►
Description
Jeff Hawkins reviews the thalamic inputs to the various layers, and discusses their importance on minicolumns, representing features vs movements, and a surprising finding regarding simple and complex cells. Discussion ensues.
B
All
right
well
great,
so
this
is
just
a
brief
update
on
the
mini
column.
Hypothesis
I've
been
talking
about
recently
and
I,
haven't
been
able
to
spend
as
much
time
as
I
want,
but
I
have
been
reading
a
few
papers
every
now
and
then
so
I'm
going
to
go
over
a
few
ideas.
This
could
be
pretty
short.
If
there's
questions
that
give
a
lot
so
I
know
you
guys
aren't
thinking
about
this
every
day.
I
am
thinking
about
it
every
day,
so
you
know
there
might
be
disconnected.
B
You
know
every
honest,
okay,
so
I
just
want
to
start
with
sort
of
the
sort
of
the
classic
view
what
they
call
the
hierarchal
interpretation
of
the
google
and
visa
findings.
This
is
showing,
of
course,
a
column,
and
when
people,
when
you
build
easel
started,
you
know
experimenting
discovering
what
sells
respond
to
it.
Initially
in
cat
v1
cortex
they
of
course
they
would
see
that
there
was
a
classic
view
that
there
was
an
input
until
therefore-
and
this
came
from
the
thalamus
and
there
they
found
these
simple
cell
response
properties,
injury.
B
You
know
lying
orientation
at
a
specific
point
and
then
and
then
they
would
find
these
layer
three
complex
cell
responses,
which
seemed
to
respond
to
the
set
of
the
lying
being
in
any
position,
and
often
they
were
emotion,
sensitive
I
saw
you
were.
You
went
from
being
primarily
a
very
specific
orientation
field
to
a
less
specific
motion,
more
motion
oriented
it's
a
mixture
of
cells
here,
but
that's
the
basic
thing
now
here.
I
also
showed
this
classic.
You
know
connection
we've
talked
about
many
times
between,
but
this
is
how
they
originally
described
it
initially.
B
They
say:
oh
and
then
the
late
tree,
three
cells
projector.
There
are
five,
and
so
you
also
see
conflict
zones
in
my
five
and
back,
then
they
didn't
really
have
any
way
of
characterizing
their
six
of
all.
So
that
was
a
classic
view
they
held
for
a
while.
So
and
the
idea
there
was
simple
cells
were
derived
from
sensory
input
and
then
complex
cells
are
derived
from
simple
cells
and
there's
some
more
invariant
to
location.
B
So
that's
why
it
was
called
hierarchical,
because
the
idea
that
the
complex
cells
in
layer,
two
three
were
derived
from
the
cell
property
mayor
for
things-
got
a
lot
more
complicated
shortly
thereafter
that
so
this
is
a
the
current
understanding,
and
this
has
been
known
for
some
time.
I,
don't
let
me
get
their
little
yeah,
it
keeps
blocking
top
of
my
screen.
We
now
know
there
are
three
different
destinations
of
input
to
a
column.
B
There's
the
original
layer
for
one
there's
a
separate
input
at
the
bottom
right
at
the
border
of
layer,
two
three
sort
of
the
bottom
layer:
three,
some
people
call
a
freebie
and
also
a
similar
input
at
the
border
of
layer,
five
under
six
really
in
layer,
the
lower
layer,
five,
so
I've
shown
those
three
here.
This
complicated
things
a
bit
and
also
I,
will
get
to
it
in
a
moment.
B
There's
there's
a
bunch
of
physiological
observations
about
this
which
I
want
to
get
to
in
a
moment,
but
first
I
want
to
talk
about
hypothesis
I'm
working
now
the
hypothesis
I'm
working
on
I'm,
just
gonna
I'm,
going
to
call
it
in
all
of
the
HTM
in
common
hypothesis.
It
says
that
many
columns,
they
don't
represent
just
features,
but
they
represented
a
graded
meaning
of
a
spiky
rate,
despite
great
determined
response
to
movement
in
particular
direction.
B
This
idea
was
derived
because
I
was
a
accident.
How
a
calm
could
discover
the
space
that
is
modeling
and
I.
Also
based
on
the
idea,
you
have
these
movement
sensitive
cells,
which
really
never
made
hell
of
a
lot
of
sense
and
the
classic,
you
know
feature
extraction.
What's
his
image
model
hierarchical
model,
it's
like.
Why
do
we
have
all
these?
Why
most
of
these
cells
responsible
direction,
but
so
the
hypothesis
here
since
oh,
what
would
work
with
the
mini
colonies?
I've
shown
them
just
one
mini
column
to
be
several
hundred
in
a
that.
B
Each
mini
column
represents
a
movement
in
a
particular
direction
and
that
would
be
learned
from
basically
flow
information
in
the
input.
So
and
therefore
like
what
that
calm
is
really
representing,
is
not
some
feature
or
moving.
It's
representing
movement
of
the
sensor
in
some
space
in
one
direction
and
I
talked
to
how
this
can
be
learned
from
a
spatial
cooler
type
of
thing.
B
So
in
that
model,
mini-com
model
says
the
settlement
accounts
defines
the
dimensionality
of
the
space
and,
and
it
defines
the
movement
primitives
possible
in
that
space
back
to
CL
way
around
first
to
find
some
movement
principles
that
are
observed.
The
movements
that
observe
primitives
excuse
me
and
then
that
essentially
defines
the
space
itself.
So
these
are
the
ways
you
can
move
it,
and
these
are
that
these
are
the
directions
that
thing
can
move
it
anymore
in
time
and
these
those
are
the
dimensions
that
that
is
a
base
set
of
the
dimensionality
space.
B
It
really
ties
the
space
in
movements
together,
which
is
a
nice
idea.
Under
this
idea,
we
know
that
the
layer,
5
cells
are
the
motor
output
cells,
behavioral
output
cells,
but
every
column
in
the
quarter,
and
so
that's
a
nice
interpretation
there,
because
we,
the
column,
can
observe
what
movements
are
possible,
but
then
it
can
generate
them
as
well
and
it
just
generates
from
the
same
way.
