►
Description
In the continuation of last week’s research meeting, Subutai Ahmad explains voting in the Thousand Brains Theory.
In this research meeting, he explains how inference is much faster with multiple columns as columns share information through long-range sparse connections to agree on what the object is. He goes over the simulation results we presented in our “Columns” paper, and shows that as the number of cortical columns increases in the network, the number of touches to recognize an object rapidly decreases, making inference much quicker.
Finally, he talks about how the Thousand Brains Theory rethinks the notion of hierarchy in the neocortex. Instead of the classic view of using hierarchy to assemble features into a recognized object, the theory states that the neocortex uses hierarchy to vote across levels and sensory modalities, and rapidly reach consensus on the objects being sensed.
Columns paper "A Theory of How Columns in the Neocortex Enable Learning the Structure of the World": https://www.frontiersin.org/articles/10.3389/fncir.2017.00081/full
Part One: https://youtu.be/XUpmN_CLOZc
- - - - -
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://numenta.com/news-digest/
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
A
When
we
talk
about
what's
going
on
within
a
cortical
column,
it's
sometimes
hard
to
get
one
thing
across
without
describing
you
know
a
lot
of
the
other
stuff,
and
I
think
we
had
a
pretty
good
discussion
last
week
on
some
of
that
and
people
were
very
willing
to
ask
dumb
questions
and
stuff,
and
I
hope
keep
please
do
that,
because
it
was
clear
that
there's
a
lot
of
different
things
that
you
know
jeff
and
I
and
marcus
and
others
kind
of
take
for
granted.
But
you
know
others
may
not
know
that
stuff.
A
We
had
this
thought
experiment
where
you
stick
your
hand
in
a
box
and
you're
trying
to
figure
out
what's
inside
that
box
in
the
first
situation,
you're
trying
to
recognize
the
object
with
a
single
finger,
that's
by
moving
that
finger
around
and
then
the
second
task
is
recognize
the
object
with
multiple
fingers.
Where
you
know.
Let's
say
you
stick
your
whole
hand
in
and
you
can
grasp
the
entire
object
and
so
on.
So
we
went
through
kind
of
the
basic
single
cortical
column
model.
I
think
so.
You
have
sensory
input.
A
We
talked
about
the
sequence
memory
where
you
have
sensory
input
coming
in,
and
you
have
the
recurrent
connections
on
the
distal
basal
dendrites,
which
act
as
predictions
for
the
for
our
cells
and
then
say.
Well,
you
could
take
the
exact
same
model
by
putting
in
a
location
signal
or
a
motor
driven
signal
that
can
act
as
context
for
the
sensor
for
the
sensory
input,
and
that
way
you
can
represent
sensory
features
at
specific
locations
and
essentially
it's
the
exact
same
code
base.
A
That's
used
for
the
sequence
memory,
and
that
means
it's
the
exact
same
neuron
model.
That's
used
to
do
sensory
motor
sequences,
and
so
the
basic
hypothesis
of
the
paper
was
that
by
adding
this
location
signal,
you
could
have
a
layer
of
cells
that
can
predict
its
input
as
the
sensor
is
moving
around
and
in
this
paper
we
didn't
actually
describe
specifically
what
the
location
signal
might
be.
We
did
the
you
know.
A
We
talked
about
that
a
lot
later
and
then
the
columns
plus
paper
proposes
grid
cells
as
a
mechanism
for
it,
but
for
the
purpose
of
this
paper,
you
can
just
think
of
this
as
just
an
sdr
that
represents
the
allocentric
location
of
the
sensor
on
the
object.
That
is
something
like
a
just
a
coordinate.
That's
representing
you
know
the
location
of
the
feature
in
the
reference
frame
of
the
object
itself.
A
So
the
basic
network
had
this
sequence,
memory
like
structure
below
here
and
as
you're
moving
you
get
different.
You
know
look
at
look
at
features
at
different
locations
that
end
up
being
activated
here.
The
second
layer
is
an
object
layer
that
this
is
a
layer
that
pools
over
different
sensory
inputs
that
belong
to
a
feature
and
that,
basically,
this
layer
is
a
lot
more
stable
than
the
one
below
this
one.
The
one
below
is
changing
with
every
movement
and
the
one
on
top
is
stable.
As
long
as
you're
sensing,
a
particular
object.
A
A
We
have
a
location
signal
coming
in
that's
going
to
predict
where,
on
the
object's
reference
frame,
we're
going
to
touch
so
based
on
this.
This
is
a
context
signal
you'll,
get
a
bunch
of
predictions,
that's
showing
up
in
this
temporal
memory,
layer,
the
input
layer
and
then,
when
you
actually
touch
the
object,
you
get
the
actual
features
coming
in.
A
It's
going
to
overlap
with
some
of
the
predicted
features
and
you'll
get
these.
These
purple
squares
here
represent
cells
that
were
correctly
predicted
when
you
got
the
sensory
input
coming
in
that
information,
then
is
sent
to
the
output
layer
and
you
will
get
because
this
layer
has
pulled
over
all
the
possible
objects.
A
B
A
So
you
got
a
a
bunch
of
here.
You've
got
a
union
of
sdrs
and
let's
say
it's
representing
at
this
point:
three
possible
objects,
this
tennis
ball.
This
cook
can
and
this
coffee
cap
are
all
consistent.
This
layer
could
have
learned
dozens
or
hundreds
of
objects,
but
these
are
the
only
three
that
are
consistent
with
this
sensory
input
at
this
location
and
now
you're
moving
to
the
next
location
as
just
as
you're
moving
you,
you
have
a
different
predicted
location
that,
in
turn,
is
going
to
cause
predictions
in
the
sensory
layer.
A
Now
this
this
set
of
active
cells,
then
propagates
to
the
output
layer
and
what's
going
to
happen
now,
is
each
of
these
cells
in
the
output
layer,
they're,
getting
lateral
support
from
other
cells
that
represent
the
same
object
that
it
represents
it's
getting
lateral
support,
and
now,
when
the
input
comes
in
the
cells
that
have
both
lateral
support
and
feed
forward,
input
are
going
to
be
the
ones
that
are
most
strongly
activated
and
those
are
the
ones
that
are
going
to
win
in
the
winner
take
off
here.
Okay,
so
that's
the
essence
of
the
algorithm.
A
That's
winnowing
down
on
the
set
of
hypotheses
so
based
on
this
feature.
Now
you
can
rule
out
the
tennis
ball,
but
you,
but
both
the
can
and
the
coffee
cup
are
still
consistent
and
then,
finally,
as
soon
as
you
move
to
the
handle
here,
exact
same
thing
happens
again,
the
sensory
feature:
that's
at
these
locations
is
sent
up,
and
now
only
the
coffee
cup
is
consistent
with
the
past
input,
as
well
as
the
current
feed
forward.
Input.
