
of
the
object.
You
truly
know
its
3d
location
before
you
know
that
you
don't
so
I
will
just
scroll
through
just
to
show
the
remaining
figures.
Just
to
give
you
a
sense.
What
I
did
so
I
defined
the
variables
showed
the
other
two
layers
and
then
the
three
remaining
figures
are
all
variations.
On
the
same
thing
where
I
is
in
this
walkthrough
section,
this
walkthrough
section
I
go
through
time.
A
series
of
events
occur
and
variables
are
inferred.
B
C
B
C
C
C
C
B
Your
yeah
well,
my
point
of
view
on
that
is
that
these
gain
factors
for
path
integration.
You
remember
them
for
a
familiar
object
for
your
coffee
cup.
You
see
it
your
coffee
cup
that
you
it
has
a
particular
size,
so
so
you're
able
to
look
at
it.
You're
like
oh,
that's,
my
coffee
cup
I
know
how
far
away
it
is
because
I
know
how
big.
B
C
B
I
guess
when
you
read
it
in
context,
that's
how
to
talk
about
the
three
walkthroughs
I'm
going
to
do.
I've!
Guess
it's
just
that
I'm
showing
you
paper
out
of
context,
so
I
mean
and
first
I
show
how
the
game
factor
is
inferred
through
sensation
in
movement,
then
that
show
how
the
spatial
relationship
between
two
objects
is
calculated.
C
Inference
so
I
would
have
been
confused
so
so
because
it
makes
a
difference
right.
It's
just
this
I,
you
know
because
it
caught
my
attention.
You
said:
hey
I,
won't
be
able
to
tell
the
distance
I've
been
through
that
by
moving
I
should
know,
I
definitely
adopted
so
I'm
just
pointing
out
there.
When
I
read
this.
D
B
I,
don't
treat
it
but
I
just
I.
Leave
it
open
to
I'm
very
much
open
to
that
possibility.
I
talked
about
that
briefly
and
once
on
the
next
page,
so
the
diagrams
I
was
giving
you
a
sentence
here.
I
think
I
should
start
now,
but
with
motivations
I
have
these
three
walkthroughs.
The
first
thing
first
thing
was
yeah
inferring,
the
game
factor
figure
out
how
large
the
object
is.
Second
part
was
learning
a
spatial
relationship
or
inferring
the
spatial
relationship
so
that
it
can
be
learned
and
then.
C
B
C
B
But
no,
it
can
even
be
much
larger.
It
could
be
a
photo,
it
could
be,
it
could
be
a
larger
or
smaller
version.
You
might
see
a
tiny
elephant.
You
guys
do
that.
This
is
more
of
that
scenario.
You
can
see
that
see
something
where
the
you
can
recognize
the
object.
Without
you
don't
necessarily
have
to
choose
its
size.
You
know
you
don't
necessarily
have
to
choose
as
actual
distance
from
you,
this
the
object
recognition
system
can
work
without
you
know.
B
B
D
D
D
D
C
A
B
That's
the
entire
division,
so
I
I
think
that
okay,
the
model
I'm
laying
out
here
is
very
much
compatible
with
what
we
just
saw.
The
idea
and
I
don't
think
it's
specific
to
faces.
I've
seen
like
art
pieces
that
play
with
this
fact
that,
like
you,
can
have
a
bunch
of
random
shapes
floating
in
air.
If
you
view
it
from
a
certain
angle,
suddenly
your
mind
sees
a
3d
shape
that
you're
confident
isn't
there.
You
had
all
of
the
path
integration
cues.
B
You
need
to
know
that
that
objects
not
there,
but
your
visual
system
will
still
see
a
3d
on
when
you're
floating
there.
If
you
you,
if
all
of
it
coherently
says
the
objects
there
like
a
face
like
Einstein
space,
inverted
you
see
in
it
sticking
out.
I
think
that
there's
good
reason
to
believe
that
this,
the
visual
system
and
how
it
recognizes
objects,
kind
of
does
it
all
about
by
by
what
images
are
on
the
retina
and
by
predicting
from
one
component.
Just
like
okay,
I
see
this
one
sub
object,
this
one
child
object.
B
Therefore,
I
probably
will
see
this
other
child
object.