B
Inverted
observers
by
activating
their
5
cells
to
say,
I
want
I,
want
to
move
in
this
direction
and,
of
course,
the
set
there
might
be
a
set
of
active
meaning,
columns
and
they're.
Together
they
represent
the
movement
sort
of
the
direction
we
this
thing
wants
move
and,
of
course,
complex
movement
would
be
a
set
of
activations
of
many
columns
over
time.
So
you
have
a
set
primitives
and
then
you
could
move
different.
B
This
ranges
from
this
route,
so
that's
basically
there
that
the
hypothesis
has
been
working
on
I,
don't
know
if
anyone's
totally
confused
by
that
or
not
before
I
go
on
to
the
next
thing
speak
up.
If
you
want
to
explain
something:
I'm.
C
B
It's
not
clear
why
most
of
the
cells
in
a
column
would
be
directionally
sensitive
when
you
think
about
what
people
do
in
terms
of
modeling
the
hierarchical
model
they
like
like
we're
doing
any
resonator
right
now,
you
don't
you
don't
take
advantage
of
motion
they're,
all
it's
just
a
spatial
information.
So
why
would
almost
all
the
Selphy
motion
sensitive
well,
there's,
obviously,
answers
that
we're
modeling,
a
sensory
motor
system,
but
I've
long
been
pointing
out
is
that
there
hasn't
been
a
really
great
explanation
for
that
I
didn't
say
it's
a
huge
problem.
C
B
B
B
It
came
from
the
fact
we
have
a
problem
to
solve
in
a
column
and
it's
very
difficult
problem
to
solve.
How
does
it
calm,
discover
the
space
it's
working
and
how
does
it
learn
the
dimensionality
of
space?
How
does
it
represent
movements
in
that
space?
How
does
a
generate
movements
in
that
space?
These
are
really
hard
problems.
As
far
as
I
know
no
one's
talked
about
in
the
past
because
they
don't
think
I'm
a
mini
column
as
a
column
representing
eccentric
motor
model.
B
So
the
the
derivation
of
this
HTM,
any
kind
of
positives,
does
not
come
from
like
oh
there's,
this
stuff
in
the
literature
is
unexplained,
it's
more
like
no.
We
have
a
new
idea
what's
going
on
in
a
column
and
it
requires
a
column
to
do
a
lot
of
really
screwy
things,
but
we
have
no
idea
how
it's
going
to
do,
but
it
has
to
do
them
so
then
I
started
looking
at
the
many
columns
and
I
said
yeah.
That's
right,
I
mean
there's
no
real
explanation
for
why
you
know
complex
elves,
a
motion-sensitive
like
that.
C
B
B
We
need
to
find
a
way
of
how
this,
how
Canton
learned
these
things
and
that
jumped
out
at
me,
like
yeah,
that
thing
I
never
explained
I,
never
could
really
justify
why
all
those
cells
and
motion
sensors
I
would
I
would
do
it
as
I
said
and
sort
of
hand,
gravy
waxing,
but
sometimes
things
in
the
world
move
so
again,
but
trees,
swaying
in
the
breeze.
Well,
then,
I
have
motion,
but
most,
but
that's
not
what
we're
doing
when
people
model
on
visual.
You
know
they
model.
Like
you
know,
classification
systems
like
we're
doing
yeah.
B
Okay,
so
this
was
the
idea
that
many
comp
now
could
really.
It
gives
a
purpose
for
the
first
time
for
like
a
real
functional
role
of
many
homes
and
that's
been
missing
other
than
this
idea.
Well,
there
feature
extracted,
but
didn't
results
much.
So
if
we
go
with
this
idea,
there's
some
implications
that
are
really
hard
to
do
it
so
we're
surprising.
Let's
put
it
that
way,
so
the
any
common
hypothesis
implies
that
many
columns
have
multiple
components
that
is
layer
force.
B
So
there's
really
two
things
you
have
to
represent
in
a
column,
you
have
to
represent
a
spatial
metric
system
which
we
are
believe
is
going
on
the
lower
layers
and
you
have
to
represent
sort
of
the
observed,
object,
features
or
object,
which
is
going
on
the
upper
layers,
and
so
it's
kind
of
convenient
to
think
of
like
oh
well.
We
have
these
upper
layers
and
these
lower
layers.
We've
pointed
us
how
many
times
in
the
past
there's
a
real
parallel
construction
here
you
can
see
it.
B
The
difference
is
that
you
don't
have
a
direct
sensory
input
bill
here
for,
but
we
wouldn't
want
that
in
this
case,
but
we
do
want
to
make
form
here
and-
and
so,
but
this
now
says
that
if
I
look
at
a
mini
column
that
it's
not
one
thing,
it's
going
to
be
multiple
things,
so
in
this
case
layer
four
represents
I'll,
just
call
it
sense
feature:
it's
got
to
be
the
thing
that
we're
actually
sensing
and
therefore
it
should
not
be
movement
sensitive.
It's
it's
more
like
this
is
the
V.
That's
at
this
point.
B
It
shouldn't
be
room
insensitive,
but
then
layer,
six
cells,
we
represent
movement
it,
but
they
shouldn't
be
feature
sensitive.
So
you
know
it's
like
how
I'm
moving
doesn't
really
matter,
what
I'm
observing
it's
just,
how
I'm
moving
so
and
when
the
classic
literature
talks
about
this.
They
talk
about.
Oh
yeah,
all
the
cells
in
the
many
columns
share
or
receptive
field
properties.
They
all
share
this
orientation
property,
but
the
HTM
any
comment
provinces
says:
that's
not
really
right.
B
If
this
theory
is
correct,
then
these
things
actually
should
be
different
and
there
are
times
when
this
thing
would
be
representing
an
orientation,
let's
say
vertical,
but
down
here
this
would
be
moving.
This
would
be
representing
a
movement,
that's
different
than
the
orientation
I
mean
it's
so
sometimes
I.
She
sells
sea
sells
here
that
are
maybe
representing
movement
that
are
that
are
that
are
at
45
degrees,
where
the
actual
feature
is
that
I
needed
and-
and
so
this
was
a
very
strong
prediction
of
the
HTM
and
economic
opposite
and,
and
so
I
said.