A
That's
coming
in
okay,
again
within
this
cell
within
this
layer,
are
all
these
cells
that
are
getting
lateral
support
from
the
other
cells
that
represent
the
same
object
and
that's
those
cells
are
assumed
to
be
more
persistent
and
then,
when
they
get
feed
forward
input.
The
ones
that
have
the
strongest
input
are
the
ones
that
that
stay
stay
on.
C
C
C
C
A
Yeah,
so
this
path,
this
location
signal
here
is
really
a
pathway,
and
so
you
might
have
let's
say
a
couple
of
thousand
cells
here.
Some
sparse
subset
of
this
is
active
and
all
of
the
neurons
in
this
input
layer
have
access
to
this
set
of
inputs
and
through
learning.
What's
going
to
happen
is
that
the
ones
that
just
like
in
the
in
the
temporal
memory,
the
cells
are
going
to
make
connections
to
the
active
location
inputs
as
you
get
sensory
input?
So
when
you
get
during
learning,
what
will
happen?
Is
you
get
sensory
input?
A
A
Doesn't
have
to
connect
to
all
of
them,
it's
just
some
subsample
yeah
and
the
sdr
math,
you
know
shows
why
that's
possible
and
then
what
happens
is
the
next
time
you
get
that
exact,
same
location
signal
now
you
do
have
a
prediction,
because
some
dendritic
segment
is
going
to
recognize
that
pattern,
and
so
you'll
get
a
prediction
on
those
cells.
So
when
the
feature
does
come
in
that
particular
cell
is
going
to
win.
B
Yeah
yeah
exactly
yeah
each
neuron
doesn't
receive
the
location
signal.
I
mean
you
can
just
imagine
physically
the
location
signals.
A
bunch
of
axons
are
just
sort
of
streaming
through
these
neurons
and
they
don't
not.
They
don't
make
connections
or
even
potential
connections.
All
the
neurons,
so
the
the
each
neuron
in
the
layer
just
has
to
figure
out
what
it
can
connect
to
and
there's
a
there's
a
whole
set
of
neuroscience
behind
that.
So
it's
not
like
every
cell's
getting
the
stuff.
B
It's
like
the
in
the
biology
they're
trying
to
make
connections
to
a
subset
that
works
for
them
and
not
all
the
connections
can
be
made
in
our
software.
You
can
do
anything
you
want.
You
can
have
every
neuron
be
able
to
connect
that
every
input
action-
that's
active,
but
in
biology,
there's
a
physical
constraint
to
this,
because
not
every
axon
can
reach
every
neuron
at
the
right
place.
C
D
A
A
A
A
B
So
super
guy
had
a
whole
page
on
this
map.
In
this
paper.
The
2007,
I
think,
was
in
that
paper
just
describing
this
and
lining
it
up,
and
you
haven't
mentioned
it,
but
as
he
said
it's
really.
It's
really
nice,
because
in
the
biology
it's
it's
somewhere
around,
you
know
15
to
30
20,
you
know,
synapses
are
sufficient
or
typically
can
be
formed.
B
E
B
B
So
you
know
there's
no
universal
principle
that
says
biology
has
to
do
this.
It's
just
it
matches
what
the
theory
says
and
and
the
biology
is
just
because
you
have
these
little
leaky
integrators
and
the
synapses
have
to
be
close
enough
so
that
they
they
multiply
as
opposed
to
if
they're
further
apart
on
the
dendrite
here,
they
give
it
this
way
a
dendrite
can
have
about
one
synapse
per
micron.
B
A
E
A
Yeah,
I
also
think
there's
potentially
evolutionary
pressure
to
get
to
as
small
a
number
as
possible
without
losing
accuracy,
because
there
are
metabolic
costs
to
having
too
many
synapses
and
large
dendrites
and
all
that,
and
so
it's
not
surprising
to
me
in
some
sense
that
evolution
would
go
towards
a
really
minimal
system.
As
long
as
he's
still
accurate
enough
yeah.
B
E
A
Okay,
now
we
can
finally
move
to
actual
voting,
so
we
just
went
through
the
single
column
version
of
the
algorithm,
this
two-layer
algorithm
and
now
imagine,
there's
multiple
cortical
columns
here,
we're
showing
three
cortical
columns.
Each
cortical
column
is
getting
a
completely
different
sensory
input.
Coming
in.
Let's
say
these
three
fingers.
A
And
then
each
cortical
column
is
doing
something
similar
to
what
I
I
just
mentioned
there.
Each
one
is
getting
a
location,
that's
on
the
object.
Now.
The
key
thing
here
is
that
each
coraco
column
is
getting
a
different
location
signal,
corresponding
to
the
sensory
input
that
it's
responsible
for
so
depending
on
where
this
finger
ends
up
on
the
coffee
cup.
A
This
is
the
root
finger
middle
yeah,
so
you
get
three
different
location
signals
coming
in
okay
and
they're
they're
in
in
principle,
they're
completely
independent
in
in
the
case
of
a
hand,
there's
some
constraints
between
them,
but
think
of
it
as
being
completely
independent
locations
coming
in
and
then
the
big
other
difference
is
now.
You
have
these
long-range
lateral
connections
that
cut
across
the
cortical
columns.
A
These
are
also
going
on
the
distal,
basal
dendritic
segments
of
these
cells
and
the
way
these
long
range
signals
work
is
exactly
the
same
as
what
I
described
previously
for
within
a
single
column.
What's
going
on
with
the
lateral
connections,
but
conceptually
what's
happening
is
that
each
each
column
is
getting
votes
from
its
neighbors
as
to
what
for
each
of
the
hypotheses
that
are
consistent
with
the
sensory
inputs
that
column,
one
has
seen
it's
going
to
send
those
signals
to
column
two
column.
A
Two
is
also
gonna
get
all
of
the
hypotheses
that
are
consistent
with
what
column
three
has
seen
so
far.
It's
going
to
get
that
transmitted
here
through
these
long-range
connections
and
essentially
the
cells
that
have
the
most
votes
for
specific
objects.
Those
are
the
ones
that
are
going
to
be
persistent
on
there.
Okay.
So
now
it's
not
just
the
information
that's
received
over
time.
It's
also
the
information
it's
getting
currently
from
the
lateral,
the
columns
that
are
on
each
side
of
it.
A
And
just
as
a
reminder,
each
column
here
has
partial
knowledge
of
the
object,
because
it's
just
getting
a
small
part
of
the
overall
sensory
inputs
that
the
animal
is
getting.
The
long
range
connections
allow
the
columns
to
vote
and
then,
of
course,
now
inference
is
going
to
be
much
faster
with
multiple
columns.
A
So
here's
here's
a
3d
hand
grasping
three
points
on
the
on
the
cup
and
each
one
of
them
is
going
to
get
a
location
signal
coming
in
that's
going
to
cause
predictions
in
each
of
those
cortical
columns.
Each
one's
then
going
to
get
sensory
input.
That's
coming
in
so
specific
sets
of
cells.
Here
are
going
to
become
active
based
on
the
prediction
and
then
those
are
going
to
get
sent
to
the
output
layer
here
and
now.