At
this
other
point
on
my
retina
at
a
certain
retinal
size,
but
a
lot
of
this
processing
lives
in
just
that
language,
purely
it
never
has
to
go
into
the
language
of
how
far
in
each
part
is
away,
and
that's
why
you
get
that's.
Why
I've
done
that
allows
the
system
to
be
more
powerful.
It
also
allows
the
system
to
be
fooled
and
basically
I'm
describing
a
model
that
works.
The.
C
C
Take
something
I've
projected
are
playing,
obviously
the
distance
between
those
opposite
sub
components.
It's
not
defining
that
image
that
I'm
staying
yeah
I
see
this
as
a
building.
You
know
it's
all
flat
fooled
with
that
I'm
being
fooled
lights.
To
mention
a
building,
I
say:
oh
yeah,
it's
nothing,
but
I'm,
not
foolin
at
the
same
idea
is
the
only
information
I
have
the
word
with.
Is
it
all
the
positions
of
those
things?
No.
B
C
The
opposite,
I
find
it
that
it's
simple
community
to
say
it
as
like.
She
doesn't
ask
it's
very
difficult
going
on,
but
if
I
just
use
a
flat
picture
I'm
just
talking
on
me
personally
personally,
it's
easier
to
explain
the
problem.
Your
argument
is:
I
can
infer
an
object
about
knowing
the
depth
of
yep
or
their
components
yep,
and
so
it's
just
the
wrong,
damn
certain.
So
to
me,
like
the
simpler
examples
you
just
take
out.
All
of
that
and
I
still
see
this
ability
yeah.
C
C
C
But
if
I
just
stop
here,
there's
no
knowledge
about
the
world
in
the
structure
the
world
I
will
show
you
a
picture
on
our
campus
ago.
The
structure,
bad
physically
possible
course,
and
so
therefore
I
have
two
different
to
argue
about
depth.
Information
society
was
recommend
with
that
definition.
I
don't
have
it
soon,
don't
have
that
by
moving
and
having
things
change,
alter
each
other.
B
C
B
Yeah,
so
we're
we're
insect,
that's
basically,
the
assumption
of
this
whole
write-up
and
model,
so
I
can
now
move
on
to
yeah
this
process
of
recognizing
objects,
and
here
I've,
given
kind
of
the
playful
minimal
version
on
a
whiteboard,
because
there's
an
accompanying
paper
where
I
talk
rash
of
linear
algebra
and
everything
and
talk
about
neural
populations.
So
this
is
like
the
fun
intuitive
explanation.
B
When
you
see
okay,
I'm
gonna
use.
These
simple
shapes
like
seeing
cylinders
that
are
composed
into
higher
objects
here,
I'm
using
the
model
of
a
hammer
which
is
composed
of
two
cylinders
that
are
arranged
in
a
certain
way
which,
just
as
an
aside,
this
model
is
particularly
appealing
when
these
quote-unquote
objects
are
these
simple
things,
these
simple
things
that
are
composed
into
bigger
things:
the
idea
that
your
eye
needs
to
learn
how
to
represent
a
smoke
Johnny.
B
Your
visual
cortex
is
to
learn
how
to
recognize
a
small
cylinder
and
a
large
cylinder
and
it,
but
it
only
has
to
learn
those
once
it
only
has
done
in
those
ones
such
that
this
big
image
on
a.
If
you
see
a
large
cylinder
on
your
retina
doesn't
matter
if
it's
actually
a
small
one,
that's
closer
a
big
one.
That's
far
away,
it's
the
same.
B
It's
the
same
partial
pose
it's
the
same.
It's
the
same
image
under
retina,
and
that
needs
to
be
learned
once
this
idea
that
I'm
speaking
in
sentence
fragments
but
I'm
trying
to
motivate
like
we're
really
coming
from
on.
This
is
not
really
it
specific
that
coffee,
cups
and
logos
it
is
that
a
little
bit
but
I'm,
really
motivated
by
you
parsing
these
simple
shapes
and
arranging
them
into
and
did
bigger,
spend
of
higher
structures.
That's
where
this
really
feels
apartment
to
me
would.
B
B
No
not
right
now,
or
at
least
my
head's,
not
in
that
place
right
so
going
back
to
this.
So
think
of
suppose
the
sensor
is
attending
to
one
part
of
its
visual
space
and
or
perhaps
looking
at
a
certain
part.
