►
From YouTube: Paper Review - GLOM: How to Represent Part-Whole Hierarchies in a Neural Network by Geoffrey Hinton
Description
Through the lens of Numenta's Thousand Brains Theory, Marcus Lewis reviews the paper “How to represent part-whole hierarchies in a neural network” by Geoffrey Hinton. By focusing on parts of the GLOM model presented in the paper, he bridges Numenta's theory to GLOM and highlights the similarities and differences between each model's voting mechanisms , structure and the use of neural representations. Finally, Marcus explores the idea of GLOM handling movement.
Paper: https://arxiv.org/abs/2102.12627
Other resources mentioned:
Numenta "Thousand Brains" voting alternate version (2017):
http://numenta.github.io/htmresearch/documents/location-layer/Hello-Multi-Column-Location-Inference.html
"Receptive field structure varies with layer in the primary visual cortex" by Martinez et al.: https://www.nature.com/articles/nn1404
"A Multiplexed, Heterogeneous, and Adaptive Code for Navigation in Medial Entorhinal Cortex" by Hardcastle et al: https://www.sciencedirect.com/science/article/pii/S0896627317302374
- - - - -
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
Yeah
today,
I'm
leading
a
journal
club
on
this
on
the
jeff
hinton
paper
on
glom,
the
the
neural
system
called
glom.
The
paper
is
called
how
to
represent
part
hole
hierarchies
in
a
neural
network,
and
the
agenda
here
is:
I'm
I'm
focused
on
the
parts
of
it
that
stand
out
when
viewed
through
the
lens
of
the
thousand
brains,
theory,
and
so
so.
The
this
presentation
has
basically
these.
You
know
these
five
bullet
points
that
I'm
organizing
it
into.
A
First
of
all,
I'm
gonna
kind
of
bridge
from
our
model
to
glom
showing
how
it
votes
on
objects
at
poses
that's
kind
of
something
we
do
in
common
or
that
versions
of
the
thousand
brains
theory
model
has
done
and
then
talking
about
how
glum
then
does
this
thing
where
it
puts
multiple
levels
of
particle
relationships
in
each
of
its
so-called
cortical
columns?
It
calls
them
hyper
columns
the
same
thing.
A
A
The
fourth
point
is
going
to
be
like
some
low
level
details
where
actually
the
neural
representations
that
it
uses
are
a
lot
more
similar
to
ours
in
our
language.
It
would
be,
it
enables
unions,
their
motivation
is
representing
uncertainty.
So
that's
an
interesting
transition,
a
different,
a
difference
from
capsules
and,
finally
I'll
talk
about
glom
and
movement
movement.
B
A
Talked
about
in
the
paper,
but
I'm
just
saying
like
what
the
low-hanging
fruit
is
here:
it
will
it
work
when
it
work
that
I'll
I'll
bring
up
some
discussion
points
on
that
and
I'll
say
that,
so
there
are
other
videos
on
youtube
about
glam,
since
we're
posting
this
publicly.
I
thought
I
didn't
start
this
blurb,
since
I'm
I'm
presenting
a
thousand
brain-centric
view
of
this.
These
other
videos
are
good
too,
though
jeff
hinton
gave
a
talk
and
then
also
this
one
from
yannick
culture
it
was
is
also
really
instructive.
A
Also,
in
the
past,
we
have
a
history
of
talking
about
like
relations
between
thousand
brains,
theory
and
hidden's
work,
both
the
his
old
his
1981
papers.
He
had
many
papers
in
1981
and
I
reviewed
them
all
in
this
presentation,
and
then
we've
also
talked
about
capsules
a
couple
times
and
a
lot
of
the
principles
of
capsules
made
it
into
glum.
So
just
letting
you
know
that,
there's
this
past
discussion-
that's
happened
on
this,
though
I
won't
assume
you
remember
or-
or
I
saw
it
all
so
yeah
jumping
in
to
first
of
all.
A
This
is
this
is
the
part
where
I'm
going
to
talk
about
thousand
brains
a
little
bit
just
to
transition
into
glom,
so
glams
cortical
columns
quote
unquote,
vote
on
object
at
pose,
and
so,
first
of
all,
here's
a
thousand
brains
diagram,
a
subset
of
it.
It's
a
subset
of
the
diagram.
There
are
some
connections
not
shown
here,
but
this
is
like
as
published
in
a
couple
of
our
papers,
I'm
showing
three
sensors
you.
We
often
think
of
this
as
like
three
fingers.
A
It
could
be
three
points
on
the
retina
and
and
what
these
do
is.
First
of
all,
they
have
a
notion.
We
usually
call
it
a
location,
but
I'm
adopting
hinton's
language
in
this
presentation
and
that's
what
these
slides
are
doing
is
introducing
the
terminology.
We,
you
have
some
notion
of
location.
I
didn't
bother
labeling
this
layer,
but
it
predicts
your
sensory
input
which
feeds
into
a
representation
of
what
object
you're
sensing,
which
then
can
be
voted
on
by
multiple
columns.
A
So
multiple
sensors
work
together
to
vote
on
an
object
id
something
we
brought
up
a
bunch
of
times
over
the
years.
Is
that
it's
desirable?
If
this
boat
can
actually
be
on
an
object
at
a
pose,
rather
than
just
the
id
that
just
makes
the
system
much
more
powerful
and
enables
flash
inference,
you
can
have
an
image,
flash
gender
retina
and
recognize
it
immediately
without
having
to
move
around
or
you
can
grasp
an
object
and
similarly
recognize
it
immediately.
A
A
So,
for
example,
you
know
you
have
your
sensors
relative
to
your
body
like
you're,
like
your
hand,
you're
reaching
your
hands
around.
You
touch
something
if
you
can
move
what
you
sense
into
a
common
reference
frame
and
your
body's
reference
frame,
your
arms
reference
frame,
some
common
frame
similar
on
retina.
If
it's
relative
to
the
full
camera,
then.
D
Marcus,
I
hope
you
don't
mind
me
interrupting
yeah,
so
I
just
want
to
make
sure
some
things.
You
said
it
was
quick
and
I
didn't
follow
it
the
way
I
thought
you
know.
As
you
point
out
in
our
papers,
we
talked
about
voting
for
objects
right
and
then,
and
then
we
later
said.
No,
it
really
we're
not
really,
that's
not
what
you
should
be.
It
really
should
be
object
to
pose,
which
is
a
different
thing
like
that,
that's
saying
what
the
object
is
and
what
its
relationship
it's
interesting.
D
E
A
A
A
Doing
it
possibly
like
this
as
hinton
discusses
in
this
glam
paper,
like
one
of
the
things
he's
he
described,
is
why
he
didn't
only
do
one
population
why
he
replicated
this
multiple
times.