C
C
D
C
B
B
So
maybe
you
can
add
on
to
what
I'm
about
to
show
you
so
so
I,
just
as
I
just
described
here
under
this
hypothesis
orientation
quote
of
late
and
for
six
will
not
match
under
most
condition
they
might
match
under
the
conditions
anesthetize
animals
being
shown,
sinusoidal
gratings,
but
but
they
made
up,
they
should
match
under
normal
conditions,
and
so
that's
the
thing
I
was
pursuing.
B
There's
another
interesting
observation
about
this
is,
as
we
know
that
there
are
small
repres
of
the
fields
that
the
width
of
your
field
properties
and
the
upper
layers
is
much
smaller
than
in
the
lower
layers.
I
showed
that
paper
a
couple
weeks
ago
that
old
paper
from
the
seventies
like
I,
can
bring
up.
If
you
want
where
we
had
these
very
very
wide,
you
know
areas
of
the
retina
which
can
impact
the
which
were
added
is
to
them
to
the
responsibly
or
6l
I'm.
Here
you
had
a
lot.
B
You
were
very
much
sorter
and
a
lot
of
10
stops
Alistair
that
were
basically
we
up
here.
They're
really
restricted
the
small
orientation,
but
then
there's
a
question
is
like
Hawaii,
so
either
so
working
hypothesis.
Now
that
these
lower
layers
are
going
to
basically
representing
movement
directors
and
the
up
here
layer
for
we're
going
to
be
representing
features
and
the
features
and
the
movement
directors
may
not
be
Co
aligned.
But
then
why
do
we
have
complex
stuff?
Familiar
to
three?
B
That's
that's
still
a
little
bit
of
a
mystery
to
me
and
these
are
smaller
receptive
field,
so
the
UC
complex
cells
up
in
layer,
two
three
there's
symptoms
to
the
topic.
So,
as
you
see
down
here
in
five
and
six
but
they're
much
smaller
and
extent
and
Phil
I,
don't
have
any
answers
that
question,
but
it
is
suggested
and
I
suggest
the
answer.
Possibly
entry
down
here,
I'm,
not
at
all
confident
I've
become
pretty
confident.
This
basic
mini
column,
hypothesis
idea
I'm
pretty
certain.
B
Perhaps
the
layer
two
three
cells
are
defying
a
smaller
space
that
is
smaller
and
the
extent
of
the
input,
like
smaller
part
of
the
retina,
a
small
part
skin
versus
down.
Here
you
have
a
much
larger
space,
which
is,
which
is
an
interesting
idea,
that
it
could
be
representing
composite
objects.
We've
talked
about
composite
objects,
so
quite
a
lot
in
the
past,
where
you
need
to
represent
an
object
relative
to
another
object,
in
fact
my
displacement,
selves
and
so
on,
but
it's
an
interesting
idea
that
it
might
actually
be
representing
two
different
spaces.
B
In
a
cortical
column,
one
is
a
smaller
space
of
the
object
than
one
is
a
larger
space
of
its
environment
or
the
larger
context
of
which
it
occurs.
I'm
not
ready
to
go
there
yet,
but
I
just
wanted
to
point
out
that
that's
a
very
interesting
idea
and
I'm
gonna
be
trying
to
which
other
internet
so
now.
Having
said
this,
I'm
gonna
have
a
jump
to
a
couple
of
papers
that
I
looked
at
this
one
this
one
here
you
see
this
connectivity
down
your
screen.
A
A
B
Heard
okay
I
heard
of
no
and,
and
yet
this
paper
is
from
1998,
this
missing
I'm,
not
gonna,
spend
a
lot
of
time
in
this
paper.
Just
I
just
want
to
show
you
what
happened
when
I
read
the
paper,
so
it's
kind
of
functional
connectivity
between
simple
cells
of
complex
cells
and
it
starts
off
with
this.
It
says:
okay
in
the
primary
motor
cortex
neurons
are
classified
into
two
main
categories:
simple
and
complex
cells,
but
you.
B
According
to
the
hierarchical
model,
which
are
just
talking
about
accomplishes
up,
the
fields
are
derived
from
convergent
input,
the
simple
cells
with
similar
orientation
preferences,
so
that
was
that
first
PowerPoint
slide
show
and
then
they
say
well.
Physiological
evidence
for
the
model,
however,
has
remained
elusive,
meaning
that
the
idea
that
they
complex
also
derive
from
simple
cells.
They
said
yeah,
there's
a
lot
of
evidence
work,
but
what
we're
gonna
do
in
this
paper.
B
First
of
all,
the
the
alternate
evidence
so
they're,
saying:
okay,
we're
going
to
show
that
there's
physiological
evidence
the
hierarchical
model
threat,
but
why
would
we
down
it
so
it
says
some
complex
cells
receive
direct,
genuine
input.
Those
of
that
layer,
two
three
and
four
I
talked
about
some
similar
I
seem
to
be
driving
by
complex
self.
B
In
that
simple
felt
when
this
means-
and
this
is
this
I'm
going
to
show
you
this
reference
in
second
year,
but
they
found
was
that
layer,
two
three
cells
which
fire
the
complex
cells
were
fired,
even
if
they
turned
off
the
input,
the
lab
for
it.
So
they
keep.
This
is
a
verbal
layer
for
so
and
Lafe
ourselves,
weren't
responding,
but
you
still
got
complex
cells
in
layer,
two
three
which
is
sort
of
says.
Well,
they
actually
three
cells
can't
be
derived
from
land
for
those
people.
B
Therefore,
cells
would
not
me-
and
maybe
oh
that's
what
they're
saying
right
here.
Oh
no
I'm,
sorry
I
misread
something.
I'm
asleep
driving
a
complex
cells
are
not
simple
cells
yeah.
This
is
under
without
turning
off.
They
did.
Sometimes
you
have
these
inputs.
It
would
drive
a
complex
cell
that
would
not
find
something
then-
and
this
is
the
one
where
they
say
they
turned
off
the
input
to
layer
four
and
lay
a
truth.