Maybe
this
is
the
fovea
there's
pointing
at
this,
and
it
recognizes
a
cylinder
as
the
skinny
cylinder
at
a
particular
size
of
its
retina
and
and
so
I
call
that
it's
recognizing
a
cylinder
at
a
partial
pose
for
today.
That's
what
I'm
calling
it
it's.
B
B
Yeah
and
it
can
recall
whatever
or
use
some
other
reason,
for
whatever
reason
this
circuit
recalls
a
model
of
a
hammer.
Maybe
it
just
knows
it's
expecting
to
see
a
hammer
somewhere
and
using
that
it
can
now
predict
like
okay.
So
if
I'm
seeing
a
handle
like
this
I
should
see,
I
should
see
the
head
of
the
hammer
up
a
little
higher
and
I
show
that
in
red,
as
a
as
a
prediction
and
I'll
argue
that,
like
I,
mean
I
I
might
seem
like
I'm
head
waving.
B
Maybe
you're
wondering
at
this
point
how
it
predicts
that
I'm,
not
sure
of
that
I
think
it
should
be
intuitively
obvious
that
it's
possible
that
who
cares
how
large
the
hammer
is,
if
it's
a
hammer
that
I'm
gonna
see
the
head
of
this
hammer
at
the
different
part
that
I
write
that
does
not
a
particular
pose
yeah
yeah
yeah,
so
the
intuition
should
make
sense
it
should.
This
should
be
intuitively
intuitively
possible
and
you'll
just
have
to
trust
me
to
some
extent
that
it
works.
B
But
so
now
this
process
of
okay,
I've
inferred
from
I,
saw
the
cylinder.
I.
Think
I
recall
my
hammer.
I
now
realized
that
if
I
attended
this
other
location,
I
should
see
another
particular
cylinder
and
it
attends
there.
Perhaps
the
cons
there,
but
here
I'm
not
really
talking
about
movement.
The
movement
works
with
everything
and
it
can
confirm
that
yes,
I
do
see
a
hammer,
I
don't
see
God,
then
you.
C
That
tells
you
something
that,
if
that's
happening
then
what's
being
communicated,
there
is
well,
let's
see
what
would
we
communicate
there?
You're
communicating
something
about.
Okay,
I
think
it's
a
hammer,
but
it's
a
hammer
out
of
a
particular
pose.
It's
not
just
saying
it's
not
just
hammer,
because
if
I'm
going
to
predict
the
location
of
the
retina
and
the
physical
size
of
the
on
the
retina
of
the
upper
cylinder,
I
have
to
know
its
pose.
C
So
that's
a
more
complicated
projection.
It's
not
just
hammer
it
hammer
a
particular
pose
and
somehow
that
has
to
be
interpreted
by
the
receiving
columns.
Okay,
that's
a
complex
problem
right
there
right
one
comments
to
say:
one
calm,
proceeding
in
our
skinny
cylinder
a
belly
column,
that's
going
to
predict
the
larvae
so
on.
So
what
answer
to
all
of
this
as
the.
B
C
C
B
Mean
like
these
columns
can
I
mean,
will
call
here:
I
haven't
defined
whether
these
are
specific
to
objects,
I'm
just
treating
them
as
pure
locations.
I'll
call
that
six
and
I
don't
know
I
guess
well.
The
first
thing
out
the
first
case
I
was
making
is
that
these
can
simply
vote
and
activate
the
same.
These.
B
This
all
of
them
now,
of
course,
this
is
specific
division,
where
everything
is
the
same.
When
you
talk
about
when
you
talk
about
you
know,
multiple
handsome
Jews,
you
now
have
to
capture
their
their
locations
on
the
different
locations
at
the
body,
using,
for
example,
the
classroom
idea,
but
the
essentially
so
somewhere
I
say
this
somewhere
object.
Id
also
get
devoted
on.
Of
course,
I
could
have
just
called
this
layer,
two
three
and
I'm
just
leaving
it
open.
B
C
I'm
a
column
and
I'm
trying
to
imagine
I'm
trying
to
predict
the
larger
cylinder.
What
information
do
I
need
to
know
when
I
say
complex
I,
don't
mean
like
that
or
difficult
I'm.
Just
saying,
there's
more
information
to
what
the
information
I
need
is
not
simple.
Thank
you.
It's
not
an
object.
Id,
it's
object.