A
Each
cortical
column
at
this
point
in
time
is
going
to
have
a
different
set
of
hypotheses,
depending
on
what
it
sends.
So
this
column,
one
its
input,
is
consistent
with
either
a
tennis
ball
or
a
coffee
cup
column.
Two,
it's
consistent
with
the
soda
can
and
the
coffee
cup
and
same
thing
with
the
column
three.
A
But
if
you
notice
there's
only
one
object,
that's
actually
consistent
with
all
three
inputs
coming
in
at
all
three
columns.
So
now
you
get
the
voting
thing
happens
and
then
what
will
happen
is
each
particle
column
is
going
to
get
input
from
its
neighbors
and
it's
going
to
win.
The
ones
that
are
going
to
win
out
are
the
ones
that
are
consistent
with
all
of
the
evidence.
Current
data
in
this
case
in
one
graph,
it
can
recognize
the
object
uniquely,
even
though
no
single
cortical
column
here
has
enough
information
to
uniquely
determine
the
object.
F
A
A
No,
no
not
fully
connected
they're,
just
axons
that
are
going
across
just
like
every
other
layer
here,
they're
just
axons
that
are
going
across
our
dendrites
that
are
going
across
and
you
each
dendrite
will
make
connections
to
this
specific
set
of
inputs
that
are
consistent
with
you
know.
It's
expensive.
B
So
here's
so
here's
something
that's
really
interesting
about
this.
If
you
just
think
about
one
column,
there
has
to
be
sufficient
connectivity
within
one
column,
so
that
that
column
on
its
own
can
eliminate
possibilities
when
it's,
if
it's
acting
as
a
single
finger,
something
like
that,
so
there
has
to
be
a
reasonable
amount
of
voting
going
on
in
a
single
column,
but
between
columns.
It
can
be
very,
very
sparse.
B
So,
but
those
connections
can
be
extremely
sparse.
You
can
think
of
it.
This
way,
I
imagine
you're
just
one
column,
and
I
want
this
this.
I
need
to
have
this
voting
mechanism
work.
Well,
the
cells
in
my
one
column
say
column
two
here
they
have
to
form
enough
connections
with
the
other
cells
in
column.
Two,
so
that
I
have
the
context.
B
B
All
that
matters
that
I'm
able
to
sample
a
you
know,
let's
say
20
synapses
from
all
the
other
cells
that
might
be
active,
someplace
that
are
part
of
this
representation
and
it'll
work,
and
if
the
different
columns
do
this
randomly,
you
could
have
a
very,
very
sparse
connectivity
and,
of
course
that's
actually
what
we
see
we
don't
see.
We
see
a
lot
of
axons.
Let's
say
this
is
in
layer
two
three.
B
We
see
a
lot
of
axons
going
long
distances
in
the
cortex
within
layer,
two
three
they
can
be
across
to
the
other
side
of
the
brain.
We
can
see
connections
from
the
left
hand,
to
the
right
hand,
you
see
those
you
see
them
from
primary
visual
cortex
to
primary
auditory
cortex,
but
these
tend
to
be
very
sparse.
It's
not
like
this.
You
know,
there's
a
lot
of
axons,
but
the
actual
the
number
of
connections
I
might
make
between
one
column
and
the
visual
cortex
and
one
column
in
the
auditory.
B
F
Yeah
yeah,
that
makes
makes
a
lot
of
things
clearer.
Definitely
I
I
think
it
also
kind
of
answers.
My
next
question
of
who
counts
the
vote.
So
if
I
understand
you're
right,
it's
basically
there's
no
higher
level
process.
That
then
counts
the
vote,
but
it's
more
all
the
neurons.
That
kind
of
recognize
cups,
looks
at
its
connections
and
says:
okay.
I
think
it
might
be.
B
A
A
E
B
Yeah
no
one's
in
charge,
just
like
no
one,
was
in
charge
in
a
single
column.
As
soon
as
I
made
the
point
that,
what's
going
on
in
the
single
column,
is
the
same
algorithm,
that's
going
on
across
15
columns
and
that's
absolutely
true.
So
within
a
single
column,
nobody
has
to
tell
it
it's
just
that,
there's
a
there's,
a
competitive
algorithm
going
on
and
and
whoever
gets
support
from
bottoms
up
in
context.
B
If
those
are,
the
ones
are
going
to
stay
active
and
if
you
didn't
get
those
two
supports
you're,
just
you're
going
to
lose
out
and
then
it'll
naturally
winnow
down
to
the
right
answer,
if
not
immediately
through
then
through
multiple
sensations
and
all
we're
doing.
When
you
do,
multiple
columns
is
you're,
basically
extending
the
size
of
the
the
object
layer,
you're
just
making
the
object
layer
conceptually
much
much
larger,
but
you
can.
Any
cell
can
still
settle
down
if
it
just
it
needs
to
get
lateral
support
and
feed
forward
support,
and
it's
good.
A
A
What's
happening
in
here
too,
if
you
look
at
the
sensory
input,
those
are
also
pyramidal
neurons
with
active
dendritic
segments
and
the
ones
that
have
support,
in
this
case,
from
a
location
signal
when
that's
the
context,
the
ones
that
have
the
support
from
location
signal
are
the
ones
that
are
going
to
win
out,
whereas
the
other
ones
will
not.
So
that's
how
you
get
these
unique
kind
of
cells
becoming
active.
So
what's
going
on
in
this
layer
below
is
very,
very
close
to
what's
going
on
here.
E
So
I
have
a
question
about
the
temporal
component,
so
if
I
grasp
I
can
get
a
certain
degree
of
synchrony,
if
I
go
one
two
three,
I
still
unable
to
recognize
it.
So
what
is
I
I
don't
want
necessarily
the
thing
fire
sends
a
single
spike
and
I'm
looking
for
them
all
kind
of
coordinating.
What's
the
integrative
mechanism
that
allows
me
to
not
miss
spikes,
you
know
coming
you
know
at
slightly
different
times
and
not
necessarily
building
up
a
coherent
pattern
on
whatever
you're.
B
A
A
If
it's
not
okay,
if
you
look
at
this
act,
if
you,
if
you
look
at
a
single
active
dendritic
segment
in
order
for
it
to
fire,
all
of
the
synapses
have
to
become
active
within
a
few
milliseconds
of
each
other,
and
so
that
means
all
of
the
output
from
the
other
coracle
column
has
to
be
arriving
within
a
few
milliseconds
if
it's
going
to
initiate
a
dendritic
spike,
and
so
how
is
that
synchronization
happening?
Is
that
kind
of
the
essence
of
your
question.
E
A
Yeah,
I
know
but
dendritic
segments
you
have
to
have
these
spikes
within
a
few
milliseconds
of
each
other
and
there
are
mechanisms
we
which
we
didn't
discuss
here
for
synchronizing
groups
of
cells.