We
shall
still
respond.
It's
fine,
so
I
said
well.
These
are
really
good
information.
B
B
B
Is
this
so
this
is
the
paper
that
says
that
you
can
have
differential
responses
between
simple
and
complex
cells,
mom
and,
and
they
say
to
visual
texture.
What
they
really
mean
by
this
is
they
get
a
whole
series
of
experiments?
Maybe
this
is
relate
to
what
arrogance
is
saying,
but
instead
of
showing
bars
or
some
sort
of
gratings,
they
showed
random
bit.
B
This
is
this
is
back
in
the
day
when
they
didn't
before
people
had
a
lot
of
computers,
though
they
had
to
generate
these
big
patterns.
It
did
a
little
interesting
inside.
They
couldn't
just
program
a
computer
to
generate
these
big
patterns.
They
had
to
create
a
mechanical
machine
to
generate
these
big
guys.
You
think
oscilloscope
there's
been
a
fun
anyway.
They
they
created
these
big
patterns
and
what
they
do
is
now
with
Yummie
draining
a
big
condom.
So
there's
no
pattern
in
your
right.
B
There's
no
spatial
pattern,
but
you
can
direct
the
bit
pattern
and
you
can
have
movement
so
you
can
cut
until
you
never
have
any
spatial
pattern,
but
you
have
movement
and
they
can
also
move
things
with.
You
can
move
us.
You
can
move
a
subset
of
the
bit
pattern,
so
you
can
move
a
bar
free,
a
bar
or
random
bit
could
move
through
the
other
random
bit,
and
so,
if
you
look
at
it
at
any
point
in
time,
it's
all
random
but
there's
still
movement
there.
B
So
there's
a
whole
bunch
of
serious
a
series
of
experiments
using
this
kind
of
input
and
I'm
not
condemning
these
two
papers
didn't
have
very
interesting
figures,
digits
lot
of
data,
but
I'll
summarize
it
for
you
here.
So
this
is.
This
is
my
summary
from
the
paper.
I
was
just
showing
you
the
differential
response
to
the
simple
cells
and
complex
cells.
Again
this
is
1977
that
were
going
back
a
long
time
here
so
goal.
This
is
the
quote
from
paper.
B
Their
goal
is
to
distinguish
between
sensitivity
to
stimulus,
orientation
and
sensitivity
to
direction
emotion,
which
are
confounded
when
not
feeling
stimulus
patterns
are
used.
So
in
the
classic,
Google
and
Wiesel
experiment,
you
couldn't
tell
the
difference
between
an
orientation
column
or
an
orientation
of
a
bar
and
sensitivity
of
motion,
and
so
by
using
these
brilliant
bit
patterns,
you
could
you
could
down
differentiate
and
they
went
to
holes
through
the
experiment.
B
From
that,
then
they
say
the
another
conclusion
is
again
a
quote:
complex
cells
which
respond
just
as
well
to
moving
noise
or
responsive
light
as
to
bars
I
mean
they
found
cells.
That
said,
I,
don't
care
of
the
bar
and
orientation
I
just
need.
The
movement
suggest
that
it
may
be
more
useful
to
regard
these
as
components
of
emotion,
analyzing,
rather
than
the
shape
analyzing
system.
So
they're,
saying
yeah,
you
know
he
says
what
but
like
they
really
do,
detecting
motion
and
less
about
shape
and
I.
Think.
B
Okay,
here's
another
coconut
paper:
it
isn't
the
deeper
layer
cells
that
are
most
responsive,
the
texture,
meaning
the
random
bit
pad
emotion
and
least
interested
in
stimulus
geometry.
So
it
is
again
the
supporting
hypothesis
that
the
lower
layers,
5
and
6
are
really
not
looking
at
orientation
bars
at
all
they're,
just
really
looking
at
motion
detection
and
they
conclude
that
complex
cells
are
not
a
second
stage
abstraction
from
simple
cells.
So
they
saying
you
know
that
hypothesis,
isn't
it
right,
and
this
is
consistent
again
with
the
disabling
of
layer
4.
So
there's
another
paper.
B
I
won't
take
you
through
here,
but
basically
the
sort
of
the
when
people
reach
this
consensus
or
some
people
reach
distance.
It's
like
here
back
at
the
time
that
the
layer
2
screen
cells,
some
of
them,
are
driven
directly
by
the
thalamus
and
they
they
are
essentially
motion
sensitive.
We
now
also
say
that
the
layer,
5
and
6
cells
or
motion
sensitive
and
that
layer
4
interacts
with
the
lower
the
layer,
2
3
cells
to
produce
a
sort
of
a
somewhat
novel
representation
in
the
upper
layer.
B
B
A
A
Years
ago
you
know
these
computational,
neuroscience
conferences
and
I
know.
There's
several
papers
that
talk
about
you
know
modeling
how
complex
cells
can
be
derived
completely
from
simple
cells
and
sort
of
quote-unquote,
proving
it
various
ways
and
experimental
papers
as
well.
So
this
is,
this
is
definitely
quite
surprising.
Yeah.
B
Well,
what's
really
great
in
some
sense
the
Asian
mini-com
hypothesis
predicted
this
and
I
didn't
know
this
either
I
had
never
heard
any
of
this
stuff,
and-
and
so
it's
like
a
really
great
prediction,
because
it's
really
unexpectedly
something
unexpected
for
me.
Maybe
some
people
like
ours,
you
knew
this
I
didn't
know
this
and
a
lot
of
people
don't
know
what
a
super-tight
you
pointed
out,
and
yet
there
was
for
what
30
some
years
ago,.
A
B
Remember
the
paper
that
I
just
showed
you
know
because
I
found
more
recent
papers.
That
still
refer
me.
That's
how
he
got
back
here
right
so
I
mean
there
could
be
evidence.
I
didn't
I
haven't
done
an
exhaustive
search,
but
it's
not
like
it's
not
like
people
stopped
referencing
these
papers.
You
know,
there's
people
still
know
I
started,
you
know.
