►
From YouTube: On the Origins of Spatial Pooling & Temporal Memory
Description
With Numenta founder Jeff Hawkins.
Aug 14 Numenta Research Meeting
B
All
right
should
we
go
yes
all
right.
So
this
this
is
very
informal.
This
is
the
continuation
of
what
I
was
asked
to
do.
Do
some
neuroscience
talks
to
basics,
so
we
talked
about
the
neuron
before
we've
talked
about
columns
and
many
columns.
I
was
asked
to
talk
about
pooling,
especially
temporal,
pooling
and
also
something
about
hierarchy.
So
I
thought
I'd
combine
those
two
today,
but
anything
you
want
to
ask
about.
Does
they
ask
and
we
can
divert
there's?
No,
it's
not
like
a
highly
structured
thing.
B
B
It
may
have
a
definition
in
machine
learning,
I'm
not
aware
of
so
I'm
just
gonna
talk
about
how
I
think
about
the
word
pooling
and
what
it
means
to
me
and
where
the
thoughts
about
it
come
from,
and
it
may
be
consistent
with
other
people
thinking
it
may
not
now
that
the
term
is
used
elsewhere,
very
heavily
and
everyone
uses
it
the
same
way.
So,
to
me,
pooling
is
a
very
simple
idea.
It's
it's
a
many
to
one
mapping.
B
It's
like
many
patterns
map
to
fewer
patterns,
and
you
know
the
brain
has
to
be
doing
cooling.
You
can
simply
think
about
this
way.
What,
if
I,
give
a
a
million
images
and
I'm
classifying
them
into
ten
categories
when
I'm
doing
this,
some
sort
of
neural
network
and
so
I
could
put
in
a
million
different
patterns
and
I'm
gonna
get
one
of
ten
outputs
of
this
set.
So
that
is
a
pooling
operation
in
some
sense
in
the
brain.
I
would
have
at
some
level
in
the
brain.
B
I'd
have
neurons
that
represent
the
classification
of
that
image
and
those,
and
that
would
be
a
pattern
and
there's
many
millions
or
hundreds
of
thousands
of
patterns
which
map
into
that
and
so
along
the
way
you
have
to
have
changing
patterns
and
a
stable
pattern,
changing
patterns
in
a
stable
pattern
and
that's
the
concept
of
pooling
it's
a
many
to
one
mapping.
I,
don't
know
if
that's
consistent
with
everyone
else's
understanding
upon
that
most
people
use
it.
B
So
clearly,
it's
happening
in
the
brain
big
time,
even
something
like
I,
just
look
at
a
scene
or
look
at
an
image.
Your
eyes
are
moving
constantly.
You
get
the
images
you
perception,
the
image
is
stable,
and
that
also
requires
some
sort
of
pooling,
because
I
was
there's
some
stable,
neurons,
representing
what
I'm
looking
at
and
the
inputs
are
changing
underneath
it.
So
there's
a
many
more
mapping
and
this
many
new
I'm
mapping,
it's
not
a
random
thing.
B
It's
very
precise
thing:
it's
not
like
I'm,
just
mapping
willy-nilly
to
something
these
are
very
precise
patterns
and
many
precise
patterns.
So
when
you
think
about
pooling
is
two
basic
ways
you
can
think
about
boiling
and
the
terms
we
use
to
that
of
spatial
pooling
and
temple
pooling,
and
there
are
pretty
simple
and
concept.
Spatial
pooling
is
the
idea
that,
let's
say
I
have
some
sparse
pattern,
some
STR
here
and
I'm
another
pattern.
B
B
If
I
just
change
one
of
these
bits,
I
generally
don't
expect
this
output
to
to
change.
So
this
you
might
think
of
that.
It's
noise,
you
might
think
of
this
like
variation,
but
if
I
have
something
a
thousand
bits
of
a
thousand
ones,
if
I
change
some
of
them
I
don't
expect
this
to
output
the
change
every
time,
I
change,
one
of
these
bits,
and
so
that
was
for
pooling.
It's
essentially
saying:
hey,
maybe
my
mapping
this
pattern
in
a
pattern,
but
any
small
variation.
B
If
you
think
of
this
pattern
as
a
point
in
a
high
dimensional
space,
if
I'm
near
by
that
point,
all
my
all
my
ones
haven't
changed
on
your
sub
small
steps
up,
my
ones
have
changed
about
a
thousand
one
hundred
of
the
100
could
change
I'm
just
moving
locally
in
that
point
in
high
dimensional
space
I
still
get
the
same
output.
So
that's
why
the
term
spatial
pooling
came
about
your
spatially
close
to
the
rich.
B
At
some
point,
these
a
bunch
of
points
that
are
spatially
near
each
other
in
this
high
dimensional
space
are
all
gonna
mount
to
the
same
point.
It
doesn't
require
any
learning
to
do
that.
You
can.
You
can
do
this
with
a
random
map.
You
get
in
there
bonds
up
here
that
randomly
get
input
from
some
subset
down
here
and
as
long
it,
and
then
you
would
enforce
a
varsity
function
on
this
say:
there's
a
competition
going
on
to
who
gets
the
win
in
place,
and
you
do
that
randomly
you'll
get
facial
pulling.
B
B
There
we
did
a
lot
of
work
on
training
the
spaceship
cooler
and
you
can
get
better
results
when
you
do
that,
meaning
these
these,
these
synapses
could
be
to
learn
and
say,
I
won't
go
into
that.
But
that's
basically,
spatial
pooling
is
essentially
just
a
very
simple
mapping
using
their
arms
from
one
time
to
another
pattern
and
naturally
there's
a
lot
of
lot
of
patterns
that
are
close
to
this
original
one.
That
would
map
onto
the
same
one.
So
that's
a
form
of
cool
all
right.
B
Here's
str1
and
str2
and
str3
and-
and
these
are
completely
orthogonal
as
much
as
any
yes,
they
are
orthogonal,
orthogonal,
vectors
and
and
they're
all
gonna
map
to
the
same
one
up
here
right
and
so
one
way
that
the
brain
does
is
to
say
well
because
there
are
temporal
proximity
this
one.