Id
plus
pose
information,
and
and
not
even
not
a
return
of
that
because
it's
like
I
have
to
know
something
about
where
the
skinny
cylinder
is.
C
There
are
many
columns
that
could
predict
the
water
solitaire,
but
only
one
should
because
it's
a
little
occasion
of
this
meeting
I'm
just
this
is
confusing
my
head
right
now,
like
all
the
things
I,
would
need
to
know.
It's
not
clear
to
me
what
those
are.
How
are
you?
How
are
you
both
out?
You
know
I'm,
saying
yeah,
there's
gotta
be
some
common
language
of
the
two
columns
can
use
so.
B
Here's
another
a
2d
sheet
of
column
yeah.
This
is
one
of
the
diagram,
all
of
them
suppose
they
all
activate
the
same
object,
ID
and
the
same
ur6,
being
Disick's,
being
the
location
relative
to
that
hammer
and,
of
course,
as
suicide.
This
object
that
ID
may
not
be
necessary
if
we
find
the
way
to
do
this
with
object,
specific
module.
B
B
B
So
I
can
just
say
a
little
bit
more.
These
going
a
little
going
back
to
these
neural
populations.
Briefly,
you
recognizes,
you
know,
points
on
surfaces
with
a
sphere,
the
one
that,
of
course,
the
one
that
I
never
know.
They're
my
way
to
draw
is
a
scaler.
Something
is
a
scalar.
I
could
have
someone
drawn
a
rate
code.
This
could
be
a
rate
code
I
here,
I'm
just
drawn
is
like
a
uses
scalar
and
code
or
something
you
should
recognize,
these
other
ones,
surface
of
a
sphere
ring
scalar.
B
It's
interesting
to
just
look
at
what
these
populations.
What
happens
if
you
change
one
of
these
and
what
happens
in
the
image,
so
the
location
of
the
sensor
relative
to
the
object?
That's
what
we
were
calling
X
hat
is
the
unit
vector
pointing
from
the
object
to
the
to
the
sensor.
That's
basically
which
projection
you're
sensing
like
it's
like.
Are
you
sensing
this?
That's
the
eraser
from
the
corner?
Are
you
seeing
it
straight
on?
It's
it
chooses
which
projection
you're
sensing.
B
The
size
is
the
most
obvious
size
on
the
retina,
just
kind
of
skills
up
that
projection.
Thinking
small
this
when
you
hold
these
constant,
particularly
when
you
hold
this
one
constant
this,
the
orientation
of
the
sensor
this
this
this
part
of
it
tells
where
the
image
is
landing
on
the
retina.
So,
as
you
see
hot
around,
if
this
is
essentially
tracking,
like
that's
like
vision,
eraser
on
the
surface
of
the
sphere,
this
is
essentially
moving
with
it,
the
tuffet
elsewhere,
it's
landing
on
the
retina
and
the
final
remaining
dimension
is
okay.
B
B
The
final
part
we
haven't
talked
about
much
is
the
game
factor
the
thing
that
it
also
represents
how
large
the
object
actually
is
the
thing
that
you
need
to
know
once
you
know
the
game
factor,
then
you
can
judge
an
object's
distance
from
one
from
one
image
of
it.
So
you
know
how
large
your
coffee
cup
is,
for
example.
So
when
you
activate
your
coffee
pub,
it's
going
to
implicitly
activate
it's
going
to
activate
some
value
here.
It
might
be
right,
but.
C
It's
gonna
activate
some
value
here
and
that's
going
to
be
stable
for
that
couple,
as
you
move
right
need
to
know
that
you
would
know
it
for
even
the
basic
inference
of
the
hammer
I
mean
I
mean
obviously
I'm
learning
than
he
ever
of
different
dimensions,
but
seems
like
the
scale
factor
has
to
come
to
play
again.
They
said
I
can
just
look,
one
I
see
the
chair
and
how
far
away
it
is
so
in
that
one's
eye
inference,
step
seemed
that.
C
C
This
before
and
so
so
I'm
fine
I
just
means
that
part
of
the
inference
process
of
the
learn
chair
is
that
scale
factor
is
involved,
otherwise,
otherwise
I
would
never
senses.
I
would
have
a
sense
of
the
actual
physical
size
of
the
other,
some
construction
of
any
dimension,
I,
don't
know
where
it
is
so
it
long
saying
is.