If
it
was
totally
discounted,
then
I
would
have
you
know
you
wouldn't
peek
people
reference,
you
I
I,
think
that's
very
unlikely.
I
mean
these
are
well
run
experiments.
B
They
it's
very
hard
to
see
how
their
conclusions,
how
their
observations
wouldn't
be
correct,
I!
Think
it's
just
more
of
like
you
know,
group
Indonesia,
you
know
there's
so
much
data
in
your
mind.
Some
of
you
may
not
understand
this,
but
there
were
these
waves
of
experimental
paradigms
have
occurred
over
many
decades
and
the
ton
of
people
would
work
on
one
paradigm
and
be
kind
of
run
of
extent,
and
they
didn't
tend
to
go
on
to
something
else.
B
And
then
they
forget
about
those
results
because
they
never
were
assimilated
into
some
sort
of
threat
occur
basis.
So
there's
hundred
or
maybe
thousands
of
papers
on
these
very
type
of
experiments
done
in
the
70s
and
80s,
and
it
just
you
know,
there's
lots
of
detail,
but
no
one
remembers
I'm
very
few
people
do
and
they
don't
think
about
it.
So
anyway,
I
I,
don't
think
it's
likely
that
this
is
incorrect,
especially
since
this
is
exactly
what
I
was
looking
for.
B
You
actually
and
you
know,
and
a
more
a
lot
of
people
refer
to
it
and
there
it
is,
and
the
data
looks
great,
I
think
and
easily
to
say.
Well
all
people
thought
about
it
because
they
got
in
this
paradigm.
Like
okay,
we
show
silence
or
big
ratings
and
that's
how
we
determine
ERF
of
a
cell
and
that's
how
everyone
does
it
it's
standardized
and
let's
go
on,
but
that's
not
really
true
and.
C
What
a
lot
of
these
papers
back
like
in
the
even
Hubal
and
weasel-like,
perform
some
of
these
experiments
that
sort
of
ran
contrary
to
this
dogmatic
model,
but
yeah
I,
think
you're,
right,
I!
Think
it's
mass
in
Asia!
So,
for
example,
when
they
record
from
freely
viewing
monkeys,
there's
like
almost
no
correlation
with
what's
on
the
screen
in
the
visual
cortex
and
yet
somehow
the
the
next
generation
of
neuroscientists
sort
of
ignored.
All
of
that
we.
B
Anyway,
I'm
almost
done
here,
I
thought
this
was
great.
I
was
very,
very
excited
to
see
this
I
I
didn't
mention
here,
but
there
someplace,
I,
read
and
I
forgot
with
paper
Mabel's
in
this
paper,
but
they
did
show
that
at
times
the
orient
the
the
motion
vector
that
got
the
complex
cells
to
respond.
The
best
was
not
always
aligned
with
the
feature
orientation
in
layer.
4
I
saw
that
I
didn't
I,
don't
know
where
it
is
I'm.
Finally
a,
but
that
was
another
part
of
the
hypothesis
I
had
we
don't.
B
E
Quick
quick
question:
yeah
if
we
remove
that
so
when
we
do
the
the
abstraction
between
motion
and
features
I'm
just
wondering
if
what
we're
seeing
here
in
this
complex
cell
is
the
ability
to
detect,
you
know
something:
a
receptive
field,
that's
undergoing
motion
and
somehow
in
lacking
the
right,
we're
not
intuiting
but
ferreting
out.
The
fact
that
this
this
text,
whether
its
texture
or
bars
or
anything,
is
moving
as
a
group
as
a
unit
and
therefore
has
some
kind
of
objectiveness.
E
B
Not
sure
I'm
saying
it
saying
Karen
but
I,
don't
think
that's
how
I'm
thinking
about
this
I'm
thinking
about
it's
it's,
it's
absolutely
nothing
to
do!
What's
in
front
of
you
or
what
you're
observing
it
just
happens.
These
complex
cells
are
representing
how
you're
moving
through
space,
whether
it's
your
eye
is
moving
your
head,
moving
your
walking
and
your
visual
move
into
space
for
it's
your
finger,
moving
to
some
space
associated
with
a
larger,
larger
area
around
it,
and
so
it's
we're
trying
to
do
the
opposite.
B
We're
trying
to
abstract
out
motion
detection
that
athletes
completely
independent
of
any
object
that
might
be
there
and
you
don't
want
it
to
roughly
be
tied
to
the
object,
not
in
its
pure
form
in
its
initial
form.
It's
purely
a
motion.
Sense,
it's
it's
basically
defining
the
space
doesn't
matter
what's
in
that
space
and
then
the
other
one
is
the
feature
which
is
observed.
B
Independent
of
you
know
initially
has
to
be
independent
of
where
you
are
in
the
space,
but
then,
of
course,
you
have
combined
with
you.
So
if
I
understood
your
question
and
I
mean
not
apologize,
we're
trying
to
do
just
opposite.
Initially,
you
want
to
initially
have
a
representation
of
space
in
motion.
That's
independent
of
the
object,
completely
independent,
and
then
you
can
combine
the
two
as
we
prophesized
in
our
column.
B
E
B
A
way
see
when
you
say
multiple
things
moving
around,
you
might
be
thinking
like.
Oh
there's,
a
bird
going
this
way
and
a
ball
going
that
way
and
a
car
going
that
way,
I'm
suggesting
it's
the
exact
opposite.
What
you're,
what
these
complex
cells
are
determining
are
your
flow
pattern.
It's
not
trying
to
figure
out
which
things
are
moving,
which
is
basically
saying
I'm
moving
through
space
Mike,
my
body
is
moving.
My
eyes
are
moving.
My
finger
is
moving
whatever
part.
B
Whatever
this
column
is
representing,
its
location
is
moving
through
the
space
that
it's
modeling
and
as
it
moves
to
the
space,
there
will
be
it's
getting
it's
getting
these
flow
bits
which
are
essentially
saying
it
doesn't
matter
like
that
random
pattern.
It
doesn't
matter
what's
here
it
we're
trying
to
dis,
extract
out
the
motion
component
only
and
that's
why
the
random
bit
patterns
work,
yeah,.