This
one
follows
out
this
in
the
follows
that
one
we
can
assume
that
it's
likely
there's
a
common
cause
in
the
world
instead
of
these
three
of
the
patterns
that
follow
each
other.
B
A
time
actually
do
all
represent
the
same
thing,
so
you
can
think
like
no
to
normality
right
well,
the
right
after
one
of
the
sleep
probably
represent
the
same
thing,
and
so
the
same
thing
would
be.
The
basic
same
thing
would
be
true
if
I'm
trying
to
touch
the
coffee,
copper
or
move
my
eyes
over
some
object,
and
here
basically
assuming
that
there's
a
temporal
correlation
corresponds
to
a
common
cause,
and
that's
where
the
wrote
mo
pooling
come
from.
B
B
B
It's
still
a
one
population
to
another
population,
but
we're
going
through
different
patterns
down
here,
and
they
have
to
learn
them
up
here.
So
that's
the
concept.
Typically
there
they
are
bound
together
because
of
temporal
proximity.
That
makes
sense.
The
next
thing
is,
we
can
break
it
down
into
two
different
parts:
higher-order
sequences
and
sensory
motor
sequences.
So
basically
these
could
be
like
notes
in
a
melody
and
that's
all
you
need
to
know.
C
B
B
B
B
D
B
C
B
B
F
F
B
F
B
B
The
point
in
four
point
years:
you've
got
these
changing
inputs
coming
into
the
part
and
they
couldn't
it
because
we're
just
the
world
is
changing
like
securing
the
notes,
invalidate
or
a
bird
flies
by
or
they
could
be
changing
because
you're
moving
and
and
not
when
you
move,
they
push
changed
so
I'm
moving,
my
fingers
of
the
coffee,
remove
them
and
I
both
case.
You
have
a
changing
input.
In
both
cases,
we
have
to
come
up
with
representations
for
that
input
that
are
unique
to
the
specific
object.
B
B
Need
to
I
need
to
make
it
unique
to
this
objects
in
word,
and
that's
the
whole
good
cell,
that
Pappas's
to
that
I'll.
Come
back
to
the
moment
here
that
you
have
to
convert
these
into
very
unique
for
its
pattern
and
then
you
can
pull
them
together
to
represent
either
the
melody
or
the
coffee
cup
or
something
else.
F
B
Something
you
could
explain
that
to
me
more
in
detail,
not
right
now,
because
I'd
like
to
understand
that
to
mean
this
was
a
guy
being
fundamental
observation
about
bringing
in
and
say
both
my
mind
was
back.
You
know
over
30
years,
like
the
brain
has
to
figure
out
how
to
take
all
these
changing
patterns
and
bring
in
two
unique,
stable
ones.
Yeah.
A
B
B
G
E
B
Representations
of
anything
good
to
know
that
these
changing
inputs
of
my
finger
are
coffee
time.
This
all
requires
cooling
into
:.
So
now
now
we're
going
to
turn
our
attention
to
the
cortex
okay.
So
we
all
know
that
the
cortex
is
a
big
sheet
right
and
it's
divided
into
regions,
but
the
reason
to
all
look
very
similar.
B
G
C
B
B
And
so
by
the
magic
you
slice
even
through
it,
and
so
these
reasonably
priced
similar.
But
if
you
follow
the
information
flow,
this
is
a
classic
view
going
back
to
the
60s,
you
might
say:
oh
the
information
comes
into
the
eye
and
then
you
million
fire
is
coming
off
the
back
of
the
other.
They
make
this
way.
They
stop
here
and
then
Thomas
that
this
relay
sells
on
Thomas
and
then
they
be
jumped
to
the
first
visual
region
of
the
cortex.
B
B
C
B
B
B
C
G
B
This
is
a
participative
terminate
primarily
later
for
dr.
animo,
their
layer
for
projects
the
layer,
two
three,
so
that's
something
that's
going
on
there
and
then
we
are
two
three
is
one
of
the
output
layers
of
the
cortex
or
the
cells
language,
reading
exit
coming
down
and
go
on
to
productive
layer
for
the
next
week.
So
this
exists,
but
this
is
a
subset
of
what
not
to
do
or
not.
So
this
makes
is
pretty
look
pretty
simple:
every
regions
got
an
input,
there's
basically
two
layers
of
cells
involved.
B
E
F
B
So
that's
great
and
then
what
we
did
in
perhaps
maybe
the
compass
remember
to
do
this.
Is
we
we
think
of
the
mechanism
that
said
alright,
we
could
model
in
just
similar
this
we
could
say
just
break
up
ever
one
layer.
Then
I
can
put
these
many
columns
in
there
and
and
we
learn
higher
order,
sequences
and
and
what
we
didn't
do
for
a
long
time.
We
actually
didn't
do
the
tempo
pooling
layer.
You
know
the
tempo
pooling
operation
is
not
necessary
to
learn
two
sequences.
B
So
we
create
a
chapel
pooling
layer
here
and
a
sequence
memory
layer
down
here
and
what
you
would
end
up
was
you'd
end
up
with
a
stable
representation
for
sequence
and
a
new
sequence
come
to
any
organization
for
the
sequences
name
of
melody
in
the
object
and
those
two
operations
could
occur
and
then
you
can
do
it
again,
but
it
wasn't
really
clear
why
you
do
it
again
and
what's
and
so
on,
there's
lot
more
to
it.
But
anyway
that
was
the
basic
idea.
B
Someone
could
argue
that
perhaps
layer,
2
3
is
a
temple
pooling
letter
now
again,
I
say
layer,
2,
3,
many
many
early
papers
don't
make
car
papers,
don't
make
a
distinction
between
layers,
2,
&,
3
layer,
1
actually
has
quite
a
few
cells
that
they
sort
of
ignore
it.
It's
more.
There
are
cells,
but
some
people
who
can
differentiate
all
kinds
of
cells
in
layer,
2
and
3.
So
it's
maybe
not
this
one
type
of
South.
B
It's
like
lots
of
cells
merits,
and
so
you
know,
there's
a
layer
of
3a
and
3b
and
2a
and
2b,
depending
who
you
ask,
but
very
often
people
say
well.