If
I'm
constantly
aware
of
the
distance
of
that
object
to
me
somehow
on
the
inference
stage,
that
information
is.
B
C
B
When
we're
talking
about
these
high
full-line
objects
like
actual
hammers
or
actual
coffee
cups,
you
have
a
very
strong
prior
about
their
actual
size,
but
when
you
want
to
talk
about
like
what's
going
on
in
there
what's
going
on
to
build
up
to
that
think,
then
those
building
block
layers
are
processing.
You
want
something
like
this.
It's
more!
That's
fine
with
me,
but
I'm
saying
it's
just
there
are
times
when
it
comes
to
the
place.
C
B
B
That
and
it's
appealing
to
me
this
idea
that
the
inference
process
is
essentially
using
these
components
to
vote
this
light
Aslam
Kevin
we
it's
dangerous
for
me
to
use
the
word
vote
right
now
here,
I'm
using
it
more
than
like
the
Jeff
Fenton
way,
the
idea
that
when
you
see
an
eye
that
votes
further
being
a
face
right
here,
when
you
see
an
eye
here,
there's
probably
a
face
you're
easier
I
hear
it
agrees,
there's
probably
a
face
here,
all
right
yeah,
we
can
all
agree.
There's
a
face.
B
The
idea
that
the
top
of
the
hammer
and
the
handle
of
the
hammer
they're
voting
for
there's
a
hammer
right
here,
but
my
new
development
that
other
people
have
probably
thought
about
before-
is
that
it's
not
voting.
There's
a
hammer
at
this
3d
location.
This
voting
on
there's
a
hammer
right
here
in
my
retina
and
that's
the
to
me.
This
is
just
what
one's
that
idea
was
in
my
head.
I'm
like
it's
one
of
those
things
that
felt
like
I
had
to
be
right.
D
B
Mom
yeah,
they
learn
rattles
exactly
exactly
in
that.
That
is
something
that
like,
if
someone
didn't
so
it's
out,
it's
not
if
that
wasn't
clear,
because
I
never
did
say
it
explicitly.
You
learned
this
object
once
I
didn't
talk
about
learning
here,
but
you
said
it's
still
the
case.
You
learn
it
once
from
one
viewpoint
and
all
of
these
mechanisms
will
work
from
other
viewpoints.
So
it's
still
a
3d
model,
but
it
just
breaks
up
the
processing
of
it
a
little
bit
different
angles
to
parts.
C
B
So
I
didn't
talk
about
what
and
where
I
was
afraid
to
to
suggest
that
this
is
divided
into
what
and
where
and
actually
with
this
funny,
I
I
arranged
this
visually
in
a
nice
way
where
I
can
now
make
my
point,
I
was
uncomfortable
putting
this
in
a
where
region,
because
it's
the
only
thing
I
would
put
there.
We.
C
Know,
that's
not
true,
so
you
can
have
both
of
these
in
both
spots
and
then
you
have
to
be
changing
your
reference
frame.
Do
that
right,
so
I
guess!
That's!
A
big
component
is
I'm
struggling
to
figure
out
how
to
parse
that
out.
So
I
can
have
both
of
these
things
and
both
sides,
and
it
would
make
sense.
B
C
C
C
C
B
Other,
just
briefly,
the
other,
the
other
state
and
I
could
make
on
that
subjective.
Well,
currently,
I'm
doing
the
movement
processing
in
buyer
six
and
what
region?
Essentially,
if
this
is
actually
over
in
the
where
region,
maybe
the
the
motor
processing
is
also
over
here.
Maybe
motor
commands
are
actually
coming
into
this
area.
Updating
locations
over
here
using
this
gain
factor
I
would
get.
That
may
be
the
case.
C
B
Other
thing,
I'll
say
about
all
this
totally
different:
putting
the
cortical
columns
aside.
All
of
these
principles
apply
the
capsules
as
well
a
system.
Someone
could
take
a
capsule
network
trained
in
a
particular
way
so
that
it
takes
advantage
of
of
this
idea
of
well.
What
call
it
this
partial
posed
idea,
because
in
there
I
see
in
particular,
it
might
be
really
nice,
because
it
has
these
multiple
layers
of
processing
with
then
the
object
once
at
the
top
and
I
think
that
it
would
benefit
those
in
between
layers
of
processing.