I
don't
know
I'm
not
presenting
very
well,
because
I
never
introduced
this
visualization.
But
the
point
is
yes.
You're
right,
I
did
do
a
little
hand
waving
here.
That's
why
you're
confused
okay.
D
A
D
E
G
So
the
question
I
have
is
on
the
top
level,
where
you
show
intercommunication
on
that
level.
Is
there
any
sense
of
what
the
level
of
connectivity
is?
I
mean
you,
you
schematic,
you
know,
you
know
just
you
know
single
connection
across
the
columns,
but
I'm
assuming
that
it's
a
it's
a
2d
set
of
connections.
Is
there
any
constraints
he
provides
on
that
degree
of
interconnectivity.
A
C
A
Cells,
they
have
the
same
activation.
Then
he
goes
into
a
later,
biologically
plausible
section
where
he
talks
about
how
you
could
get
something.
That's
a
little
more
realistic.
So
he
does
so
he
the
way
they
do
it
in
glom
is
literally.
Have
it
be
the
same
vector
the
same.
You
know
1024
bits,
let's
say
and
the
same
one
is
going
to
vote
for
the
same
one
I
see
so.
G
E
D
G
D
The
adjacent
ones,
no,
no,
no,
it
could
be
it
could
be.
These
are
the
medium
and
long
range
connections
in
the
cortex.
So
you
know
if
you
look
at
connections
between
the
the
left
side
of
the
brain
and
the
right
side
of
the
brain.
Well,
you
see
a
lot
of
connections
between
the
parts
of
the
somatocentric
cortex
represent
the
left
hand
and
the
parts
represent
the
right
hand.
D
So
in
theory
I
mean
that
would
we
would
say
that's
because
often
the
left
and
the
right
hand
are
touching
the
same
thing
and
they
have
it's
fruitful
for
them
to
vote,
and
so
any
any
part
of
the
cortex
that
could
fruitfully
vote.
You
might
have
connections
the
the
other
interesting
thing
about
these
connections
and
I've
mentioned
a
few
times.
I
don't
think
we've
ever
really
tested
this.
Maybe
you
haven't
remember,
but
you
don't
need
as
many
as
you
think
you
need.
D
That
is
if
I
was
a
single
column
and
I
need
to,
I
need
to
be
able
to
we're
voting
on
something
I
maybe
need
to
get
a
hundred
connections
to
vote.
You
know
from
someplace
else.
Those
hundred
connections
could
be.
Could
be
randomly
distributed
among
100
different
columns,
that
is,
every
column,
doesn't
have
to
be
very
much
connected
to
every
other
column.
It's
as
long
as
I
just
have
a
sufficient
number
of
I.
D
I
can't
prove
this
right
now,
but
I'm
sure
it's
true,
I'm
pretty
certain
too
that
that
the
the
densities
connections
can
be
much
smarter
than
you
might
imagine.
Initially.
As
long
as
each
column
has
enough
bits
coming
in
from
someplace
else,
the
whole
system
will
settle
so,
but
so
you
would
see
these
going
in
locally
you'd
see
them
within
a
within
a
region.
You
would
see
them
between
regions
that
are
adjacent
to
each
other.
You
might
see
them
across
the
cortex.
You
know
different
size,
corpus,
callosum,
kind
of
thing.
G
D
E
That's
okay,
just
to
make
sure
I
understand
here
kind
of
one
big
difference
when
we
have
in
the
thousand
range
theory
the
the
top
population
was
stable
for
an
object,
regardless
of
the
pose
right.
So
it
was
sort
of
an
invariant
representation
of
an
object
here
when
we
have
an
object
that
pose
that
representation
is
obviously
changing
as
the
pose
changes
correct
right,
and
so
you
don't
necessarily
have
a
stable
representation
of
an
object,
but
you
might
have,
I
guess,
he's
used
the
term.
D
So,
that's,
that's.
Obviously
that's
the
thing
I
was
saying.
I
haven't
really
kind
of
worked
this
out
in
my
head.
Maybe
marcus
you've
done.
I
don't
know
where
I
still
know
the
identity
object,
although
the
poses
change
one
way,
one
way
to
think
I
think
the
one
where
I
think
about
it
is
that
we're
only
perceptually
aware
of
these
voting
neurons
and
so
what
you
perceive.
What
am
I
you
know,
I'm
looking
right
now
I
see
my
computer
in
front
of
me.
It's
not
you
know.
D
My
eyes
are
moving,
but
the
computer
doesn't
seem
to
be
moving
and
so,
but
I
perceive
the
computer
at
being
its
own
post
to
me,
I
don't
just
say
computer,
it's
it.
Oh
it's,
my
computer
at
some
particular
position
and
orientation
to
me,
and
that
is
the
thing
I'm
perceiving
so
that
is
those
are
going
to
be
voting
neurons
that
represent
that.
Yet,
at
the
same
time
I
do
know
it's
my
computer.
It's
not
just
like
it's
not
unique
from
my
computer
and
other
positions.
D
F
Yeah
just
a
quick
question:
I'm
just
scheming
the
paper.
There
is
this
notion
of
connections.
I
think
between
parts,
something
like
that
yeah.
So
is
it
as
sparsely
as
jeff's
explaining
that
he
feels
that
you
only
need
some
of
the
connections
and
since
you
know
they're,
you
only
need
some
of
the
connections
to
make
sure
they're
all
connected
to
each
other.
So
is
that
notion
of
spicy
in
this
paper
as
well
or
are
all
this
connectivity
then.
A
A
G
D
A
D
C
D
So,
but
but
okay,
but
so
you
look
at
your
objective,
pose
or
voting
well.
That
implies
in
this
picture
that
it's
like
one
population
is
voting.
This
is
the
thing.
That's
always
confused
me
about
this.
If
we
can
say
that
oh
there's,
two
populations
that
are
voting
orientation
and
location.
That
makes
it
easier
for
me
to
think
about
it.
So
is
that
what
you
think
is
that
the
way
to
write
that
would
you
think
about
it?
Then
there
is
no.
There
are.
A
B
A
It's
valid
yeah,
okay,
so
one
last
thing
relating
the
two
and
the
voting
on
the
left.
I'm
gonna
I'm
showing
like
what
happens
at
one
level
of
glom
roughly.
This
is
just
bridging
the
gap.
I
had
to
contort
this
a
little
bit
to
make
this
point,
but
but
at
its
core,
it's
taking
sensory
input
and
the
location
of
the
sensor
in
a
shared
reference
frame,
the
its
most
common
one
it
uses
is
simply
where,
on
the
retina
is
this
this
pixel?
Basically
but
it,
but
it.