E
I
get
that
what
a,
but
in
the
example
where
you
had
the
moving
bar
of
you
know
where
the
the
bits
changed
right.
So
it's
it's
not
your
whole
flow
field.
It's
there's
a
specific
flow,
that's
happening
and
you're
in
it's
that's
being
detected
so
I'm,
not
saying
that
it
makes
it
an
object.
It
makes
it
a
region
of
coherence
from
which,
maybe
later
on,
you
can
say,
there's
an
object
to
it.
I.
B
Should
I
should
go
back
to
I?
I
didn't
have
to
read
that
caper
baby
paper
a
bit
more
carefully
again,
because
I
mean
the
thing
I
latched
onto
was
that
these
cells
respond.
When
there's
no
bond,
there's
nothing,
they
seem
and
many
cells.
We
seem
to
spawn
the
best
when
there's
no
features
at
all,
and
so
the
idea
that
you're
moving
a
bit
bar
through
a
bit
pattern.
You
know
to
ran
a
bit
bar
through
a
random
bit
field.
B
I,
don't
remember
if
that
was
an
important
thing,
or
was
that
sort
of
like
halfway
between
what
people
typically
do
and
what's
actually
going
on
so
I?
Don't
I,
don't
remember,
that's
actually
a
significant
thing
or
that
sort
of
like
say
yeah,
that's
what
you
would
expect
to
see
and
these
cells
really
just
prefer
to
have
motion
flow
altogether.
I
I
should
read
that
again.
I
I,
latched
on
to
the
part
I
wanted
to
latch
onto
which
was
these
cells
seem
to
prefer
no
pattern.
B
B
So
if
I'm
walking
forward,
then
there'll
be
diverging
flow
patterns
on
left
and
right,
I'm
walking
backwards,
they're,
converging
flow
patterns
if
I'm
sipping
sideways
I
get
a
different
pattern
than
if
I
turn
my
head,
and
so
the
idea
here
is
that
if
you
had
a,
if
you
had
a
field
of
flow
bits
that
you
put
them
through
a
spaceship,
Euler
you'll
end
up
basically
discovering
the
ways
you
move
through
this
field.
You're
discovering
like
oh
I,
am
I.
Have
many
cons
representing
going
forward?
I'd
really
come
back
with
that.
B
Many
comes
up
steam
train,
my
intellect
they
represent
many
concepts,
I'm
stepping
sideways
or
raising
up
and
looking
down
whatever
that,
whatever
motions
the
body
observes
it
takes,
you
will
discover
those
motions
as
some
some
combination
of
flow
bits
and
and
that's
that's
all
it's
trying
to
discover
it's
try
and
discover
the
column
itself
does
not
initially
generate
behaviors.
It's
just
trying
to
observe
what
behaviors
are
being
that
it's
observing
some
space
and
it's
trying
to
determine
what
behaviors
that
the
column
is
performing
in
that
space.
Again,
this
doesn't
have
to
work
on.
B
It
has
to
work
for
things
like
language
and
high-level
thought.
It
has
to
work
throughout
the
entire
amia
cortex.
So
the
theory
has
to
be
not
tied
to
specifically
vision
or
touch
or
hearing
it
has
to
work
with
us,
but
it
also
has
to
be
very
generic
model
of
how
a
Collins
is
I,
don't
know
what
I'm,
looking
at
but
I'm,
giving
a
little
bit
I'm,
giving
the
arms
a
little
bit
and
I'm
going
to
build
a
model,
all
right,
I'm,
sorry
I'm!
E
No
I
I,
don't
think
we're
misaligned
with
it
I.
You
know,
I
think
you
know
moving
beyond
the
fact
that
you've
got
the
entire
field
of
view
moving
and
selector
parts
of
it
moving
could
still
be
driven
by
the
same
underlying
mechanism.
It's
what
I'm.
What
I
was
interested
in
was
the
fact
that
you
might
be
able
to
have
the
ability
to
differentiate
that
the
flow
is.
You
know
in
this
direction
over
here
in
this
direction,
and
this
other
more.
B
Exactly
right,
yeah,
so
what
I
don't
want
to
do
is
associate
it
with
the
underlying
start,
object
there
or
the
I'm
drawing
feature
there,
but
you're
right,
I.
Think
in
general
this
is
a
very
general
purpose
idea:
I've
got
several
hundred
mini
columns,
they're
all
going
to
be
competing
in
a
spatial
cooler
like
way.
B
Yeah
in
the
abstract,
yes
yeah,
so
exactly
it's
still
done.
I
haven't
resolved
the
whole
issue
about
how
you
go
from
egocentric,
the
allocentric,
yet
there's
still
a
lot
of
questions
there,
but
it's
just
basically
I'm.
This
is
like
a
core
idea
about
how
space
and
represented-
and
you
know,
there's
tons
of
questions
still
about
it.
I
want
to
just
just
there's
a
few
more
things
in
this
paper.
Just
are
worth
mentioning
exist
like
I,
had
forgotten
some
of
these
things,
so
you
know
I
just
put
them
down
because
they
oh
yeah,
I
forgot.
B
The
stuff
like
interesting,
like
simple
cells
are
silent
before
input
they
just
they
don't
respond
at
all
to
get
in,
but
but
complex
cells
have
a
resting
firing
rate
as
a
general
rule
yeah
and
says:
oh,
that's,
pretty
interesting,
I'm
arguing
that
complex
cells
are
representing
motion,
and
this
this
least
gives
you
the
possibility
that
they
can
represent
to
motion
in
two
directions.
You
know
you
can
additive
and
subtractive
I'm,
not
saying
that's
true,
but
it's
interesting
idea:
the
simple
cells,
none
of
them
responded
the
noise
fields,
the
meaning
that
none
of
them
responded.
B
These
motion
ideas,
so
it's
gotta,
be
the
whole
field
or
a
bar
moving,
neither
hopefully
the
bar
actually
I,
don't
know
what
I
meant
by
that
comment
anyway.
They're
never
responded
the
noise
village,
of
course
we
know
they
have
small,
represent
our
s,
which
is
consistent.
This
is
in
layer
truth.