You
have
these
oscillations
and
stuff
and-
and
you
know
we
speculate
in
the
past
one
of
the
reasons
to
have
that
is
so.
You
can
have
these
like
really
tight
synchrony
across,
but
I
would.
B
Say
I
would
say
overall
we
don't
know
the
answer
to
that
question
kevin
which
is
it
or
if
it's
both
I
mean
the
there's
evidence
that
subatai
said
that
hey
you
know
these
these.
These
action
potentials
have
to
arrive
on
a
dendrite,
pretty
close
together
in
time,
but
I'm
not
sure
that's.
We
know
the
answer.
There's
a
lot
of
stuff,
that's
unknown
here.
I
think
we
can
say
that
this
voting
mechanism-
I
I
feel
comfortable,
saying
this
voting
meccan
is
happening
in
the
brain.
B
This
is
very,
very
clear,
but
there's
a
lot
of
details
about
it
that
we're
you
know
we
can
only
read
so
much
in
the
empirical
literature
and
you
can't
trust
everything
and
there's
some
conflicting
things
and
there's.
Sometimes
these
systems
seem
to
be
persistent.
Sometimes
they
don't,
and
so
I
I
get
uncomfortable
saying
I
know
the
answer
to
this
question.
E
Yeah,
so
I
was
I
was
you
know
I
earlier,
based
upon
the
last
research
meeting
that
we
had
on
this
thing.
I
asked
subatai
what
what's
the
conditions
under
which
a
spike
train
is
is
generated,
and
so
that's
another
way
of.
If
you
can't
make
it,
you
know,
you
know,
if
you
want
to
create
a
window
in
which
you
know,
there's
a
greater
chance
of
synchron
overlap
between
things.
You
might
generate
a
spike
train,
so
at
least
you
know,
there's.
B
B
We
do,
and
we
do
know
that
we
do
know
the
synchrony
thing.
You
know
the
research
that's
been
done
on
grid
cells,
for
example,
is
really
really
fascinating
because
it
shows
in
the
good
cell
mechanisms.
Not
only
is
it
is
it
timed
with
the
end
of
you
know,
with
the
rhythm
of
the
of
of
a
you
know,
the
time
with.
Let's
say
it's
a
theta
rhythm,
but
it's
it's
timed
with
specific
phase
of
the
theta
rhythm,
like
you
know
the
different,
and
so
it's
even
it's
more
precise
than
just
you
know.
B
Oh
on
the
peak
of
the
theatricism,
though
it's
it's
like,
oh
at
the
90
degrees,
with
the
faderism
and
everything
right.
So
we
now
we
know
that
neurons
have
this
ability
to
very
precise
timing
relative
to
rhythms
and
so
that
that's
a
fact
and
whether
they're
doing
that
everywhere.
It's
interesting
question
we
don't
know,
but
I'm
so
I'm
I'm
guessing
it's
going
to
be
ignorance.
The
terms
of
neural
mechanisms.
We
have
are
really
not
very
blunt.
They
tend
to
be
very
precise
and
people.
Think
of
them
as
blunt,
but
they're
not.
B
B
B
E
B
B
Synchronize
yeah
yeah,
I'm
not
saying
that's
happening
here,
I'm
just
we
know
that
neurons
are
capable
of
that
and
also
there's
evidence
that
there's
sort
of
a
memory
trace
for
individual
synapses
that
that
one
individual
synapse
becomes
active,
the
dendrite
knows
it
happens
for
a
while.
So
there's
a
lot
of
stuff
going
on.
D
B
B
Very
you
know
given
the
the
world,
and
so
it
forms
all
these
connections
which
get
lost
and
then
later
on,
you
won't
be
able
to
form
them
very
easily.
So
there's
a
pruning
thing.
We
didn't
model
that
also
their
individual,
axons
and
dendrites
are
always
growing
they're
like
little
little
worms
that
are
sort
of
sticking
their
nose
out
in
different
places
and
trying
to
form
new
connections
and
retracting
if
they
don't
find
them,
and
so
there's
all
this
kind
of
growth
that
goes
on.
B
We
modeled
that
in
a
very
simple
way,
we
just
said:
oh
well,
we'll
just
form
connections
when
we
need
them
and
we
won't.
We
don't
have
to
worry
about
the
physical
constraints
of
the
neural
tissue,
so
we
do
model
forming
a
new
synapses,
but
we
didn't
have
to
model
these
other
things
that
are
actually
going
on
in
the
real
brain.
There's
no
point
in
modeling
them
did,
were
you
going
to
say
something
different
everything.
A
Yeah,
I
can
talk
about
the
specific
learning
rules
that
we
did
implement
because
those
are
all
unsupervised
learning
rules.
You
know
very
similar
to
the
temporal
memory,
one
where
and
what
we
talked
about
with
dendrites
last
time.
I
don't
know
if
that
would
be
helpful
or
not,
but
it
doesn't
model
the
stuff.
You
just
talked
about
jeff
of
the
initial
profusion
of
growth
and
things
like
that
yeah,
but
we
did
have
a
very
simple
set
of
learning
rules
for
each
neuron
and
how
it
made
connections
to
the
to
the
active
cells.
D
A
D
A
No,
no,
no,
no!
No!
No
yeah!
It's
different!
It's
not
like
we're
setting!
Well,
it's
not
like
a
deep
learning
system
where
you
set
up
a
random
set
of
connections
and
then
through
back
propagation,
you
know
tune
each
wait
until
you
get
to
a
final
set,
it's
very
very
fast.
It's
when
a
cell
becomes
active,
it
looks
around
to
see
what
other
cells
are
active
nearby
or
whether
the
axons
are
active
nearby
and
it
proactively
forms
connections
to
those
and
then
over
time.
A
D
B
These
are
these
voting.
These
are
all
long
distances.
These
are
long
distance
connections.
My
point
is,
if
I,
if
I
look
at
the
visual
cortex
of
the
somatosentry
cortex
or
any
part
of
the
cortex
you'd,
see
adjacent
columns
like
just
like
shown
here-
adjacent
columns,
making
connections
with
one
another,
but
actually,
where
those
you
know,
but
but
they'd
also
go
long
distances.
As
I
said,
you
go
from
the
left
side
of
the
brain
to
the
right
side
of
the
brain.
You
go
from
the
somatocentric
cortex
to
visual
cortex.
B
There
are
connections
from
v1
to
s1,
which
don't
make
any
sense
in
the
traditional
hierarchical
models,
but
they
make
sense
here.
It's
just
that
those
long
distance
connections
there's
lots
of
them,
but
they
they
are
still
sparse.
If
you
can't
have
a
fully
connected
network
between
all
columns
in
the
brain,
that
would
be
ridiculous.
A
B
It
there's
a
there's,
a
thing
called
the
white
matter
in
the
brain
right,
which
is
which
is
when
I
want
an
axon
from
a
cell,
wants
to
go
short
distance,
maybe
a
millimeter
two
or
so
on.