A
C
A
Deep,
this
is
and,
and
it
outputs
an
object
at
a
pose
and
bringing
in
our
model
on
the
right
here.
Ours
takes
sensory
input,
figures
out,
object,
pose
relative
to
sensor
and
then
takes
this
location
input
and
puts
it
into
a
shared
reference
frame
and
just
to
be
cute.
I
made
them
vote
together,
but
the
the
statement
I'm
making
here
is
that,
like
from
a
high
level
like
from
afar,
you
can
look
at
this
alternative
version
of
the
3000
brands
model,
because.
C
A
Stupid
pointed
out,
we
usually
vote
on
identity,
so
I'm
going
to
always
call
this.
The
alternet
version,
the
one
that's
voting
on
an
object
at
a
pose.
You
can
look
at
it
as
like.
What
we're
doing
at
a
single
level
is
a
highly
specified
version
of
what
the
glom
network's
doing
at
a
single
level.
It's
leaving
it
open
to
just
train
how
whatever
it
wants,
where
ours
in
in
principle,
it
could
train
this
network.
A
A
So,
first
of
all,
I'm
going
to
talk
about
how
glam
includes
multiple
levels
of
part,
whole
relationships
in
each
cortical,
column
or
hyper
column,
and
and
I'm
going
to
also
talk
about
how
the
voting
and
glam
doesn't
all
try
to
settle
on
one
thing
at
least
not
on
the
not
on
most
of
the
levels.
It
tries
to
it'll
be
better
with
a
picture
which
is
in
the
next
slide.
It
creates
these
islands
of
agreement.
A
This
is
showing
five
different
populations
across
across
six
different
columns,
and
it
hierarchically
goes
up
these
vectors
here.
He
just
draws
these
simple
2d
vectors,
but
it's
actually
a
stand-in
for
a
high-dimensional
population
just
making
up
a
number,
maybe
a
thousand
cells
each
I'm
just
making
that
up
and
and
within
the
so-called
column.
As
you
go
up
it,
you
arrange
these
low-level
parts
into
subparts
into
parts
into
objects
and
and
what
you
calls
the
scene
level
embedding
at
the
top
and
as
you
go
up
it,
they
they
agree.
A
More
and
more
until
they
are
agreeing
on
the
single
until
they
are
voting
on
a
single
thing
typically,
but
in
the
intermediate
steps,
which
are
all
voting
horizontally
at
their
at
their
particular
level.
It
kind
of
by
design
creates
these
islands
these
islands
of
agreement,
where
these
three
columns
will
agree
on
this
level.
These
three
columns
will
agree
on
this
level
and
then
on
the
next
level.
They
they
all
six
agree
I'll,
throw
in
a
terminology
point
here.
We
usually
use
the
word
level
totally
differently
from
this.
A
So
I
made
my
point
basically
the
two.
The
two
points
I'm
getting
across
here
are
it's
it's,
this
five
level.
He
chose
five
because
it
just
seems
realistic
from
what
we're
capable
of
doing
like
if
you
like,
psychologically
and
the
five
levels
in
in
one
column,
with
voting
at
every
single
one
of
those
levels,
all
all
resolving
into
a
a
coherent
scene
level.
A
Vector
at
the
top,
this
is
a
it's
a
recurrent
network,
where
it's
processing
input
over
time
and
every
level
is
receiving
input
from
the
bottom,
that
from
above
it
below
it
and
from
the
side
and
all
just
the
system
is
trying
to
resolve.
It's.
It's
very
much
like
a
voting
system
where
it's
also
receiving
input
from
the
bottom
and
from
the
bottom
and
top.
A
D
Where
are
the
sub-components
of
this
picture,
where
they
are
relative
to
other
sub-components
of
this
picture
and
he's
doing
that
in
parallel
without
any
movement
of
the
sensor,
and
this
figure
reflects
that
this
figure
says:
oh
in
one
part
of
the
image
the
we
can,
these
columns
on
the
left
here
kind
of
blue.
You
agree
on
the
blue
thing.
That's
you
know
some
some
part
of
the
image
and
then
we
go
to
the
third
level
well,
slightly
larger
part
with
the
red
arrows.
D
Oh
there's
now
there's
more
columns
of
the
green
on
that
the
the
blue
and
the
yellow
were
part
of
our
reddit
red
thing
and
so
on
we're
in
the
thousand
brainstorm
releases.
I
know
in
brains
it's
not
just
a
thousand
there's,
but
in
brains.
What's
the
difference
is,
is
that
we
attend
two
parts
of
the
image
and
our
eyes
are
moving
and
attending
the
one
part
to
the
next
part
to
the
next
part
and
each
attention
each
movement
of
the
eye.
D
If
we
then
figure
out
the
relative
positions
of
the
thing
that
was
attended
to
at
that
moment,
so
in
the
brain,
you
would
build
up
this
hierarchical
composition
through
movement.
You
do
one
piece
at
a
time
just
like
walking
through
your
house,
you
build
up
one
room
at
a
time
that
you
do
this
as
you
move.
D
Oh
the
upper
left
corner
of
the
image
is
going
to
be
one
thing,
and
the
bottom
right
corner
is
another
thing
and
that's
where
you
might
see
the
red
and
the
green
arrows
here.
They're
like
saying:
oh,
we
see
a
boat
in
the
bottom
right
and
we
see
a
person
on
the
upper
left,
but
that's
all
happening
in
parallel,
as
opposed
to
any
kind
of
movement.
I
know
that
helps
with
your
confusion
about
this,
but
that's
I
that's
my
interpretation
of
this
is
that
that's
all
correct.
Okay,.
H
I
think
he
mentions
at
the
start
that
he
uses
kind
of
an
auto
encoder
in
in
his
kind
of
original
interpretation.
I
was
just
wondering
exactly
as
in
is:
it
is
the
idea
that
the
kind
of
low-level
representation
is
being
taken
to
a
kind
of
through
potentially
multiple
levels,
in
a
neural
network
to
some
hidden
representation
and
then
which
is
in
in
his
view,
is
lower
dimensional
or
or
how
is
I
don't
know
yeah?
Does
he
get
concrete
about
how
that's
implemented.
A
So
well,
just
first
little
note
just
detail
that
I
didn't
mention
the
the
connections
from
here
to
here
from
from
one
level
to
the
next,
and
also
from
the
the
top
down
level.
Each
of
those
has
a
mini
deep
network
involved.
It's
it's
unspecified
how
many
layers
it's
going
to
be
so
it's
kind
of
it's
going
to
be
like
a
little
feed
forward
network
from
here
to
here.
A
little
feed
forward
now
we're
from
here
to
here
and
the.
A
H
G
A
H
C
C
E
B
D
No,
you
don't
talk
about
that.