It
looks
like
there's
a
lot
of
connections
to
layer
3
those
become
that
put
the
net
reads
in
a
song
that
was,
that
was
a
classic
who
won
bezel
or
back
to
the
60s
feet,
but
we're
pathway.
There
is
a
also
a
classic.
B
In
this
view,
there's
a
feedback
pathway
and
the
feedback
pathway
starts
with
cells
in
layer
6
and
they
project
back
through
the
white
matter
into
the
previous
layer,
a
previous
region,
so
I'm
not
showing
I'm.
Here
with
me,
a
backward
production
and
in
the
original
classic
do
what
these
axons
do.
Is
they
spread
very
long
distances
in
layer
one?
So
it's
a
feedback
and
two
cell
types
that
connect
to
it
would
be
the
layer,
five
and
layer
3
these
cells.
These
are
parental
cells,
which
have
these
long
apical
dendrites
extend
up
into
this
layer.
B
H
B
B
Stacked,
so
you
know
you:
can
there
are
50
when
you
know
what
you
look
at
the
hallway
at
the
top
of
the
stairs?
That's
a
picture
from
cahal
than
1900
and
that's
a
picture
that
Kahala
fall
through
the
microscope
and
he
said,
look
there
seems
to
be
layers
here
and
why
they
see
them
because
the
cells,
some
of
them
cells,
a
bigger
some
of
them,
are
small
or
somewhat
densely
packed
some
I'm,
not
densely
packed
or
not.
The
first
thing
they
did.
They
just
looked
at
me.
B
They
looked
at
these
stains
of
the
cortex
and
they
could
say.
Look
look
at
those
stains
have
looked
like
these
letters,
these
letters
players.
Nowadays
you
can
look
at
the
connectivity
and
the
connectivity
implies
a
different
set
of
layers
or
an
additional
sets
players
or
they
can
look
at
gene
expressions
and
so
on.
So
there's
many
many
different
ways.
You
can
determine
what's
a
unique
cell
type,
but
in
theory
you
would
say:
there's
a
population
itselves
and
look
similar.
They
project
someplace,
similar
thing
and
input
some
function.
Similarly,
they
express
similar
types
of
chemicals.
B
Therefore,
that's
a
layer
so,
but
there's
got
to
that.
Yet
almost
all
the
complexities
in
the
brain
is
in
here,
and
these
connections
represent
a
small
percentage
of
the
all.
The
connections
actually
exist
in
the
brain,
but
if
we're
talking
about
region
the
region,
so
in
a
convolutional
neural
network,
your
layers
are
very
simple
and
there's
a
hundred
of
them
here.
This
this
region,
which
it's
sort
of
equivalent
to
maybe
a
layer
of
cognition,
is
very
complex
and
if
there's
only
a
few,
that's.
B
F
B
I
mean
there
may
be
in
a
human
right,
there
might
be
a
hundred
regions,
it's
debatable
what
what
some
unique
region
it's
very
hard
to
tell
sometimes
but
let's
say,
there's
100
regions
and
there
might
be
any
region-
might
get
input
from
seven
or
eight
other
regions.
Actually
it
turns
out
to
be
more
than
they
thought,
and
so
the
conductivity
of
the
region,
the
region
graph,
is
close
to
the
40%,
but
that
doesn't
tell
you
how
many
axons
and
synapses
it
just
says.
I
just
look
at
the
continuity
of
the
graph
between
these
hundred
regions.
B
B
Everywhere,
before
pretty
much
and
the
conductivity
sparsity
is
very
difficult
to
put
your
finger
on
because
some
of
these
cells
project
long,
just
luckily
so
sparse
compared
to
what
of
what
population
are
you
looking
at
I?
Think
I?
Don't
find
it
fruitful
think
about
that
connectivity
right
now,
like
I,
can
say
something
concrete
about
this
for
city
in
general,.
C
B
G
B
B
So
that
is
all
true,
but
it's
really
really
not
the
whole
picture,
not
even
close,
but
it
is
instructive,
because
this
these
connections
are
large
and
they
do
exist
and
they
were
proven
and
they're
still
there
and
what
you
can
say
well,
there's
something
going
on,
because
I
get
an
input
left
more
like
more
particular
three
layers
we
took
fillet
for.
That
is
this.
We
have
to
explain
that.
Okay,
there
was
another
feed-forward
pathway
which
was
ignored
for
a
while,
but
it
wasn't
understood
and
two
scientists
we
talked
about
here.
B
A
lot
Sherman
and
Sherman
and
mcgillivray
have
been
promoting
this
idea
and
and
here's
one
way
to
think
about
it-
see
over
here
there's
the
input
from
the
I
oppose
to
the
sequel
to
the
house
before
those
that
we
want
well.
This
is
true
for,
but
from
the
ears,
through
the
skin
and
everywhere
else.
In
fact,
every
every
primary
century
region
it's
an
input
from
the
thalamus,
so
pretty
important.
It's
obviously
something
important,
but
they
pointed
out
was
that's
actually
the
wrong
way.
Think
about
it.
B
Take
a
look
at
the
thalamus
that
was
has
lots
of
these
little
submarines.
In
and
it
turns
out
that,
if
I
follow
a
different
way
to
look
at
this
conductivity
conductivity
is
that
I
get
the
input
from
a
sensory
organ
to
define.
It
goes
to
the
thalamus.
It
then
projects
to
some
region,
the
output
of
this
region,
a
different
output,
not
the
same
one.
B
We
have
over
here
a
different
one
leaves
this
region
goes
to
a
different
relay
Center
in
the
thalamus
which
projects
to
the
next
region
yeah,
but
that
one
was
a
different
relay
center
of
the
thalamus.
We
projected
that
so
this
thing
this
is
the
this
might
be
the
main
feed
forward
path.
The
main
feed
forward
path
is
that
every
cortical
region
gets
its
input
from
the
thalamus
and
it
projects
to
another
region
through
the
thalamus,
and
although,
numerically,
these
are
smaller
number
of
connections
than
these,
they
are
much
stronger
and
much
more
impactful.
B
B
D
B
B
G
B
We
take
a
very
powerful
driver
and
this
is
copacetic
with
the
whole
idea
of
a
common
core
algorithm.