It
can
do
it
right
in
the
cortex,
but
if
it
wants
to
go
longer
distances,
it
leaves
the
cortex
and
goes
into
the
white
matter
and
enters
someplace
else,
and
so
most
of
the
volume
of
the
new
cortex
is
in
the
white
matter.
B
But
not
it's
not
most
of
the
connections,
the
most
divine
and
so
those
it's
interesting
like
if
you
well.
I
mentioned
this
before.
But
if
you,
if
you
cut
the
corpus
callosum,
which
is
the
white
matter,
that's
connecting
the
left
side
of
the
brain
to
the
right
side
of
the
brain
and
so
you're,
literally
separating
the
two
halves
of
the
brain
you're,
basically
preventing
any
distance
connections
between
the
left
side
of
the
brain
and
the
right
side.
B
It's
surprising
at
first,
because
it
doesn't
make
you
disabled
in
any
severe
way.
It
doesn't
like.
You
know.
If
you
cut
a
computer
in
half,
it's
not
going
to
work
right,
but
here
brains
continues
to
work,
but
what
you
find
out
is
that
the
two
sides
of
the
brain
are
reaching
different
conclusions
or
they
can
reach
different
conclusions.
B
So
if
I
show,
if
I,
if
I
take
a
person
who
has
their
brain
cut
in
half
and
and
and
they
grab
something
with
both
hands,
they
say
yeah,
that's
right.
I
know
what
that
is,
but
if
you,
if
you
control
it
so
that
the
left
hand
in
the
right
hand
or
actually
could
reach
different
conclusions,
they
do
reach
different
conclusions
and
the
left
hand
of
the
brain
may
say
this
is
the
coffee
company
right
after
the
brand
say
this
is
a
tennis
ball
whatever,
and
it's
it
it's
just
because
they
can't
vote.
B
That's
not
the
real
examples,
but
there
are
real
examples
that
they've
done
this.
Where
they,
you
present
something
to
two
sides
of
the
brain
and
they
they
both
reach
the
wrong
conclusion:
that's
consistent
with
the
left
side,
input
and
the
right
side
input,
but
they
don't
reach
the
same
conclusion,
whereas
a
normal
person
who
has
an
intact
corpus
callosum
will
reach
the
same
conclusion.
They'll
say
no,
there's
only
one
answer
here
and
that
answer
is
it's
a
coffee
cup,
so
those
those
are
long-range.
B
They
don't
envision,
whether
so
where
they
show
something
on
the
left
side
of
the
right,
the
visual
field,
the
right
side
of
the
visual
field-
I
don't
know,
I
don't
know
if
they've
done
the
touch,
they've
done,
vision
and
and
the
and
and
they'll
show
that
the
split
brain
patients
will
reach
two
different
conclusions.
It's
a
little
tricky
because
if
you
just
ask
them
a
question,
the
void,
the
the
the
language
is
only
on
the
left
side
of
the
brain.
B
E
E
The
one
the
one
I
remember
takes
advantage
of
the
dimorphism
of
the
brain
and
they
have
a
black.
You
have
a
black
box
where
you
reach
in
at
in,
like
have
a
coffee
cup
in
there,
but
you
can't
visually
see
it,
and
I
forget
whether
it
was
the
left
hand
or
right
hand,
but
one
of
them
will
you
grasp
it
and
they
ask
the
person.
Do
you
know
what
it
is?
Yes,
can
you
say
what
it
is
and
they
can't
they
can't.
B
B
E
A
E
B
Well,
there
are
stitchings
if
the
experiment
is
specifically
isolating
inputs
from
the
left
side
of
the
brains
on
the
right
side
of
the
brain
or
like
left-hand
input
from
right-handed
for
the
left
visual
field
from
right
visual
field,
then
it
could
be
what
you're
saying
super
type
pointing
out
there.
You
know
these
these.
What
and
where
pathways
where,
if
you,
if
you
damage
one
a
person,
can
reach
for
something
and
but
they
don't
know
what
it
is
they
can't.
You
know
they
just
don't
know
what
it
is,
but
I.
B
No,
that's
a
different
issue:
that's
damaged
the
one
where
pathway.
I
think
I
don't
want
to
get
off
topic
here
too
much,
because
I
think
we
want
to
focus
on
the
voting
mechanism
here
and
I
think
it's
important
to
get
a
sense
for
this
voting
mechanism.
You
know,
as
pubesce
points
out
very
eloquent.
I
never
really
realized
this,
but
the
columns
themselves
are
voting.
B
B
You
know
yeah.
I
I
like
to
make
a
cautionary
note
about
this:
the
columns
paper,
that's
what
this
paper
is
introduced.
Two
big
ideas,
one
big
idea,
so
each
column
is
able
to
learn
a
sensory
motor
model
of
of
its
input,
a
complete
model,
so
every
column
is
only
complete
models,
and
to
do
so,
it
needs
to
have
a
reference
range.
That
was
a
big
idea.
B
The
other
second
big
idea
is
the
voting
between
columns
all
the
other
details
in
the
paper.
Some
are
true,
or
some
are
not
true.
Some
there's
a
lot
of
things.
The
paper
didn't
explain
at
all.
It
didn't
explain
how
the
orientation
of
the
sensor
isn't
integrated.
It
didn't
explain
how
this
column
knows
how
things
are
moving.
B
B
Yeah
we
did,
we
did,
but
even
then
we
didn't
understand.
So
the
point
is
that
the
two
big
things
that
are
really
really
big
ideas
that
are
going
to
stay
true
are
that
there
are
that
columns
have
reference
frames
and
they
build
structured
models
through
sensory
motor
integration
and
they're
going
to
use
grid
cells
for
reference
frames.
B
B
I
guess
the
example
was
we
don't
show
how
orientation
is
represented
here
at
all
and
we
don't
show
how
objects
how
we
learn.
Object,
behaviors
or
object
composition.
So
that's
one
to
one
get
everyone
to
realize.
This
is
just
there's
a
lot
of
other
details,
some
of
which
came
out
in
the
frameworks
paper,
some
of
which
we
still
haven't
published
so
much.
We
still
don't
know,
but
these
two
big
ideas
are
really
big
ideas.
They
were
worth
publishing.
A
I
think
this
this
slide
has
kind
of
the
exact
rules
that
in
our
simulation,
which
again
jeff
has
mentioned.
That's
not
there's
a
lot
more
there's
a
lot
of
stuff-
that's
not
explained,
but
there's
this
is
in
the
simulations,
what
we
actually
implemented,
including
the
learning
rules.
So,
as
you
can
see
it's
super
simple,
it
can
be
stated,
I
kind
of
stated
it
all
verbally.
So
I
don't
know
if
it's
neat
to
go
through
it
or
not,
but.
A
We
had
a
bunch
of
these
simulation
results
in
there.
This
is
looking
at
the
set
of
cells
that
are
active
in
the
up
in
the
pooling
layer
and
the
object
layer.
So
on
the
left
here
it
it
shows,
you
know,
let's
say,
for
10
different
objects.