B
A
B
Actually,
learn
something
like
this.
In
that
case,
what
this
can
provide
to
capsule
people
is
giving
them
a
theoretical
understanding
of
lot
of
what
might
get
learned,
get
it
giving
giving
them
a
sense,
giving
them
the
intuition,
for
what
is
the
system
gonna
learn?
I
can
show
like
well,
here's
one
thing:
that's
elegant,
maybe
that's
why
these
capsules
in
a
calculator.
B
Correct
the
first
one
in
the
first
one
is
essentially
fed
in
locations
the
first
paper
they
did
so
for
that
it's
quite
it's
for
that
they're,
like
a
stone's
throw
away
from
incorporating
more
recently,
they
got
away
from
movement
they
just
they.
They
have
this
notion
of
no
compositionality
and
learning
these
features
and
learning
arrangements
of
them,
but
without
incorporating
movement,
and
so
it
is
a
little
bit
more
like
black
magic.
What
it
actually.
C
D
B
Okay,
one,
so
you
just
pretty
much
said
that,
like
okay,
if
you
look
at
this
figure
on
on
the
projector,
these
each
of
these
is
two
cell
population
and
I've
specified
exactly
what
that
population
could
be,
and
so
basically
it's
a
complete
model
that
is
saying
what
these
populations
represent
at
any
particular
time.
What
it's
not
saying
is
how
am
I
actually
computing,
one
from
the
other
I.
D
B
Do
through
through
matrices,
but
the
neural
circuit
that
does
all
that
I
could
do
a
group
force
neural
circuit.
That
has
like
a
very
large
number
of
synapses
that
does
these
mappings
from
one
to
the
other.
It's
all
possible,
but
developing
an
actual
point
of
view
on
what
that
neural
circuit
is,
is
still
an
open
question.
So
so
this
is
ready-made
model,
as
is,
if
you
accept
me,
using
magic
to
using
matrices
using
using
math
to
compute
one
of
these
neural
populations
from
the
other
eye,
because.
D
B
D
B
C
The
other
thing
I'm
trying
to
starting
to
think
about
is
a
vision.
How
does
it
relate
to
touch?
What
are
the
equivalents
these
have
to
these
have
to
project
onto
the
same
fundamental,
so
you
know
seeing
processing
so
how
do
I
actually
published
in
this
point
of
view,
I've
made
some
arguments
about
that
in
the
past.
How
do
you
like
these
two
but
I?
Think
about
that
here?
I
guess
you
mean
you
might
have
something
this
model
here,
but
I
think
it's
I
think
it's
got
enough
holes
in
it
not
close
enough.
C
It's
missing
components
that
that
if
you
simulated
he
you
gain
something,
but
not
as
nearly
as
much
as
you
try
to
solve
to
fill
some
of
those
holes.
First,
don't
be
my
intuition,
you
know,
could
simulate
something.
This
is
a
heated
my
work
and
you
could
go
to
work,
but
then
you
know:
where
are
the
end
of
that?
You
still
got.
It
could
be
problems
to
me.
My
intuition
says
you
gotta
identify
the
big
problem,
still
the
big
missing
pieces.
C
It's
me
you
just
mentioned
something
said
the
time
and
I'm
trying
to
mention
it
myself
right
now.
So
I
I
feel
like
it
would
be
better
than
you
do
for
me.
Well,
personally,
not
a
lot
of
dissimulation,
so
better,
just
sort
of
dig
deeper,
just
to
try
to
figure
out
what's
missing
opponents
here
and
trying
to
figure
out
where
and
how
they're
solved
and
so
money.
But
my
sentence
we're
going
to
that
is
not
the
intention.
C
C
C
D
D
B
B
A
We
are
we're
moving
from
neuroscience,
based
meetings
now
to
more
machine
learning
based
meetings.
So
in
the
future,
our
research
meetings
will
be
focused
on
the
application
of
HTM,
3
ornamental
theory
to
machine
learning
and
deep
networks.
That
sort
of
thing
so
stay
tuned
for
that,
if
you
are
interested
and
thanks
for
watching
I
will
be
streaming
this
afternoon
and
at
1
o'clock
and
about
an
hour
and
a
half
or
so
doing
more
building
HTM
systems
where
we'll
be
talking
about
overlap,
duty
cycles,
after
active
duty
cycle
periods,
etc.