We
could
stay
like
look
I
region,
a
region
of
the
cortex,
doesn't
know
what
it's
looking
at
it
doesn't
they
all
look
the
same
and
they
all
work
on
the
same
principle.
So
if
one
of
them's
going
to
be
driven
by
Salama
chemical
button,
and
and
this
is
how
it's
connected-
these
are
often
called
our
relay
cells,
because
it's
only
one,
it's
almost
yeah.
C
B
It's
almost
like
you
know:
one
fiber
comes
in
here.
It
makes
this
one
cell
fire
and
again
another
fire
here.
It's
like
it
relays
the
circuit,
but
we
know
that's
not
true.
We
do
this
very
complex,
but
it
look
like
that
and
suit
I
did
a
paper
was
published
whose
it
cosine
a
cosine.
We've
spent
a
lot
of
time
speculating
what's
going
on
here.
What
are
these
cells
doing?
What
are
these
relay
still
do?
We
can
talk
about
that,
so
this
is
an
ultimate
view.
It's
not
like
both
of
these
exist.
B
F
C
B
They
don't
they
don't
again,
they
don't
look
like
they're
doing,
processing
it's
not
like
they,
they
will
function,
that's
why
they
call
them
real.
It
excels.
There's
not
enough
these
there's,
not
enough
they're,
all
machine
ER
here
to
look
like
it's
doing
a
lot
like.
Can
you
get
up
here
again
in
terms
of
cell
wall?
Going
all
this
complex
stuff
down
here?
It's
just
like
I
mean
if
it
or
not-
and
we
know
it's
not
true.
B
We
know
that
it's
doing
more,
but
if
you
just
look
at
around
girl
machinery
here,
it's
tiny
compared
to
the
mountain
almond
shaker
here
until
we
know,
but
it
seems
to
be
critical
and
I'll
come
back
to
what
it
might
be
doing
in
a
moment,
but
they
is
a
general
rule.
There
hasn't
been
a
lot
of
like
focus
on
well,
I
should
say
general
rule
like
oh.
The
function
of
these
things
is
X,
it's
more
like
they
look
like
on
this
team.
The
response
properties
are
they
they
they
do
a
lot
of
analysis
of
it.
E
B
These
are
associated
with
attention.
I'll
come
back
to
in
a
second,
by
the
way,
if
you
lose
these
connections,
you're
dead
I
mean
you're,
a
vegetable
yeah,
the
thalamus,
although
it's
small,
it's
like
a
size
of
a
small
bird
egg.
It's
to
a
moment,
son
Brandon,
it's
relatively
quite
small
compared
to
the
neocortex,
it's
you
know,
there's
small
structures
but
they're
so
critical
to
the
processing
information
for
the
brain
that,
if
you
remember
the
woman
in
Florida
who
was
brain
dead,
there
was
a
big
debate
about
letting
her
off
life
support.
B
B
B
Then
the
question
is:
why
are
they
located
here
and
not
up
here
and
it's
clear
that
the
brain
wants
to
put
these
all
very
close
together,
at
least
like
it
wants
to
have
these
relays
are
all
close
together?
So
up
here
remember
he
one
could
be,
could
be
a
big
big
area
right
and-
and
there
are
no
cells-
are
there's
no
cells
in
here
all
the
way
across
he
wanted.
B
So
he
wanted
a
bunch
of
separate
little
things,
we're
not
really
connected
together,
but
here
these
things
are
all
very
close,
and
so
it
gives
it
the
brain
ability
to
do
something
together
all
the
inputs
in
this
at
once,
and
so
we
it's
basically
bringing
the
lower
layer
of
all
these
things
down
to
the
center
area
here,
where
it
can
do
something
on,
and
we
have
some
idea
what
that
is,
but
it's
really
think
about
it.
Even
though
birds
may
have
styles,
every
everything
exists
in
some
form
in
a
previous
animal.
B
B
B
Here's
a
here's,
a
syllabic
relay
this
is
part
of
the
sounds
here.
Okay,
this
is
one
of
me.
This
is
one
of
the
little
guys
here
and
I'm
just
rolling
this
one
right
here:
I'm
such
I'm,
showing
you,
okay,
and
so
this
this
where's
that
that
cells
in
layer,
five,
okay,
so
that
layer,
three,
it's
layer,
5
so
letter.
Five
in
this
case
is
the
feed-forward
output
of
this
region.
Where
later
three
was
the
feed-forward
output
over
there.
This
is
their
font
and
the
projection,
this
relay
cell
and
then
project
back
up
to
layer
4.
B
Primarily
these
all
these
things
to
make
other
connections
primarily,
therefore,
and
I'm
so
layer,
four
again
layer,
3
and
layer
3.
In
addition
to,
in
addition
to
going
forward
layer,
3
sells
projects
become
to
layer
5.
So
if
you
follow
the
activity,
it
goes
from
found
this
to
live
forward
away
a
3d
tool,
a
5
mm
and
then
back
to
another
denser
the
neck,
because
repeat
that
again
to
the
next
video
next
combo.
So
there
are
two
feet
forward.
Pathways
is
one
thing
that
exits
from
layer,
3
and
one
thing
exit
somewhere.
B
B
That's
one
I
showed
over
there
and
that's
repeated
again
and
again
so
two
for
pathways,
1
2001
direct
one
seems
to
be
the
driver
more
important,
one
of
the
things
to
be
more
modulate
or
less
important.
The
textbooks
mostly
talk
about
this
one,
but
this
seems
to
be
the
most
important
one
and
now
to
make
things
really
confusing
these
same
when
they're
5
cells
that
project
to
the
Damas,
which
then
project
to
the
next
region
that
same
axon,
which
is
coming
off
the
cell
splits
into
two
parts.
B
B
That
is
the
with
the
code
of
motor
cortex.
There
are
these
cells
in
layer,
five
that
are
really
big
and
fat
cells
in
layer
five
closely,
they're,
just
big
the
reason
they're
big
is
gonna
have
to
protect
all
the
way
down
to
your
spinal
cord.
So
when
people
talk
about
motor
cortex,
they
said.
Oh,
this
is
motor
cortex
right
here
and
it's
got
these
big
cells
only
or
five
and
then
conjecture
the
spinal
cord.