This
is
layer
for
not
only
the
simple
selves
made
for
small
or
apps,
which
is
consisting.
It's
you'd
want
this.
Without
that
to
represent
that
orientation,
that's
different
than
haps
the
moving
orientation.
B
They
were
primarily
monocular,
yet
sharp
orientation
tuning.
They
also
responded
the
flash
lines
you
know
either
like
you
flash
line,
if
you
said,
would
respond
on
on
for
the
line
offset
align.
They
also
responded
mostly
motion,
mostly
in
either
direction,
so
that
it's
a
little
bit
mysterious
they're,
not
all
responding
the
motion
and
they
respond
when
you
don't
have
Marsh
inside
that
may
be
some
sort
of
complex
cells.
B
Have
this
resting
bright
firing
rate,
they
all
respond
at
the
noise
fields
and
in
all
cases
it
was
more
important
than
the
bars
and
some
only
some
respond
to
the
noise
field.
Some
only
responded
in
ratio,
so
something
like
you
could
show
moving
bars.
They
just
went
fire
up
and
some
of
the
grading
they
did
they've
large
RF
I,
don't
know
what.
Why
would
I
guess
most
of
you.
It
was
more
binocular,
three
cousin,
binoculars,
I
guess
it
makes
sense.
If
you're
trying
to
detect
motion.
B
Oh
here
here
the
Sun
had
different
orientation
preferences
for
her
noise
or
versus
the
bar,
so
that
is,
if
you
show
a
moving
bar,
some
cells
respond,
be
sure
the
motion
they
wouldn't
respond
in
the
other
way
around
and
then
it
said
most
of
directionally
sensitive,
but
exclusively
so
in
layer
five,
so
they
found
some
cells
were
not
rexis.
Complex
cells
would
not
try
four
senses,
but
I.
Don't
know
why,
but
in
layer
five
they
were
all
directly
sensitive.
So
it's
just
consistent
with
their
five
representing
movement
and
down
here.
B
This
is
a
quote
about
unknowns
about
how
these,
how
the
motion,
detecting
cells,
are
they
getting
the
exact
same
input?
There's
a
question
is:
where
is
it
slow
detected
is
detected
in
the
retina?
Is
it
detected
in
the
thalamus,
or
is
it
checking
the
cortex
and
they're
just
they're?
Just
talking
about
that
question?
Yeah?
Okay,
that's
it
for
me.
Today,
I.
D
One
thing
that
has
always
bothered
me
with
our
columns
Plus
model,
the
one
with
layer,
four
layer,
six
back
and
forth
is
that
those
layer,
six
cells
are
often
characterized
as
being
simple
cells,
not
complex
cells,
the
the
layer,
six,
a
ones
and
they
received
the
lamech
input
and
and
there's
just
these
various
things
we
wouldn't
expect
if
they
were
a
pure
location
and
I
noticed
that
in
your
slide
here,
when
you
were
showing
where
the
salamis
products
u---boat,
you
kind
of
you're
characterizing
it
as
being
lower
lower
layer,
five
admit
a
little
bit
of
layer.
D
B
D
I,
remember
all
right,
so
I
guess
I'm
bringing
up
that
you're.
Some
of
the
things
you've
said
I've
found
that
maybe
we
have
to
flip
it.
Maybe
five
is
doing
some
of
the
things
that
we've
been
saying.
Six
is
doing.
Maybe
six
is
doing
some
of
the
things
we're
saying.
Five
is
student
yeah
based
on
where
I
I
have
not
surveyed
all
the
literature
very
closely
I'm
a
lot
of
this
I'm
just
getting
straight
from
the
Thompson
later
six
paper
and
and
that
that
paper
paints
a
story.
D
B
B
I'm,
not
sure
I
do
when
skimmers
I'm,
not
sure
my
cameras,
so
remember
the
Gilbert
paper
I
talked
about
a
few
weeks
ago.
The
one
yeah
this
guy
here
was
this
thing.
I
mean
this
there's
all
kinds
of
conflicting
stuff
going
on
here.
So
here,
if
you
look
at
layer
six,
these
are
the
simple
cells
in
layer
six
and
they
show
up
a
handful
of
simple
cells
in
layer
six.
But
when
you
go
when
they
go
to
to
the
complex
of
this
show,
you
know
they
show
many
times
more
complex
cells
in
layer.
B
Six,
so
I,
don't
know
what
to
make
of
that.
It's
just
like
okay,
I
can't
explain
the
simple
cells
right
now,
but
but
there
are
a
lot
more
complex
cells,
so
I
mean
I,
don't
know,
there's
a
lot
of
wonky
stuff.
We
have
to
try
to
figure
out
here,
but
I
guess
I'm
excited
that
this
new
hypothesis
being
supported
not
eliminated
by
these
findings,
so
I
think
you
know
Marc
is
it's
sort
of
like
when
we
think
about
in
Toronto,
Port
Jackson.
B
Let
me
see
now
we
see
you
know,
border
cells
and
itself,
and
this
console
not
gonna
sell
and
you
know
better
cells.
This
complex
milieu
of
response
properties
out
of
discovery,
and
we
might
find
the
same
thing
going
on
here
in
cortex.
We
don't
really
know
why
it
so
I,
don't
know
it's
it's
clearly,
not
as
simple
Oh.
D
B
I'm
saying
the
Thompson
paper
she
talks
about,
you
know:
there's
two
sets
of
layer.
Six
thousand
did
two
different.
You
know
layer
4
in
layer,
five
was
at
left
four
layers,
so
we
have
to
work.
I
mean
I
still
think,
as
I
said
earlier
right.
This
is
idea
that
we
might
be
doing
composite
objects
right
in
the
mini
column,
right
in
the
column
and
there's
a
whole
bunch
of
stuff
going
on
there
right
and
I.
E
B
B
Well,
we
do
know
that
there
are
long-range
connections,
not
just
between
many
comms.
Remember
many
comms
are
really
really
skinny.
They're,
like
you
know,
30
40,
50,
mic
Lendl
either.
Oh
I
read
another
paper,
just
a
little
slide.