What
is
the
sdr
corresponding
to
each
of
these
10
objects?
So
I
think
we
had
roughly
4
000
cells
in
the
object
layer
and
so
for
object,
number
one.
These
are
the
cells
that
would
be
active
and
for
object.
Two,
it
be.
A
You
know
this
column
of
cells
active,
so
there's
a
very
sparse
subset
of
cells,
that's
active
for
each
object.
I
think
we
have
40
out
of
4
000
active,
so
it's
like
one
percent,
very,
very
sparse,
and
then,
if
you
look
at
panel
b
here,
this
shows
the
convergence
of
those
of
that
layer
as
a
function
of
the
number
of
sensations.
A
A
A
A
If
you
have
more
more
unique
features
in
your
cortical
column,
it's
going
to
get
it's
going
to
be
quicker
and
quicker.
So
if
you
look
vertically
here,
if
you
only
had
five
unique
features
that
are
available,
it
would
take
you
longer,
whereas
if
you
have
30
unique
features,
you
can
sort
of
understand
that
each
feature
is
going
to
really
narrow
things
down
a
lot
more.
So
when
you
have
a
fixed
number
of
objects,
let's
say
40,
you
have
30
unique
features.
A
Maybe
I'll
I'll
show
this
one.
We
talked
a
little
bit
about
some
of
the
neurons
that
have
long-range
lateral
connections.
Here's
there's
many
papers
that
that
talk
about
it.
Here's
one
particular
image
that
I
really
like
that's
showing
this
particular
cell.
This
is
looking
from
the
top
down
this
particular
cell
and
the
connections
it
makes
locally.
So
this
is
essentially
within
one
cortical
column,
and
you
can
see
that
it's
making
connections
further
and
further
away
to
different
cortical
columns,
there's
more
subtlety
here,
which
I
I
won't
get
to
unless.
B
B
B
No
all
all
parameters,
not
stellate
cells,
but
all
parameters.
I
believe
it's
true
for
every
parameter.
I
I
think
that's
why
they're?
This
is
why
you
know
when
they
talk
about
inhibitory
cells.
They
sometimes
distinguish
between
interneurons
interneurons
are
basically
any
cell
that
doesn't
send
its
axon
elsewhere,
whether
it's
inhibitory
excitatory
most
of
them
are
inhibitory,
but
like
spiny,
solid
cells,
don't
send
their
axon
out
of
the
cortex.
Therefore,
they're
considered
interneurons
technically
so,
but
if
it's
a
pyramidal
cell,
I
believe
every
parameter
cell
sends
its
axon
into
the
white
matter.
B
Someplace
else
now
it
doesn't
mean
it's
going
elsewhere
in
the
cortex.
It
could
be
going
to
subcortical
something
like
that,
but
I
believe
these
let
all
their
two
three
cells.
I
could
try
to
check
this,
but
I
believe
this
is
true:
send
their
send
an
accent
also
into
the
white
matter,
which
will
go
someplace
else
in
the
cortex
and
you're
not
going
to
see
that
in
this
picture,
which.
B
This
is
all
you
can
see
the
little
scale
bar
there,
one
millimeter.
So
this
is
all
within,
like
you
know,
10
millimeters
or
seven
millimeters
and
that's
kind
of
the
that's
generally
considered
the
extent.
Maybe
eight
millimeters
the
extent
of
what
is
what
a
local
axon
will
do,
but
it
doesn't
mean
it's
not
connecting
much
further
away.
It
is
almost
certainly
connecting
us
for
the
right-hand
cortex.
B
It
just
won't
be
very,
very
sparse.
It's
you're,
just
not
gonna,
you
might
see
one
teeny
little
red
dot,
someplace
else
and
another
teeny
little
dot,
someplace
else
or
a
bunch
of
teeny
dots
someplace.
You
know
just
I'm
just
pointing
that
out.
Yeah,
that's
what
the
cortical
cortical
connections
are
yeah.
The.
A
Label
yeah,
so
maybe
one
last
slide
so
the
whole
idea.
You
know
the
two
two
big
ideas
that
jeff
mentioned:
really
change,
how
you
think
about
hierarchy,
how
you
could
think
about
hierarchy
and
we're
still
trying
to
figure
all
this
out,
but
I'll
mention
you
know
one
aspect
of
it.
This
is
sort
of
again
the
core
to
the
whole
thousand
brains.
A
Theory
yeah.
So
if
you
look
at
hierarchy
in
the
traditional
viewpoint,
you
have
some
sensory
input
coming
in
from
the
retina,
you
have
some
simple
features
that
are
detected,
let's
say
in
v1
and
the
first
layer,
more
complex
features
and
v2
and
so
on,
and
then
somewhere
up
at
the
top.
You
have
an
object.
The
objects
are
actually
recognized.
A
A
B
B
A
B
From
here
to
there,
yeah
yeah
they're
not
a
lot
but
they're
there,
which
make
absolutely
zero
sense
in
the
classic
view
like
what
would
a
you
know?
How
could
a
a
line
orientation
in
one
party,
visual
field
affect
have
anything
to
tell
you
about
a
particular
sensation
in
your
off
of
one
finger.
You
know
it.
It
doesn't
make
any
sense
in
the
classic
view,
but
anyway,
I
might
point
out
this
40
number
is
much
higher
now,
because
they
have
new
ways
of
discovering
these
things.
B
E
B
But
it's
very
biased
by
the
techniques
that
were
technically
available
at
the
time
they
did
this.
Also
where
they
looked.
You
know
there
was
no
like
magic.
You
know,
camera
that
says,
show
me
all
the
regions
they
they
had
to
do
this
very
painstakingly
back
then,
where
they
would
inject
the
dye
into
one
side
of
the
region
and
then
see
if
that
thing
showed
up
in
another
set
of
regions,
and
it
was
very
great,
so
it
was
a.
E
B
It's
very
difficult,
it's
it's
really
challenging.
I
don't
believe,
that's
true
kevin.
I
could
be
wrong.
I
don't
think
that's
true.
You
have
to
you,
have
to
you,
know,
go
down
and
you
know
reject
the
dye.
You
have
to
make
sure
what
layers
are
you
injecting
it
into?
Then
you
have
to
figure
out,
you
know
how
far
does
it
many
of
these
died?
I
don't
know
the
rabies,
but
many
of
them
don't
go
very
far.
They
get
limited
and
then
you
have
to
then
carefully
look
everywhere
else
in
the
cortex.
B
E
B
Anyway,
the
techniques
for
met
for
duty
determinants
are
very
difficult,
very
limited
back
when
fellman
and
vanessa
did
this
paper
they
had
to
co,
they
had
to
combine
results
from
different
experimental
techniques,
different
ways
that
people
did
this.
That
was
one
of
their
challenges,
because
the
different
techniques
show
different
types
of
connections
or
have
different
limitations,
and
so
they
admit
in
their
paper
like
well,
you
know
we're
combining
them,
but
maybe
that's
not
the
right
thing
to
do,
because
each
one
has
its
biases.