B
That's
it
that's
the
motor
output
of
the
printer,
but
now
they've
discovered
that
that's
not
really
true
in
terms
of
every
region
seems
to
have
layer.
Five
cells
there
just
like
that
cells
are
just
smaller
and
they
do
the
same
thing
so
in
in
visual
cortex
back
here,
the
cells
in
layer
five.
This
is
like
this
is
a
v1
primary
visual
cortex.
It
has
a
motor
output
in
v2
as
a
motor
output,
the
cells
in
layer,
five
project
two
in
this
case
they
project
to
the
part
of
the
brain
that
controls
on
them.
B
So
it's
not
like
the
information
goes
to
Chunkin
now,
I
have
an
output,
it's
more
like
everybody
is
doing
sentry
put
around
the
sensory
motor
system
everywhere,
and
this
is
a
very
confusing
question.
How
could
the
output
of
a
region
be
both
a
Morgan
man
and
a
feed-forward
coming
into
the
next
week?.
B
B
B
So
I,
don't
think
I
think
in
general
they
say:
there's
some
simple
examples
where
they
say.
Oh
this,
we
even
projection
the
body,
the
bigger
region,
pinups
of
the
superior
colliculus
with
the
division,
one
of
the
regions,
projects
to
focusing
is
also
related
to
movement
related
to
auditory.
But
I
bet
you
I,
don't
really
know
beyond
that,
and
many
of
these
I
think
many
of
the
regions
in
the
cortex
are
uncharacterized.
They
don't
know
it's.
B
B
You
don't
have
to
use
your
cortex
to
Jeru.
Is
the
old
one?
Doesn't
you
don't
think
about
it
right?
You
know
what
you're
talk,
but
the
court
Jake's
can't
control
it
too.
So
you
know
the
cameras
on
them
say
all
I'm
gonna
think
about
the
children
right
now.
So
those
two
almost
everything
on
her
body
and
so
who
know
these
productive.
F
B
B
B
B
That
we
can
talk
about
the
another
framework
in
this
view
here,
and
there
are
cells
in
layer
six
which
project
back
to
the
thumbs.
This
is
not
a
complete
feedback.
These
cells
do
not
go
to
the
thalamus
and
back
to
the
region.
It's
not
a
downward
chain,
it's
just
back
to
the
pounds
and
so
the
same
reason
that's
projecting
the
same
relays.
Balance
here
that
are
projected
to
this
region
get
a
feedback
from
this
region
back
to
them.
So
this
is
the
way
where
a
particular
region
like
this
region,
can
control
its
inputs.
B
B
Go
here
to
here
here
to
here
to
here,
don't
do
that.
It's
just
like
a
it's
like
some
sort
of
control.
On
top.
There
are
a
lot
of
these
counter
cells.
There
are
ten
times
as
many
cells
projecting
back
to
the
thalamus
as
there
are
from
the
that
was
projection
to
this
region.
So
the
one
that's
projected
from
the
thalamus.
This
is
this
view
throat
pathway
here
remember
these
are
drivers.
B
This
is
like
a
very
powerful
pathway,
there's
ten
times
as
many
neurons
projecting
back
and
it's
not
a
driver,
it
does
it's
a
modulator,
each
other
change
the
input,
but
it's
not
like
driving
there,
but
it
doesn't
it
doesn't.
It
doesn't
do
right
and
make
these
cells
fire
it
changes
which
ones
are
gonna,
sell
fire
and
how
that
fire,
but
there's
a
ten
to
one
ratio.
G
B
Really
convolutional
neural
network
ideas
came
from
us
a
super
type
of
session.
This
is
more
of
the
reality
what's
going
on
and
that
these
this
sort
of
sensory
motor
every
system-
this
is
hoses.
Every
system
is
a
sensory
motor
influences
every
column,
every
every
Regis
and
every
column.
It's
essentially
different
system.
It's
not
just
a
feed-forward
spatial.
You
know
cooler
and
very
complex
functions
going
on
in
here
and
it's
creating
behaviors
everywhere,
and
so
now
we
can
see
the
system.
B
E
B
Extra
layer,
three
layer,
three
projects
with
a
5-2
that
in
here,
but
about
40
to
50
percent
of
the
synapses
on
layer,
four
cells
come
from
layer,
six
black
40,
because
then
of
the
inputs,
these,
like
four
cells,
is
coming
from
layer.
Six,
these
these
don't
appear
to
be
the
main
drivers,
but
that's
a
lot
of
synapses
and
it
turned
out
that
layer,
four
cells,
also
project
back
to
the
latest-
excels
it's
a
little
odd,
mostly
the
apical
dendrites
elastic
cells
are
here,
but
the
projection
is
the
same.
B
You
can
think
about
it
for
a
little
bit
like
play,
cells
or
sensory,
driven
representations
of
location
or
information,
and
and
and
so,
but
in
order
to
get
me
to
do
my
in
order
to
do
the
sensory
motor
sequences
so
that
I
can
pull
in
layer,
three
I
would
need
to
combine
it
with
a
grid
cell
location.
So
this
kind
of
mechanism
right
here
with
a
la
layer
for
not
only
two
layer,
four-
could
learn
high
order
sequences,
but
it
could
also
learn
sensing
bomber
sequences
based
on
the.
B
If
there
was
a
green
cell
like
cells
in
layer,
six
I
say
this
because
it's
more
complication
in
this,
but
that's
the
basic
idea,
and
so
we
focused
a
lot
on
a
couple
of
papers
which
sort
of
highlighted
these
massive
connections
between
these
two
layers,
because
it's
ready
it
could
be
key
to
understanding
everything
that's
going
on.
So
now
we
have
the
ability
to
say:
okay,
well,
layer.
Four
theory
could
be
learning
high
order
sequences,
it
could
be
learning
sensory,
motor
sequences
layer.
Three
could
still
be
doing
pooling
over
them.
B
It's
more
complicated
that
and
now
I've
shown
that
here's
the
feedback
to
a
syllabic
region
from
layer
six
and
here's
the
feed
forward
from
the
layer
five
cell
to
the
the
next
lamech
region
in
the
next
region
of
the
cortex
and
motor.
So
it's
a
bit
more
complete
picture
than
I
had
before
so
major
I.