I
can't
remember
if
you
don't
you
like
to
find
it
more
recent
paper
where
they
were
using
two
photon
imaging
to
try
to
find
out
how
precise
feature
RF
changes
occur
in
the
cortex
like
like
there's
this
idea
that,
as
you
move
across
the
cortex,
maybe
your
maybe
your
orientation
just
changing
slightly.
B
You
know
gradually
and
there's
a
bunch
of
overlap,
or
maybe
it's
very,
very
precise
and
what
they
showed
and
that
that
thing
is
that
it's
super
precise,
like
one
cell
wide
they've,
got
this
response
and
the
next
LOI
over
it.
You've
got
a
different
response
and
it's
consistent
all
the
way
down,
meaning
they
can
then
go
down
to
multiple
layers
and
see
this
one
singular
cell
differentiation
during
that
very
very
precise,
like.
C
B
You
know
you
know
remember,
for
in
had
came
back
from
cosine
I
think
it
was
in
and
he
had
that
he
pointed
out.
Those
is
some
new
papers
related
to
the
rodents
that
were
showing
they
were
saying
they
were
arguing.
That
actually
is
this
organization
and
and
people
hadn't
observed
it
I
need
you
to
find
those
papers
again.
I
ripped
them
quickly,
but
he
sent
references
out
to
it's
a
slack
completely
it
is
it
archived.
It
was
my
be
able
find
that
isn't
a
depends.
How
far
back
you
go.
B
B
Class,
this
is
Lauren,
came
back
from
cosine.
He
said
there
was
this
presentation
about
this,
and-
and
you
should
Jeff
you're
gonna-
be
excited
about
this,
because
it's
showing
that
there
is
this
organization,
people
weren't.
The
argument
was
that
people
weren't
looking
precisely
enough
I
mean
I
and
and
and
he
gave
some
links
to
some
papers.
No
one
else
remember
this
yeah.
D
B
C
So
both
layer,
2,
3
and
layer,
5
6,
have
complex
and
simple
cells,
and
this
interesting
thing
you
know
you
probably
know
Jeff
they.
They
have
like
a
sort
of
different
evolutionary
origin
sort
of
layer.
2
3
was
formed
after
layer,
5,
6
yeah,
so
like
the
horizontal
connectivity
between
so
like
the
connectivity
between
layer,
5
6,
between
deep
layers
between
South
and
deep
layers
is
denser
than
across
layers,
meaning
that
it's
sort
of
like
it's
almost
like
there's
two
computational
units
and
there's
like
less
talk
between
them
and
there
is
within
them.
C
B
B
Like
cells
with
a
special
type,
complex
cells,
okay,
some
I
can't
I
couldn't
see
how
what
the
hell
over
there
drummer
and
here's
the
mapping
of
the
special
type
complex
cells-
and
these
are
very
precisely
only
only
in
layer,
five
and
layer-
three,
that's
it
the
only
place
to
show
up.
They
don't
show
up
anywhere
else,
and
so
there's
there's
another
cut
again.
There's
there's
variations
in
this.
It
sort
of
reminds
me
of
like
when
say
Oh
in
hippocampal
complex.
B
Where
do
you
find
or
in
the
orientation
head
direction
to
himself,
but
you
find
them
everywhere
and
and
there's
all
kinds
of
mixtures
of
cells
you
find
you
can
find
pure
head
Direction
cells
and
you
can
find
all
kinds
of
cells
that
have
dry
head,
Direction
properties
and
I.
Think
something's
probably
happening
here
too,
but
at
least
there's
this
one
class
of
cells
which
they
were
able
to
differentiate
with
a
purely
layer,
five
and
surely
layers
three
and
and
they
have
certain
pockets.
So
I
the.
B
But
that's
another
thing
that
people
I
think
stopped
referring
to
it's
kind
of
a
complex
definition.
I
couldn't
follow
initially,
but
they
they
have
a
way
of
this
differentiating.
You
saw
some
I,
the
other
one
interests
and
I.
Think
it's
just
one
more.
Those
complexities
ever
lost
the
group
in
Egypt.
He.
E
B
I
haven't
found
evidence,
so
imagine
you're,
a
layer,
five
cell,
that's
getting
directly
input
from
the
talents
and
you're
gonna
create
a
comp,
a
large,
complex
off.
Okay,
we
don't
know
how
that
happens.
The
easiest
way
for
me
to
imagine
happen
is
that
there's
a
much
wider
convergence
of
flow
bits
until
they
have
five
cells.
Then
there
are
two
non
flip.
It's
until
they're
four
cells
that
is
lay
five
cells
are
just
looking
at
a
larger
part
of
the
retina
and
I.
B
Think
the
evidence
would
that
idea,
because,
because
he
sells
respond
very
rapidly,
it's
not
like
it's
some
processing
that
occurs
and
then
it
shows
up
yourselves
bingo
they
get
just
like
biliary
cells.
Complex
cells
can
appear
without
any
input
from
layer
four.
So
where
would
that
contraire
that
that
consolidation
occur
I,
don't
think,
there's
many
places
it
could
occur.
B
Once
asked
me
several
times,
as
Murray
shrim
in
this
question,
I
said:
are
the
same
cells
that
project
to
layer
for
the
exact
same
cells,
the
projected
layer,
two
three
until
they're
five
six
and
he
said
he
thinks
they
were,
but
that
is
not
inconsistent
because
you
would
find
some
cells
projecting
the
both,
but
you
might
be
some
other
cells
that
don't
so
I
I.
Don't
know
he
wasn't
certain
about
that.
He
thought
so.
So
my
my
first
hypothesis
would
be
it's.
The
simple
one
is
that
that
a
column
gets
a
wide
input.
B
B
Okay,
well,
it's
a
simpler
explanation,
so
there's
something
they
said
for
that
it
would
be
if
I
had
no
other
evidence.
I
would
say
that's
more
likely
than
a
more
complex
explanation.
It's
very,
very
simple!
To
achieve
that,
you
just
run.
If
you
just
don't
decide
our
standard
spatial
pool
or
you
just
need
to
have
a
wider
divergence.
That's
it.