Remember
in
this
case
the
columns
are
associated
with
a
physical
part
of
an
image
and
again
because
there's
no
movement
involved,
the
left
marsh
column
is
always
going
to
be
associated
with
some
multiple
scales,
black
being
the
smaller
scale,
blue,
large
and
so
on,
but
it's
always
associated
like
imagine.
You
just
got
a
picture.
It's
a
two-dimensional
image
and
that
column
is
always
focused
on
one
part
of
that
image.
D
A
Here
I
should,
I
should
throw
in
the
next
slide
here
then
we
can
continue
this
just
in
case,
so
I
consider
both
possibilities
here.
One
is
if
it
is
all
in
one
column,
first
image
that
came
to
mind
is
this
idea
of.
If
you
look
at
cells
and
the
cortex
as
you
go
through
the
layer
depth,
you
get
more
like
complexity
going
up
or,
and
they
get
larger,
so
it's
it's
evocative
of
that.
A
A
D
But
his
his
use
of
these
terms
suggests
to
me
that
he
he
he's
just
using
very,
very
casually
related
to
biology
it's
it's
not
like.
There's
a
he's
thought
about
it.
Much
of
it
doesn't
seem
like
he's
thought
much
about
the
biological
aspects
of
this
at
all.
He
says
he
had
people
talking
about
calling.
I
call
this
a
column
too.
You
know
that
kind
of
thing,
as
opposed
to
the
kind
of
questions
you're
asking
without
you.
D
A
D
But
even
just
ways
I
mean
this
system
that
he
proposed
it
couldn't
possibly
work
for
touch,
for
example,
or
for
hearing,
because
you
don't
have
this
sort
of
a
regular
mapping
idea
that
you
have
like
a
picture.
You
know
that
you're
processing,
where
columns
are
always
in
the
same
location
in
that
picture,
you
know
what
does
that
mean
in
touch
with
fingers
and
moving
and
body
parts
and
moving?
It
doesn't
make
any
sense.
So
again,
I
think
it's
I
think
he
comes
out
from
the
computer
vision
perspective.
D
D
D
He
has
a
lot
and
he's
showing
the
cortical
column
he's
just
sort
of
breaking
it
down
into
like
these.
This.
These
networks
markers,
are
talking
about
these
communications
between
the
levels
in
a
column
and
then
there's
voting
across
columns
where
we
would
say.
Oh,
you
have
a
network,
you
have
a
large
neural
network
and
there's
some
sort
of
operation
that
goes
in
one
layer,
and
then
we
go
up
to
the
next
one.
There's
some
sort
of
operation
goes
on
there.
I
mean
it's
just
instead
of
parsing
it
differently.
A
E
D
E
D
A
So,
coming
back,
I
want
to
just
once
again
bring
up
the
islands
point.
Take
that
hierarchical
view
say
that
maybe
this
one
that
with
the
blue
and
the
pink
is
v1
the
what
this
model
does
by
it
does
intentionally
as
it
has
like.
Your
visual
field
represent
different
islands
of
so-called
objects,
and
that's
we
don't
explicitly
not
do
that.
But
more
often
when.
B
E
But
that
is
that's
typically,
that's
a
traditional
hierarchical
view
of
visual
processing.
This
is,
you
will
get
these
islands
even
in
a
traditional
feed
forward
hierarchy
without
lateral
connections.
You
know
just
because
there's
convergence,
you'll
you'll
literally
get
these
these
types
of
levels
and
you'll
get
one
part
of
it.
Talking
about
the
boat,
the
other
part
of
it
kind
of
representing
the
forest
or
whatever
the
you
know,
the
mountains
or
whatever
you'll
you'll,
get
these
quote-unquote
islands,
and
I
think,
marcus
you
you're.
He
has
a
more
probably.
E
F
Here,
but
isn't
there
a
difference
here
that
the
post
convolution?
Now
you
have
these
columns
at
each
patch
of
the
image
so
yeah?
You
have
object
level
embeddings
for
each
pose
that
pauses
on
for
each
patch
of
the
image.
I
don't
know
how
he
calls
that
yeah.
Isn't
this
different
than
convolution
in
the
traditional
I.
E
F
E
A
A
E
A
E
D
E
D
My
point
yeah,
I
think
I
think
I
would
agree
with
that.
This
is
not
like
a
biological
protocol
and
it's
I
mean
marcus.
You
point
out.
There
is
some,
you
know,
scale
difference
within
up
and
lower
layers,
but
I
have
a
totally
different
interpretation
for
that
so
to
I
think
what
I
would
say
yeah
this
is.
This
is
not
like
our
cortical
columns
at
all.
This
is
another
thing.
He
calls
an
abstract
concept
and
calls
it
a
column.
D
A
So
my
reason
for
coming
back
to
this
slide
was,
I
wanted
to
bring
up
voting
horizontally
and
say,
and
I
just
the
discussion
topic
here-
is
in
thousand
brains.
Do
you
see
it
as
having
these
islands,
or
do
you
see
it
as
everything
voting
on
one
thing,
as
you
look
around
as
you
look
at
an
object,
does
your
entire
sheet
of
your
entire
sense.
D
The
way
I've
been
thinking
about
it
is
that
that
you
with
each
attention
you
narrow
down
your
attentional
field
until
you
are
able
to
recognize
it
as
one
thing
right,
and
so
I
could
may
say,
oh,
I
can
recognize
this
bicycle
as
one
thing
it's
occupying
this
percent
of
my
visual
field
and
that's
good
enough.
If
I
didn't
know
what
a
bicycle
was,
then
I
wouldn't
be
able
to
recognize
it
as
one
thing
and
then
I
would
narrow
my
attention
down
to
oh
there's
a
wheel
and
there's
a
pedal
and
there's
a
handlebar.
D
I
know
those
things
are,
and
so
it's
like
you
only
you
vote
on
one
everything
is
voting
at
one
things
at
a
time
and
you're
doing
it
sequentially
the
extent
of
the
receptive
field.
That's
involved
changes
depending
on
what
you
can
recognize
as
one
thing,
but
to
me
it's
not
you're,
not
building
these
things
up
simultaneously.
D
D
My
current
sense
there
is
some
weirdness
about
it
because,
like
what,
if
you
know,
we
talked
about
v1
what?
If
what?
If
the
thing
you're
looking
at,
is
so
large
scale
that
v1
really
can't
contribute
as
much
as
we'd
like
you
know,
we
do
know
that
the
retina
projects,
the
v2
into
v4
as
well,
and
so
what
is
v1
doing?
If
I'm
looking
at
something
very
large
that
that
is
suitable
for
v2
and
before,
but
not
you
know
like
what
does
it
be?
What
is
the
one
processing
at
that
point?
Is
it?