Think
I'm
gonna
throw
this
one
idea
that
we've
talked
a
lot
about
the
thalamus
and
how
it
it's
it's
implicated
in
attention.
B
B
There's
a
stop
when
you
attend
to
things
is:
there's
an
established
frequency
at
gamma
frequency
I
mean
these
two
cell
populations
cycle,
the
rapid
wait
together,
eight
Hertz
and
it
so
clearly
when
you
attend
to
something
when
these
things
are
working
together,
they're
synchronized,
somehow
and
one
of
the
hypothesis
I
had
one
of
the
things
we
know
about
about
grid
cells.
So
we
hypothesized
that
those
grid
cells
in
the
New
York
cortex,
where
we're
prophesizing
the
tunnel
and
they're
gonna,
be
down
there
later.
Six
grid
cell
equivalents.
B
One
of
them,
the
question
is:
how
did
with
cells
work
and
one
of
them
hypothesis
how
grid
cells
work?
How
do
they
integrate
over
time?
How
do
you
integrate
motor
behavior
to
update
a
location,
I'm,
not
gonna,
talk
about
this
in
detail
today.
There's
two
basic
theories
about
how
good
cells
might
work.
One
is
based
on
oscillation
frequency
and
how
there's
a
certain
frequencies
of
these
dollars
that
allow
you
to
when
you
move,
how
do
the
good
cells
know
to
update
their
location?
B
B
B
You
may
know
if
you
might
think
about
it,
like
oh
I'm
leaking
subitize,
he's
really
close
to
me
or
falling
away
from
me,
but
how
I
would,
if
I
think
of
him
as
object,
that
I
can
move
my
eyes
or
our
sensory
motor
object
recognize
them.
The
angular
movements
of
my
eyes
will
be
different
if
he's
further
away
than
these
clothes
I,
don't
have
to
learn
that
different
ways,
and
what
the
things
can
be
happening
here
is,
is
that
essentially
they
could
be.
B
The
cortex
could
be
controlling,
essentially
the
the
scale
of
movement
like
a
movement
command.
How
much
it
actually
moves
could
be
controlled
by
the
falmouth.
This
oscillation,
so
I've
already
speculated
this
back
in
my
book
14
years
ago
about
how
the
assignment,
if
some
of
these
cells
McDonald's,
could
be
changing
scale
of
time.
B
So
that,
like
when
you,
when
you
recognize
and
melody
you
can
either
faster
or
slower
or
I,
can
I
can
play
back
the
sequence
that
I've
memorized,
faster
or
slower
so
I
have
to
have
a
way
of
changing
scale
time
in
the
cortex
and
these
cells.
Look
like
they'd,
be
perfectly
positioned
to
do
that
and
changing
scale.
Time
is
really
the
same
as
changing
scale.
Space,
they're,
really
kind
of
the
same
thing,
and
so
the
idea
that
the
accomplish
could
be
this
role.
B
That's
not
you
tend
to
sort
of
things,
but
attention
means
you're,
also
you're
scaling
space
time
to
adjust
to
the
new
new
location
of
this
object
or
the
new
temp
of
this
object
or
the
new
position
of
this
object.
It's
a
little
more
company.
It's
it's!
It's
a
complex
thing
that
absorb,
but
that
the
idea
that
this
is
really
a
scale
and
function
in
a
digital,
attentional
function,
I
think
I.
Think
there's
a
lot
of
Merit
to
that,
because
the
scale
property
has
to
occur
to
be
able
to
scale
a
tiny
space.
H
H
B
So
we
almost
like
a
smartphone
pocket,
you're
someplace
right
and
there
was
a
company
that
came
in
there
recently.
Actually,
they
happen
they
now
own,
the
name
pop
and
they're,
making
the
try
to
solace,
and
it's
like
you,
look
at
it's
like
a
teeny
version
of
this,
and
it
does
everything
just
like
a
version
of
smaller
and
well.
I
have
learned
how
to
use
this
farme
I
now
type
on
you,
my
thumbs
and
I
know
that
stole
things
up
and
down
I've
learned
how
to
interact
with
this
phone.
You
give
me
this.
B
B
Learned
a
certain
movements
here
tight
and
now
you
give
me
a
smaller
keyboard:
I
can
type.
How
does
the
data
curve?
You
know
how
do
I
know
only
the
room
I
think
are
a
little
bit
and
not
as
much
as
I've
been
rolling
of
it.
So
this
kind
of
stuff
is
happening
everywhere
in
the
world.
We're
constantly
scaling
things
both
in
space
and
time.
B
So
this
kind
of
stuff
you
can
you
can
convince
yourself-
is
current
everywhere
right.
So
it's
an
it's
an
inherent
property
that
must
be
addressed.
It's
system-wide!
So
that's
that's!
What
the
beauty
of
having
is
in
some
central
location
I
can
scale
the
entire
input
to
this
region
in
one
spot,
even
though
I
can't
do
it
up
here,
because
our
no
long-range
connections
that
go
from
this
side
to
the
side,
but.
H
B
B
They
are
not
the
related
cells,
okay
and
the
matrix
cells
also
project
to
the
court,
but
when
they
project
to
the
cortex,
they
projectory
broadly
over
a
single
sensor
modality.
This
is
what
got
me
a
long
time
ago.
So,
for
example,
if
I
look
at
the
vision
that
matrix
cells
in
the
visual
regions,
they
projected
cost
all
of
these
guns.
B
B
So
what
they're
doing
is
providing
is
timing,
signal
that
the
different
cells
in
the
different
layers
can
use
to
regulate
time
and
and
if
I
were
going
to
adjust
the
time
I
would
adjust
it
for
all
of
them.
This
is
consistent
with
a
thousand
branches,
their
hypothesis-
it's
not
like
this
guy
is
running
at
one
time
scaling
this
we've
got
another
time
covered
they're
all
doing
the
same
thing,
I'm
all
running
at
the
same
time
scale.
He
said,
maybe
in
the
vision,
this
one's
looking
at
details
in
one
part
of
the
visual
space.
B
This
is
when
you
have
more
water
punch
in
the
visual
space
and
so
on.