C
A
A
A
B
B
But
that
rule
that
there's
two
layers
of
cells,
two
types
upper
layers
and
lower
lasers
spread
is
very
well
known.
Can
I
make
a
side
point
about
that
right
now,
I'm
interrupting
your
hierarchy
thing
here,
that's
fine!
We
talked
about
object,
identity,
right,
we're,
recognizing
objects,
and-
and
we
know
that
this
is
happening
because
you
you
when
I
look
at
my
coffee
cup-
it's
stable
even
though
my
eyes
are
moving.
It's
I
perceive
the
object
identity.
B
I
don't
perceive
the
movement
of
my
eyes,
necessarily
unless
I
attend
to
it,
and
and
so
we
know
that
what
we're
perceiving
is
this
voting,
however,
we're
not
just
voting
on
the
object,
identity,
we're
also
voting
on
the
objects
orientation
in
location
relative
to
my
body.
I
want
to
see.
Oh
I'm
seeing
a
coffee
cup.
I
see
the
coffee
cup
at
a
particular
location
in
a
particular
orientation,
and
so
that's
stable
too,
and
so
that
also
has
to
be
voted
on.
B
We
haven't
talked
about
this
much,
but
the
voting
is
more
than
just
object,
identity.
It's
also
the
object's
posed
to
the
body
and
we've
not
really
incorporated
that
into
our
theories,
but
but
so
there's
more
than
one
thing
being
voted
on
here,
and
we
don't
really
understand
that
too
well,
yet
we
talked
about
it
a
bunch,
but
I
just
want
to
point
that
out
that
you
can.
B
You
can
know,
what's
being
voted
on
by
what
you
are
able
to
perceive
consciously
and
and
talk
about,
because
you
don't
get
to
see
what's
going
on
in
most
of
what's
going
on
in
the
column.
That's
not
that's
not
available
to
the
other
parts
of
the
brain,
but
the
part
that
is
available
to
your
memory
and
to
into
your
conscious
what
you
can
talk
about
is
the
that
has
to
be
voted
on
because
it's
stable.
B
B
C
B
B
C
F
A
F
So
that
that
means
that
not
every
vote
is
equal
basically,
so
it's
not
really
a
democratic
system
where
every
column
puts
in
one
vote
and
then
it's
counted
or
something
it's
more
like
how
opinions
spread
on
social
media.
Some
sources
are
more
credible
and
reach
more
people
and
concerning
different
types
of
questions.
If
I
may
make
that
matter,.
B
For
well,
I
think
I
think
the
difference
between
this
is
that
you
can
try
to.
In
this
case
the
system
settles
on
a
single
answer.
It
it
it's.
It's
never
happened.
You
know
that
you
don't
find
the
brain,
the
you
know,
like
you,
know,
liberals
and
conservatives,
or
something
like
that.
You
know
it's
like
it.
They're
all
gonna
it.
It's
gonna,
it's
gonna.
Do
it's!
It's
not
happy,
not
having
a
single
answer.
So
it's
kind
of
keeping
yeah.
B
B
So
I
got
thousands
and
thousands
of
columns,
modeling
this
coffee
cup
and
they're
all
agreeing
it's
a
coffee
cup,
and
now
I'm
holding
the
carpet
cup
in
two
hands,
and
so
I
have
I
have
hundreds
of
patches
of
skin
that
are
touching
the
coffee
cup
at
the
same
time
and
they're
all
in
agreement
they're
all
saying
yeah.
This
is
the
same
thing,
but
now,
if
one
small
part,
one
teeny
part
of
my
visual
field
or
one
small
part
of
my
skin
feels
something
that's
inconsistent.
B
You
know
like
that.
Shouldn't
be
there
you,
don't
it
doesn't
get
you
you
notice
it.
It's
like.
Even
a
single
incorrect
vote
can
say
uh-uh
what
you're
telling
me
it's
inconsistent
with
what
I'm
sensing
something
is
wrong
here
and
the
whole
system
gets
flipped
around
it's
it's
not
like
a
it's
not
like.
Oh
majority
wins.
It's
like
everybody's
got
to
agree
and
if
one
person
disagrees,
one
person
has
solid
evidence
that
this
is
not
what
they're
sensing
there's
a
funny
sensation
on
my
one
my
index
finger.
B
Then
then
the
whole
system
attends
to
that
and
we
say:
okay,
what's
wrong
here,
something's
inconsistent:
we
have
to
figure
it
out
and
what
we'll
do
you'll
do
is
you'll,
then
visually
attend
and
tactfully
attend
to
that
one
spot
on
the
cup.
That
was
not
right
and
you
say,
oh
there's
a
funny
little
hole
here
or
it's
got
a
crack
or
you
know
this
part
of
the
cup
is
made
of
liquid.
I
don't
understand
that.
So
it's
not
a
majority
rule.
B
B
Similarly,
if
the
only
two
differences
between
two
objects
is
maybe
temperature
or
or
some
sort
of
texture
on
the
surface-
and
maybe
the
visual
cortex
can't
can't
resolve
that.
So
I
I
think
it's
the
way
more
look
at
it.
It's
not
like
some
are
better
or
more
trustworthy.
I
think
it's
this.
They
all
have
different
abilities,
and
but
everyone
has
to
consid
agree
based
on
their
abilities,
but
there's
no
more
expert
than
someone
else.
F
B
A
There
are
a
lot
of
other
terms
that
you
could
use
and
some
people
have
have
used
like
you
know
you
can
frame
it
in
terms
of
prediction.
You
know
the
different
columns
are
predicting
what
the
object
would
be
and
the
and
the
columns
that
have
the
most
predictions.
If
you,
I
guess,
that's
voting,
but
you
know
it's
it's.
So
it's
kind
of
like
a
predictive
signal.
You
know
we,
you
know
you're
getting
predictions
from
neighboring
columns,
so
it's
a
type
of
predictive
coding.
If
you,
if
you
want
to
think
about.
B
It
that
way,
I
I
think
again,
it's
not.
I
think
the
thing
we
have
to
avoid
is
like,
like
oh
majority,
wins,
that's,
that's
not
right.
It's
not!
The
majority
wins
it's.
Everyone
has
to
agree.
Even
if
I
don't
have
a
lot
to
say
about
it.
Maybe
I
have
just
a
limited
ability
to
predict.
You
know
to
know
what
this
object
is,
but
it
has
at
least
be
consistent
with
what
my
models
the
world
are.
So
I
think
the
idea
that
it's
not
like
a
majority.
That's
the
thing
I'm
trying
to
get
over
it's
it's.
B
E
B
You
know
another
example
one
I
use
in
the
book
is
you
know
the
bistable
images
or
bistable
perceptions
where
you
can
have
like
you
know.
The
example
was
the
the
face.
The
two
faces
versus
the
bonds
and
there
you
have.
You
know
the
all.