D
F
Can
I
have
a
quick,
more
machine
learning
question,
switching
the
gears
a
little
bit,
so
you
mentioned
recurred
connections
between
the
ladders
at
some
point,
so
this
I'm
implying
there
is
some
notion
of
time
step
or
cycles
here,
and
that
is
true
also
for
inference
time.
Is
that
correct,
correct,
yes,
okay
and
then
do
you
back
propagate
through
time
as
well
propagate
through
the
whole
recurrent
connections?
Yes,
okay!
All
right.
A
If
you
fill
in
these
12
numbers,
you
can
represent
any
pose
and
in
fact
or
or
you
can
reduce
that
down
to
six,
because
they're
really
only
six
dimensions
of
pose
in
3d
and
so
so
capsules
kind
of
abstracted
away
the
neural
hardware
and
just
allowed
these
so-called
cells.
These
units,
these
numbers,
to
be
trained
by
backprop
and
to
just
do
matrix,
multiplications
and
glom
interestingly,
gets
away
from
that
it.
A
A
If
you're
representing
location,
there
are
two
ways
to
do
this:
the
capsule
way
of
doing
representing
a
2d
location
would
be
you
have
an
x
cell
and
a
y
cell
siri,
you
have
an
xl
and
a
y
cell
and
in
the
glom
way,
is
more
of
like
the
2d
attractor
sheet
way.
Where
you
have
a
bunch
of
sheet
of
cells
that
just
represent
different
2d
locations,
they
they've
moved
in
this
direction,
so
the
general
the
general
view
of
it
is
you
take
all
your
degrees
of
freedom?
A
E
Yeah,
it's
interesting
in
the
original
capsules
talk
that
hinton
did.
He
actually
talked
about
the
right-hand
way
of
doing
it
and
it's
it
comes
from
an
old
technique.
Called
huff
transforms
where
which
are
basically
exactly
that
you
represent
points
in
configuration
space
and
then
you,
you
accumulate
votes
for
the
different
areas
of
configuration
space,
and
he
he'd
mentioned
that
in
his
first
talk.
So
I
think
the
original
implementation
of
capsules
had
the
thing
on
the
left,
because
I
couldn't
figure
out
exactly
how
to
do
the
half
transform
way
on
the
right.
E
C
A
Basically,
a
union
of
two
possible
locations
and
orientations,
for
example,
yeah
or
like
a
location,
on
a
specific
object
versus
a
location
on
another
specific
object
and
again
this
is
similar
to
how
an
attractor
sheet
can
represent
multiple
locations
simultaneously,
whereas
x
y,
looking
the
coordinates
can't,
but
as
subite
just
mentioned,
the
downside
is
coordinate.
Transformations
are
now
more
difficult
in
the
system.
A
We
face
the
same
issue
when
you,
when
you
represent
these
axes
directly,
it's
just
a
set
of
scalar,
multiplications
or
dot
products
to
perform
a
coordinate
transformation
when
you
do
it
in
more
of
this
population,
distributed
throughout
the
space
of
the
space
of
all
of
these
axes.
It's
now
more
of
a
like
brute
force,
it's
harder.
It's
it's,
no
longer
a
trivial,
a
trivial
cheap
operation
and
so
they're
gonna
he's
gonna
face
the
same
issue
and
he
disclaims
that
in
the
paper
that
that
is
a
thing.
That's
going
to
be
harder.
A
Well,
it's
it's
the
thing
that
he
that
he
explicitly
wants
to
work
like
it's.
It's
gonna,
be
this
feeding
into
a
multi-layer
perceptron
of
some
kind
into
the
next
level,
and
that
that
multi-layer
perceptron
needs
to
perform
a
coordinate
transformation
on
this
somehow
and
part
of
it's
gonna
happen
through
like
here.
E
D
D
So
as
long
as
you
as
long
as
you're
doing
the
transformations
via
movement
or
movement
vectors
involved,
I
think
you
could
have
the
best
of
both
worlds.
I
think
you
can
have
like
the
easy
transformation
according
to
transformation
and
you
still
get
the
the
internal
representation.
That
is
not
like
that.
You
can
get
unions
and
things
like
that.
I
I'm
sorry,
I'm
not
saying
it
very
well,
but
I
think
you
can
have
the
best
of
both
worlds
and
I
think
the
the
method
that
I've
been
proposing
to
the
thomas
would
let
you
do
that.
E
I
mean
one
possible
thing
here.
What
I
don't
know,
if
you
mentioned
in
this
paper,
is
you
could
take
the
point
of
configuration
space
that
has
the
most
votes
attend
to
that,
and
at
that
point
you
can
do
all
your
transformations
with
respect
to
that
point,
you
know
like
if
you
want
to
move
to
a
particular
point.
You
can
just
pick
the
point
here
which
has
the
most
votes
and
then
now
you
can
do
whatever
you
need
to
now.
You
can
do
the
previous
method.
D
Isn't
the
point
of
the
transformations
are
like
you,
have
these
black
circles
here
and
what
you
want
to
do
is
you
want
to
just
like
say
you
know,
just
transform
all
imagine
if
the
reference
frame,
rotated
or
or
something
like
that,
and
you
know
you
want
to
know
where
they
are
in
the
new
reference
frame.
Is
that
right?
That's
what
we're
talking
about
here
right,
we're
new
transformations.
D
D
I
think
it's
as
long
as
you
that
the
the
transformations,
if
you
think
about
if
the
transformations
are
going
through
the
thalamus,
you
basically
do
the
transformations
in
the
1d
vectors
of
space
and
and
then
the
blobs
would
follow
I'll
just
have
to
leave
it
that
I
think
you
can
make
this
all
work
even
without
attention.
I
think
I
think
I'd
have
to
think
about
it,
because
bugs
means
that
you
could
have
both
best
of
those
rules.
E
D
D
Yeah
again
you're
doing
the
transformations
in
these
1d
things.
It's
just
like
you
know
that
think
about
the
the
voltage
control
of
oscillators
and
the
grid
cells
they
represent.
You
know
movement
in
one
direction,
so
then
we,
but
then
you
translate
it
into
you
know
later
you
end
up
with
you
know,
with
good
cells
and
place
cells,
I
mean
I
I
want
to
show
up,
because
I
don't
haven't
go
through
this
enough.
I
think
you
can
it's
I'll,
just
say
it
again.
A
D
I
thought
that
was
one
of
the
weaker
aspects
of
the
paper
where
he
just
kind
of
relied
on.
You
know
classic
neural
networks
to
solve
all
these
different
problems.
Yeah,
let's
take
another
one
here
or
stick
another
one
here
you
haven't
shown
the
diagram.
Yet
with
this
you
know
this
time
diagram
with
the
three
levels.