Yet
there's
still
this
sort
of
hierarchy
to
this.
We
have
to
confine
all
that,
but
anyway,
that
idea
that
the
idea
that
there's
easy,
that
you
be
scaling
time
or
scaling
attentional
time
and
space,
would
it
be
applied
to
all
the
different
regions
in
that
modality,
but
I
don't
see
matrix,
L,
jumping
from
the
visual
region
to
the
third
so
auditory.
B
All
the
ortho
system
is
going
to
run
on
one
sort
of
time
frame
and
all
the
visual
stuff
is
gonna
run.
Another
total
time
frame
you
can't
relate
a
visually,
tend
to
two
different
things,
and
you
can't
really
auditorially
tend
to
do
jump
back
and
forth
and
there's
a
lot
of
evidence
that
support
this.
When
you
change
attention,
these
cells
change,
they're,
actually
changes.
B
Are
all
hypotheses
on
the
existing
I
think
anywhere
else,
but
I'm
pretty
confident
that
their
basement
right
so
I?
Don't
you
can
ask
more
questions,
but
it
just
took
the
bottom
line
of
all.
This
is
I.
Think
the
most
important
thing
to
our
work
right
now
is
is
that
when
we
think
about
the
ultimate
AI,
is
it's
not
going
to
be
a
hundred
layers
of
so
cooling
or
match
boiling,
or
something
like
that?
B
It's
going
to
be,
you
know
complex
sensory
motor
functions,
of
which
you
might
have
multiple
levels,
a
few
levels
in
the
hierarchy,
but
everyone,
the
thousand
brain
theory,
says
every
one
of
these
system,
so
you
can
build
a
complete
accident.
It
may
be
limited,
but
you
could
do
that
and
where
you
can't
do
that.
We
should
know
that.
So
this
is
a
much
more
powerful
system,
and
today,
when
I
talk
without.
C
B
Today
is
right:
now:
that's
what
we're
trying
to
choose
a
property,
the
sparsity
but
I
believe
that
in
the
future,
we're
gonna
have
to
choose
all
these
concepts
to
do
it.
It's
true
AI,
it's
optional!
You
don't
have
to
done
this
way.
You
do
all
the
ways,
but
these
functions
have
to
be
they're
scaling
the
time
motor
centric
order
and
inference
to
know
all
something's
got
to
be
there.
Okay,
I'm
done.
H
B
F
B
B
There's
a
lot,
that's
known
about
the
detailed,
really
detailed
architecture,
and
it's
very
unusual,
and
so
we
have
postulated
we've
looked
at
that
detailed
architecture
and
postulated
how
this
rerouting
stuff
could
work.
So
the
idea
that
you
know
literally
when
I
attend
to
something
I,
don't
I'm
really
trying
to
turn
off
the
inputs
from
elsewhere,
like
I.
Don't
really
want
to
see
everything
in
the
background
here.
I
just
want
to
see
this
thing,
and
so
there's
a
physical
routing
attention,
in
addition
to
the
sort
of
scale
and
properties
appropriate.
B
B
B
But
it
only
appears
if
one
of
them
is
actually
doing
anything,
so
they
said.
Well,
that's
the
real
lifestyle.
I
don't
want
these
are
connected
to
that
and
well
we
think
it's
going
on
we're
saying
well
under
the
correct
conditions
this
could
be
rather
and
this
one
can
be
turned
on,
so
you
really
can
multiplex.
You
can
just
read
out
these
signals.
It's
still
one,
the
one
but
the
routing.
There's
no!
There's
no
like
a
convergence
on
these
cells.
F
B
E
F
B
B
All
we
know
that
there's
these
oddities
there's
oscillations
that
are
occurring
when
we
look
at
you
know,
area
potential,
potentiation,
I'm,
not
sure
they've,
measured
them
in
natural
real
is
on
bottom
of
this
true.
But
if
you
look
at
the
overall
activity
of
this
area
used
to
get
a
probe
what's
going
on,
is
there
it'll
finish?
You
know,
wave
see
this
way
so
I'm,
not
sure
I,
don't
know.
If
you
know
you
know
I,
don't
they
know?
B
That's
not.
My
theory
is
that
that
oscillate
that
somehow
the
oscillation
you
wanna
you
want
to
establish
an
oscillation
because
all
these
regions
that
are
the
uniformly
impact
time
and
in
space
scalar
motor
scattering
from
every
every
region
that
Laval.
So
it's
more
like
how
listing
projects
back
to
here
with
that
frequency,
which
we
didn't
then
impact
how
the
grid
cells,
when
you
get
a
vigil
to
get
at
a
motor
command?
How
much
do
actually
both
it's
the
question
and
also
about
the
timing
of
sequences,
hey
I,
look
good.
B
There's
not
a
lot
of
literature
on
this,
but
we
did
do
something
about
that.
I
can't
remember
many
democracy,
but
I
was
testing
about
the
year
because
they
have
to
be
able
to
Twitter
it's
sort
of
at
the
same
rate,
which
you
can
switch
attention
yeah
and
my
recollection
was
where
they
can
change
it's
really.
When
you
attend
to
something
it's
quick
as
you
can
attend
something
that
that
frequencies
of
stauch
memory
a
lot
of
time
you
be
not
attending
some
there.
Isn't,
though,
there
isn't
a
cohesive
camera
frequency,
the
channel.
B
H
H
B
F
F
B
C
E
B
Now
what
you
can
hope
is
they
can
do
this
type
of
fMRI
and
F&R
you
use
on
my
patients,
but
problem
with
fMRI
is
two
things.
One
is
the
temporal
resolution
is
not
what
you
want,
meaning
you
can't
really
pick
up
small
time
difference.
It
took
definitely
I
picked.
It's
a
blur
time
see
and
also
the
size
of
the
region
you
can
measure
is
getting
smaller
there
now
down
the
sub
millimeter,
but
which
is
still
a
lot
of
neural
tissue.
You
can't
look
at
individual
cells.
B
A
lot
of
bloggers
get
better,
so
these
experiments
were
really
hard
to
do
and,
although
the
telco
cross
would
say
this
could
be
testable,
it
says
to
you
this
theory
about
that
was
a
dummy
testable.