D
You
know
every
one
of
these
he
throws
in
like
every
every
time.
Anything
has
to
happen.
There's
a
you
know,
oh
sticking,
a
multi-layer
neural
network
here
and
like
this
that
ends
up
with
hundreds
or
thousands
of
them.
You
know
all
over
the
place,
but
I
cuz
I
we
know
the
brain
doesn't
do
that.
The
brain
has
specific
mechanisms.
Good
cells
are
a
specific
mechanism.
You
know
they
work
on
a
very
specific
terms
and
head
direction,
cells
and
so
on.
So
it's
you
know,
I
think
he's
he.
D
D
D
E
E
G
D
Yeah,
that's
true,
that's
true,
I
mean
who
knows
I
mean
the
thomas
has
to
learn
to
do
these
things
too,
and
it's
a
very
simple
learning,
and
so
I
guess,
if
you,
if
you
reduce
the
problem
the
sufficiently
and
the
neural
networks,
get
pretty
damn
simple
anyway,
but
I
think
it's
a
general
philosophy
that
that
hinton
was
taking
as
we're
going
to
use.
You
know
classic
neural
networks
for
all
these
things
and
we'll
use
them
liberally
and
leave
that
to
the
to
the
reader
to
decide
exactly
how
that's
going
to
be
done.
A
F
A
A
F
F
E
D
H
D
D
Yeah
at
least
yeah
even
toy
versions,
but
but
you
know
yeah,
as
you
all
know,
I
mean
we
like
to
think
about
these
things
purely
abstract
things.
We
could
we
kind
of
know
they're
going
to
work
before
we
implement
them.
So
you
know
so
it's
like
it's
all
like
well
is
is
an
idea.
Well,
if
we
can
do
it
like,
I
would
kind
of
think
it
through
before
you
know
at
least
a
certain
amount
make
sure
people
don't
work.
D
F
C
C
A
So
the
last
point
I
wanted
to
make
on
this
whole
degrees
of
freedom
thing
is
so
we
usually
think
of
these
as
location
orientation
and
all
those
degrees
of
freedom
or
an
entry
rhino
cortex.
We
think
of
head
direction,
cells
and
grid
cells
and
border
cells.
Maybe
a
couple
others.
I
one
way
where
hinton's
takes
this
even
further,
where
glom
takes
this
even
further,
is
it
blurs
everything
together
even
object,
identity?
A
Everything
is
kind
of
conjunctively
encoded,
though
the
network
may
choose
to
separate
them
by
default,
it's
all
kind
of
conjunctively
encoded.
So
I
call
this
out
in
a
dedicated
slide,
one
difference
from
our
model,
that's
kind
of
worth
consideration,
because
it
reminds
me
of
a
little
bit
of
biology.
A
Make
is
that
the
the
literature
on
grid
cells
head
direction
cells
all
of
that
often
kind
of
paints.
This
cartoon
world,
where
everything
is
these
separate
populations
and
the
truth,
resembles
this
a
little
bit
more
than
the
cartoon
version.
I
call
it
this
one
paper
that
makes
the
point
that
everything
kind
of
exists
on
a
continuum,
and
there
are
the
many
cells
that
are
kind
of
mixes
of
of
all
of
these
cells,
and
so
I
just
I
this
was
worth
calling
out
to
me
in
some
ways.
D
Yeah
I,
although
I
mean
two
ways
you
can
look
at
that.
I
mean
you
know
attention
like
when
you
think
about:
oh,
let's
think
of
classic
pure
grid
cells
in
classic
pure
head
direction,
cells
and
then,
as
we
think,
about
how
they're
created
and
we
think
about,
like
the
oscillatory
interference
models.
You'll
end
up
with
these
conjunctive
cells,
they
don't
that
are
part
of
the
creation
of
the
pure
grid
cells
and
part
of
the
creation
of
the
pure
orientation
cells.
D
So
you
could
say:
oh
that's
one
way
of
explaining
it.
We
actually
are
creating
true
grid
cells
and
these
conjunctive
things
are
just
the
components
that
naturally
occur.
When
you
use
the
methods
of
oscillatory
interference
to
create
them,
or
you
could
say
no,
they
really
aren't
pure
anything
and
it's
all
a
mixture,
I'm
more
of
a
former.
D
It
seems
more
structured
and
more
believable,
and-
and
you
know-
but
I
see
your
point
but
again,
I
I
think
it's
too
easy
an
excuse
to
say:
oh,
we
don't
need
to
worry
about
any
of
those
details.
You
know
how
good
cells
come
about
because
we'll
just
do
this
and
they'll
something
will
be
gritty
and
some
things
will
be
orientation,
some
things
and
so
on.
So
I'm
I'm
not
convinced
by
this
argument.
I'm
open
to
it,
but
I'm
not
going
to,
but.
A
Good
fair
point:
okay:
this
brings
me
to
part
five,
my
final
part
on
movement
and
but
and
well.
What
I'll
talk
about
here
is
that
movement
partially
works,
because
this
has
a
location
of
the
sensor
representation,
but
it
doesn't
have
path.
Integration
built
in
there
is
kind
of
an
obvious
extensibility
point,
so
here
I'm
probably
going
to
make
the
statement
that
with
glam
movement
is
low-hanging
fruit,
but
it's
not
something
they've
focused
on
so
what
first
I'll
say
like
what
partially
works.
So
here's
jeff
mentioned
this
figure
a
little
bit
earlier.
A
But
the
other
one
you're
right,
you're
correct
that
there
is
another
figure,
and
I
guessed
wrong
so
here
time,
knowing
this
audience
visually
parsing.
This
is
going
to
be
hard
at
first,
but
it's
pretty
simple
figure.
It's
just
time
is
moving
from
bottom
to
top
we're
looking
at
the
same
population
of
cells
across
time
and
and
this
is
showing
that
there
is
both
bottom-up
input
from
level
l
minus
one
to
l
and
there's
top
down
input
from
the
above
level
to
the
lower
level.
A
It's
just
a
confusing
image
to
see
arrows
pointing
up
saying
top
down,
but
but
that's
what
this
just
just
helping
you
parse
this
image
visually
a
little
bit,
so
it
has
this
location
of
the
sensor
relative
to
a
shared
reference
frame.
Typically
in
the
text
here,
that's
just
simply
where
on
the
on
the
retina,
this
pixel
is
but
but
it
really
can
generalize
to
anything
it
could
be.
A
You
have
to
operate
under
one
constraint:
prediction
with
movement
will
work
as
as,
if
the
top
level
the
highest
of
these
levels,
the
scene,
one
is
unchanged
by
the
movement
if,
if
your
scene
level,
so
so,
for
example,
if
if,
if
your,
if
this
is
a
multi-sensor
array
with
vision
and
touch-
and
you
move
your
hands
around,
I
would
say
this
kind
of
does
support
path.