It
doesn't
may
not
be
testable
easily
and
may
not
be
testable
using
today's
technique,
but
in
zeroes
testable
it
makes
predictions
that
can
be.
D
B
They
can
measure
that
you
can
measure
some
of
those
because
you
with
scott
recordings,
you
can
put
the
EEG
scott
recordings
on,
and
you
know
these
massive
like
machine
massively
your
when
you
go
through
the
scalp
and
them
in
the
skull.
You
can
measure
the
activity,
the
upper
layers
of
the
cortex
in
some
regions
over
large
areas,
so
they
can
see
they
can
see
frequencies
that
are
occurring
sawing
out
through
sleep
you'll,
see
ones.
That
are
it's
another.
B
If
you
want
to
get
something
more
refined
than
that
more
specific,
like
okay,
what's
going
on
in
this
little
area
right
here
and
then
you
guys
take
a
program
and
and
then
and
then
you
yeah,
you
can
measure
the
potential
the
local
field
potential.
They
call
it,
but
just
like
all
the
cells
around
this
probe,
I
can
solve
their
electrical
activity
and
sometimes
I
got
a
dividual
spikes.
B
So
there's
there's
a
slew
of
different
techniques.
All
vary
in
terms
of
their
their
difficulty,
the
preparation.
What
can
be
done
about
animals,
whether
slice
or
vivo
or
the
animals,
keep
your
feet
moving
are
not
free
moving
because
they
some
of
many
customers.
Even
if
they
stick
a
card
with
the
brand,
they
have
a
sweet
lock.
The
breakdown.
B
B
B
B
C
B
Of
scaling
I'm
talking
about,
has
to
occur.
It's
not
an
option.
I
prove
that
it
has
to
occur.
Then,
when
you
try
to
map
that
like
well,
where
could
these
things
be
occurring
and
what
Anatomy
supports
that?
And
so
we
can
make
a
lot
of
progress
without
I
can
doing
these
experiments.
We
can
just
look
at
all
the
existing
data
and
deduce
what
what
are
the
options.
You
know
we're
working
this
beaker.
C
B
I
should
mention
something
about
technical
reasons.
I
was
wondering
freaking
spider.
This
idea
there's
many
to
one
mapping.
Typically,
it's
very
memory
intensive,
so
you
can't
map
a
million
patterns
into
one
that
can't
do
it.
There's
not
sit
on
to
do
that.
So
we
did
a
lot
of
testing
with
the
Spartan.
This
is
and
this
Parsons
these
are
the
more
you
can
do
the
threshold
of
the
dendrites
matter
to,
but
there
is
a
lumen
and
one
of
the
things
I've
thought
about.
This
is
when
you
learn
like
melodies
and
melodies,
be
quite
long
and
I.
B
B
It's
like
this
always
work
on
the
edge
of
puppet
abilities.
It's
a
very
honest.
It's
probably
head
we're
pushing
it.
So
as
things
decay,
you
lose
that
at
first
you
know,
you
don't
lose
the
you
know.
You
know
the
word
you
just
can't
think
of
you
can
picture
the
person
you
can
picture
what
they
do.
You
can
think
of
all
the
things
that
you
know
about
oppression.
You
can't
remember
the
name.
C
B
Exactly
yeah
boys
matter
like
name
it
as
long
as
I,
know
what
it
is
to
do
about
it,
what
to
do
about
it
right,
yeah.
So
again,
it's
like
it's
like
melody.
I've
got
all
these
knowledge.
I
just
can't
put
the
name
on
em,
that's
a
that's!
A
placeholder
for
recognizing
people's
faces
other
people's
when
they
are
in
ability.
C
B
B
H
F
B
But
you
can't
write
all
that
up
and
on
papers
you
know.
Yes,
it
was
quite
enough
to
write
those
papers
even
covering
what
we
did,
because
they
one
of
the
first,
the
first
rejections.
I
got
on
them
there
on
paper,
we
got
paper
we're
covering
too
many
things.
I
talk
about
too
many
stuff.
You
need
to
focus
on
one
little
thing
and
problem.
B
Is
you
it's
hard
to
focus
on
another
thing,
because
it's
the
system
that
comes
together-
and
this
is
a
problem,
basically
that
neuroscience
and
doesn't
hasn't-
had
an
equivalent
to
theory
but
like
physics
or
cosmology,
hands
where
people
in
cosmology
can
come
up
with
grand
theories
about
time
and
space
and
that
once
it's
wonderful,
we
can.
We
come
up
with
grand
drains
about
the
brain
people
set
off
too
much.
You
shouldn't
do
that
it's
impossible.
E
E
B
B
B
When
you
mix
your
sentences,
I
probably
happen,
I'll
break
it
into
the
block.
When
I
was
growing
up,
my
father
didn't
like
peace.
I
said
why
don't
you
like
beats
and
he
goes
are
they?
They
have
an
audio
a
text,
and
that
was
like
the
running
joke
only
when
he
was
like
90
years
old
or
85
years
old,
then
I
understand
that
he
sounded
like
a
a
or
t-28
beats.
He
heard
boy,
it
wasn't
like
they
had
a
description
of
the
taste,
that's
what
he
heard
so
the
Edison
decision
anyway.
B
What
Ramachandran
proposed
is
that
movie
would
want.
You
would
not
want
any
connections
between
these
two
and,
and
so
when
he
suggests
it
was
perhaps
was
gone,
a
lot
of
synesthesia.
You
do
have
these
connections
that
are
just
Galan
crossing
over
between
the
same
visual
and
auditory,
and
he
gave
evidence
to
show
that
the
classic
synesthesia
of
symptoms
aligned
with
regions
that
are
the
two
different
types
that
are
right
next
to
each
other
and
and
so
when,
when
you're
getting
activity
here
in
activity
here
and
that's
why.
B
Other
stations
may
not
be
the
sentenced
anesthesia,
the
evolution,
Nations
you're,
imagining
things
are
out
there
synesthesia
is,
is
real.
These
people,
you
know,
that's
the
sound
of
beat.
It
was
not
like
you
to
loosen
it
right,
it's
that's
what
it
is
right.
You
know
just
like
we
say,
Beats
or
red.
Well,
why
are
they
red
point.