It
does
support
movement
in
that
sense.
If,
if
it's
all,
if
you're,
if
it's
like
vision
and
touch
and
grabbing
things,.
D
A
I
would
say:
yeah,
it's
gonna
infer
across
time.
It's
it's
gonna,
it's
gonna
be
able
to
more
quickly
recognize
that
things
represent
the
things
that
are
grasped,
you're,
asking
a
valid
question.
Now
I
am,
let's
see
I
am
putting
I'm
showing
that
this
almost
works
as
okay,
but
it's
not
a
focus
of
the
page.
D
D
I
got
it
but,
as
you
point
out,
there's
no
path
integration,
so
there's
no
there's
no
knowledge
at
all
about
where
your
sensor
is
going
to
be
after
you
move
it's
more
like
it's
like
your
eyes,
are
doing
random,
cycads
and,
and
nobody
knew
about
it.
No
part
of
the
brain
knew
it's
where
your
eyes
were
moving.
A
C
D
Yeah
and
about
path
integration,
you
don't
really
know
where
you're
going
to
be.
If
you
like,
be
like
walking
around
your
house
and
every
time
you
open
your
eyes,
you'd
be
in
a
new
room,
but
you
had
no,
how
you
didn't
know
how
you
got
to
that
room.
You
know
you're,
just
like
you
know,
I
don't
know
I'm
in
the
bathroom
now.
Oh
I'm
in
the
kitchen.
Now
you
know
as
opposed
to
oh
I
went
down
the
hallway,
and
now
I
know
I'm
going
to
be.
D
D
F
A
F
D
Talked
so
we
share
with
that
the
idea
of
hard
work
hierarchies
the
whole
idea
that
there's
compositional
structure
and
you're
learning
a
model
of
the
world.
Where
things
are,
you
know
the
relative
poses
of
objects.
You
know
the
displacement
zone.
We
focus
on
that
too.
We're
doing
it
through
movement
and
attending
the
one
object
at
a
time
and
building
this
up
sequentially,
which
is
clearly
what
brains
do
but
he's
trying
to
solve
it
in
a
flash
as
marcus.
D
Maybe
yeah,
but
you
know
it's
interesting,
it's
a
different
approach.
I
mean
personally,
I
would
never
have
written
that
paper.
I
would
never
said.
I
always
felt
that
the
flash
inference
problem
is
such
a
red
herring.
It's
going
to
lead
you
down
the
wrong
path
of
thought,
and
so
I've
always
personally
rejected
that.
But
but
that
doesn't
mean
that's
right.
I
mean
it's
alternate
to
try
it
the
other
way.
So
you
know
we've
struggled
around.
D
We
struggled
not
understanding
how
this
stuff
worked
and
we
started
off
with
simple
things,
like
sequence
memory,
which
is
a
long
way
from
understanding.
You
know
visual
scenes,
but
at
least
he
had
that
temporal
component
to
it
and
relied
on
that
where
he's
starting
from
the
other
end
start
with
flash
inference,
it's
what
most
vision.
Researchers
do
right,
that's
what
vision
research
has
been
about
for
the
last
50
years,
primarily
is
about
flash
inference
and
he's
saying:
how
can
I
get
this
particle
hierarchy?
Thinking
about
it?
D
D
H
An
interesting
it's
an
ongoing
debate
in
the
psychology
literature,
a
lot
of
yeah
research,
I
think
in
the
kind
of
80s
on
which
suggested
that
a
lot
of
things
we
we
only
process
in
like
a
serial
manner
and
then
very
simple
features.
We
can
process
in
a
parallel
manner,
but
I
think
if
you
look
at
the
actual
plots
for
things
like
search
times,
it's
more
of
a
kind
of
a
gray
answer
that,
obviously
we
process
things
better
when
we
direct
our
attention
to
them.
D
Yeah,
no,
that's
true,
but
again
those
things.
If
you
think
about
like
search
time,
you
know
it's.
You
have
to
look
very
carefully
at
the
experiments,
because
some
of
these
things
are
popping
out
these
pop-out
experiences.
You
know
well
the
different
colors
that
doesn't
mean
you're
processing,
multiple
things
at
a
time
or
or
a
different
shape.
Or
you
know,
the
real
question
is:
if
I
show
you,
you
know,
five
objects
on
the
screen
and
I
flash
it
in
front
of
you
and
those
objects
have
no
relationship
to
each
other.
D
H
D
D
D
Anyway,
I
think
it's
it's
an
open
question
and-
and
I
was
hesitant
a
little
bit
when
I
said
to
the
attentional
mechanism-
it's
only
one
thing
at
a
time
because
I
think
at
different
levels
of
the
corporate
hierarchy.
It
gets
more
complicated
and
we
don't
understand
it
well.
So
I
don't
think
we
know
the
final
answer
to
that.
A
So
I'm
nearing
the
end
of
this,
but
on
the
topic
of
path,
integration.
One
thing
I
do
point
out
is
the
pose
is
already
represented
in
here.
It
just
needs
to
be
updated,
and
so
using
movement,
update
using
movement
and
photo
updated
oppose
is
a
doable
thing
I
think
like
like.
Although
movement
is
not
built
into
this,
I
think
it's
probably
intentionally
kept.
You
know
close
at
hand,
it's
it's
it
at
the
moment.
It's
easy
to
add
movement
to
this,
and
so
it's
it's.
D
E
A
A
A
D
What
is
is
another
thing
that
comes
you
know,
one
of
things
you
get
from
having
sort
of
reference
frames
for
storing
your
knowledge
and
grid
cells,
and
so
on.
You
get
the
ability
to
manipulate
objects,
so
I
could
you
know
I
could
I
can
see
an
object
from
one
orientation
tend
to
its
different
components,
and
I
can
now
visualize
it
from
a
different
orientation.
No
problem
does
anything
like
that.
Does
this
have
any
capabilities
like
that
here.
C
D
D
How,
but
how
is
it
I'm
missing
that
I
mean
here
I
I
what
I'm
imagining
he's
doing
here
is
saying:
here's
a
scene
and
I'm
going
to
tell
you
what's
in
the
scene.
These
are
the
different
components
in
the
scene,
and
this
is
how
they
grow
together
in
a
scene.
But
where
is
the
structure
that
allows
me
to
rotate
the
entire
scene
and
now
know
where
everything's
going
to
be?
A
A
D
It
just
seems
like,
without
a
reference
frame,
it's
almost
impossible
to
do
that.
I
mean
you
can
always
try
to
lower
in
everything,
but
that
doesn't
you
know
that
this
gets
impractical,
so
it
just
I
don't
know
I
don't
I
kind
of
like
I'm
not
saying
convince
myself,
but
I'm
working
on
the
assumption
that
that
there's
no
other
way
to
do
that
without
reference
range.
How
do
I
stretch
something
in
one
dimension?
How
do
I
rotate
it
mentally?
B
I'd
like
to
I'd
like
to
chime
in
here,
so
in
some
of
his
earlier
capsules
work.
I
think
he
mentioned
that
you
know
specific
capsules
were
designated
for
to
recognize
certain
parts,
and
then
there
were
higher
level
capsules
to
recognize
certain
holes,
and
in
that,
when
you,
when
you
have
it
like
that,
you
know
what
the
relationship
is
between
the
part
and
the
hole
so
the
so
the
transformation
between
the
part
and
the
hole
can
be
encoded
into
the
weights.
E
Information
yeah,
I
guess
the
one
question
there
following
on
from
that
is
sort
of
mechanistically.
How
does
that
happen?
If
you,
if
you
reckon,
if
you're
trained,
to
recognize
a
particular
object
at
a
certain
pose,
and
now
you
see
a
completely
novel
view
of
that
at
a
different
pose,
will
you
automatically
be
able
to
recognize
that
object
and
recover
the
pose
right?
That's
the.
E
D
F
A
E
Yeah
so
yeah
I
was
going
to
say
I
mean
the
hope
would
be
that
you
would
not
need
to
train
on
different
poses
right.
You
would
only
treat
if
it's
a
particular,
let's
say
assume
it's
a
rigid
object.
You
train
on
one
pose
and
now
for
a
small
number
of
poses,
and
now
you
can
recognize
it
in
in
very
different
poses,
and
the
thing
that
quran
was
mentioning,
at
least
in
the
original
capsules.
E
I
think
the
idea
was
that
the
relative
locations
of
everything
is
encoded
in
the
weights,
and
so
you
you
don't
you
all
you
need
to
do
is
recover
their
relative.
Okay,
you
don't
you
know
at
a
different
orientation
or
a
different
scale.
You
would
automatically
be
able
to
recognize
that
same
same
object,
but
it's
not
clear
how
that
happens
here
with.
B
Two
here
in
the
I
think
the
argument
is
here
that
information
that's
encoded
in
the
weights
in
the
in
the
in
the
previous
capsules.
Work
is
now
encoded
as
part
of
the
embeddings.
So
that's
used
to
go
from
like
the
capsule
at
level,
the
part
at
level
whole
at
level,
l
plus
one.
I
think
that
that's
the
argument.
E
F
A
That
the
capsules
went
through
this
transition,
where
the
first
paper
did
all
this
explicitly.
It
was
a
train,
yes
and
then
the
later
one,
the
apparent
I
think,
the
one
that
performed
the
best
the
stacked
capsules,
auto
encoders
paper
got
away
from
this,
and
I
I
never
really
wrapped
my
head
around
how
that
satisfied.
The
original
goals
of
capsules-
but
maybe
it
did-
I
didn't
do
my
homework
on
this,
so
so.
A
I'm
also
lost
you've
raised
a
good
question
here
and
I
think
the
reason
this
is
coming
up
now
is
because,
when
the
system's
processing
movement,
it
has
to
solve
the
same
problem
as
to
solve
this
and
that
that's.
Why
we're
talking
about
this
right
now
and
that's
where
movement
comes
into
this,
and
so
I
don't
know
in
some
ways,
maybe
our
a
contribution
or
a
point
we
could
make
here-
is
that,
like
that
problem
gets
solved
once
you
bring
movement
into
it,
so
punting
on
movement
might
not
actually
benefit
you
right
now.
A
Okay,
I'll,
come
to
an
end
pretty
quickly.
Here
I
mentioned
I
mentioned
here
like
strictly.
You
would
only
need
to
path,
integrate
one
level
and
you
could
inherit
it
from
the
rest
or
if
you
have
path
integration
on
every
level,
it
will
improve
sensory
motor
inference
final
slide
here,
I'll
just
go
by
quickly
and
say
you
could
also
rearrange
these
and
put
the
arrow
elsewhere.
A
A
Frame
where,
if
you
as
you
move
through
the
scene,
the
scene
level
vector,
doesn't
change
if
it's
actually
like
allocentric,
then
path
integration
would
happen
in
this
box,
and
this
would
be
like
a
nice
unifying
theory
where
moving
moving
always
updates
this
box.
I
don't
expect
all
of
you
to
follow
that
right
now,
where
we're
going
along
so
I'll
just
go
to
my
summary
slide.
A
Unless
someone
else
wants
to
interrupt
me
here
so
summary,
I
would
say
that
the
current
thousand
brains,
theory
models
and
glom
use
essentially
the
same
building
blocks
or
at
least
a
lot
of
the
same
building
blocks.
The
voting,
part-to-whole
reference
frame,
transformations
and
newly
like
neural
representations
that
can
handle
uncertainty
or
ambiguity.
A
Sensor,
there's
going
to
be
displacement,
cells
and,
while
glam
just
kind
of
lets
a
vector
of
cells,
learn
whatever
it
wants,
may
learn
these
it
may
not.
They
may
find
a
different
way
of
doing
it.
We
emphasize
movement,
glom
doesn't,
but
it's.
I
still
see
it
as
kind
of
low-hanging
fruit
and
maybe
it'd
be
worth
them
focusing
on
it.
Now,
for
the
training
reasons
we
talked
about
and
then
glom
emphasizes
multi-level
part,
hole
hierarchies
and
islands
of
agreement.
A
D
Yeah
and
I'm
I'm
hoping
you're
right,
but
I
I'm
not
convinced
about
movements
of
low-hanging
fruit.
I
think
sort
of
an
impoverished
version
of
the
movement
might
be,
but
going
back
to
that,
you
know
the
rotating
of
an
object
in
3d.
It's
part
of
movement.
In
my
mind,
it's
not
clear
how
you
do
that
at
all
here,
but
but
I
think
it's
interesting.
You
know.
I
think
it
is
interesting.
D
F
A
Question
really
what
what
just
happened
is
I
I
worked.
I
worked
to
prepare
this
presentation
and
I'm
gonna
sit
with
it
for
a
little
while
after
this,
and
that's
my
that
is
my
real
answer
is
I
haven't,
sat
with
it
for
for
some
time.
Yeah
I,
I
guess
just
one.
I
do
think
that
hinton
is
definitely
coming
in
with
a
with
a
strong
psychology
background
and
has
done
many
experiments
just
trying
to
interrogate
what
the
visual
system
is
doing
from
like
an
experimental
outsider
standpoint,
but
yeah.