The
columns
can
be
consistent
with
either
interpretation
all
the
visual
columns.
You
know
it
could
be
the
one
and
so,
but
you
don't
maintain
both
of
those
thoughts
in
your
head.
At
the
same
time,
it's
very
you
can't
perceive
that
image
as
both
faces
and
bars
at
the
same
time.
B
It's
just
it's
got
to
be
one
or
the
other,
and
it
can
bounce
back
and
forth
every
few
seconds,
but
there
is
no
like
hey.
You
know
we're
not
consistent
here.
So
we'll
just
keep
two
two
things
going
at
once:
nope
everyone's
gonna
agree
or
not
agree,
and
if
there
was
a
single
feature
that
differentiated
between
the
vowels
and
the
faces,
then
you
wouldn't
see
the
other
one.
It
would
just
be
you're
done.
You
know,
we
know
what
it
is.
E
A
A
As
long
as
they're
modelling
the
same
objects,
it
doesn't
matter
whether
it's
somatosensory
or
visual,
they
can
all
because
they
can
all
have
stable
representations
of
objects.
They
can
easily
vote
on
what
the
object
is,
and
so
you
know
think
about
like
a
a
lemon
versus
a
line.
You
know
if
you're
touching
it,
you
may
not
be
able
to
tell
the
difference,
but
your
eyes
can
tell
you
what
the
color
is,
and
so
you
can.
E
B
B
If
it's
a
very
skinny
straw
and
it's
a
big
image,
it
will
be
hard
for
you
to
recognize
the
big
image.
It'll
just
be
hard
for
you
to
integrate
over
time
the
the
sensations
you
need
to
do
this,
and
but
if
I
let
you
look
through
that
image
with
a
big
straw,
you'd
be
able
to
do
it.
Similarly,
if
I
had
a
very
very
small
thing,
I
was
looking
at
a
continuous
letter,
you
could
write
well.
B
The
looking
at
to
the
narrow
straw
would
work
very
well,
but
if
I
look
into
a
big
straw,
assuming
I
had
the
same
resolution
for
straw,
meaning
I
would
now
be
very
fuzzy,
I
wouldn't
be
able
to
recognize
it.
So
one
hypothesis
we
put
out
is
that
you
know
at
least
in
the
first
several
regions
of
the
visual
system,
they're
they're,
looking
at
the
world
through
different
scales,
and
that
makes
it
so
so
the
sensory
motor
integration
can
work.
B
Yet
there
still
have
this
hierarchical
connection.
One
possibility
of
the
hierarchical
connection
is
that
that
region
v1
can
project
the
v2.
Not
only
you
know,
not
only
does
v2
get
input
directly
from
the
retina,
but
it
can
also
get
process
data
from
v1,
so
v1
can
be
saying
you
know,
I
think,
we're
looking
at
a
coffee
cup.
That's
where
your
input
to
your
your
region.
B
Another
thing
is
that
these
connections
that
go
through
the
thumbnails
look
like
motor
connections,
and
so
they
may
be
just
passing
motor
information
between
regions
which
is
hard
to
describe.
So
I
guess
I'm
saying
kevin,
there's
a
lot.
We
don't
understand
about
this
and
we
need
to
it's
just
it's
a
lot.
A
What
percentages
of
percent
of
the
synapses
in
the
neocortex
actually
correspond
to
this
particular
model?
I
mean
clearly
there's
a
hierarchy
embedded
in
here
in
the
sense
that
there's
always
a
hierarchy
in
a
complete
graph.
That's
almost
thing,
but
if
you
look
at
just
this
model
and
what
they
propose
as
being
the
main
kind
of
feedboard
pathway,
this
hierarchical
pathway,
what
percentage
of
synapses
in
the
neocortex
are
actually
part
of
this?
E
E
My
understanding
was
well.
Okay,
that's
that's!
I!
I
was
running
off
the
fact
that
there
was
suggests
at
one
point
that
there's
ten
times
as
many
connections
going,
the
reverse
direction,
isn't
the
feed
forward
connection,
but
that
eliminates
totally
the
aspects
of
of
hierarchy
and
and
and
lateral
connections.
A
F
A
A
B
A
Yeah,
so
let
me
point
out
some
facts
so,
as
we've
discussed
before
a
very
small
percentage
of
the
synapses
are
actually
proximal.
So
it's
you
know,
these
are
definitely
proximal
connections
that
are
going
up
the
hierarchy
here.
So
I
think
we
talked
about
two
to
five
percent
of
a
pyramidal.
Neurons
connections
are
proximal
below
that
you're
gonna
be
below
that,
but
you
know
so
it's
bounded
by
two
to
five
percent.
So
it's
going
to
be
less
than
five
percent
of
the
connections.
A
So
I
think
it's
dramatically
less
than
one
percent
of
the
connections
in
the
neocortex
are
explained
by
the
speed
forward
connections.
What
this
means
is
that
there
are
100
times
more
synapses
explained
here,
and
so
the
feedboard
hierarchy
explains
one
percent
or
less
of
the
neocortex,
and
why
do
people
think
this
is
the
way
the
neocortex
works?
It
just
baffles
me
well,
just
very
ignoring
99
of
what's
going
on
in
the
neocortex.
Well,.
B
E
A
C
A
C
Okay,
you
have
to
have
some
like
feedforward
stuff
to
compute
features
so
that
those
neurons
in
layer
four
have
future
representations.
But
now
I'm
thinking,
maybe
even
those
connections
or
those
processing
steps
are
not
necessary
because
the
neurons
from
layer,
two
and
three
are
also
feeding
back
down
to
them
and
they're
also
receiving
information
from
other
columns.
So
is
it
right
to
think
of
the
neurons.
D
B
Not
really
priority
stuff,
not
really
again,
you
have
so
it's
really
complicated,
but
the
general
thought
is.
You
have
to
keep
track
of
proximal
synapses
versus
distal
synapses
right.
So
when
we
say
like
oh
I'm
getting
feedback,
you
know
it
does,
does
v1
get
feedback
from
v2
absolutely,
but
is
that
a
driver
meaning
does
it
actually
make
the
cells
in
v1
fire
like?
B
If
I
give
you
a
sensory
input
from
the
retina
to
v1,
it
makes
the
cells
fire
in
v1,
and
it's
and
and
one
of
the
some
of
the
cells
are
in
layer
four.
But
if
I
have
a
feedback
projection,
it
typically
wouldn't
go
to
the
proximal.
Synapses
typically
would
go
to
the
distal
synapses
and
therefore
it
doesn't
drive
those
cells.
Doesn't
it
you
can't
make
v2
can't
make
v1
fire?
You
know
that
kind
of
thing.
B
B
So
yeah
exactly
so,
for
example,
here's
what
happens
classically
receptive
fields
are
defined,
especially
in
business
research,
which
is
most
of
this.
They
will
be
able
to
start
off
with
an
animal.
They
will
stick
a
probe
in
v1
they'll,
show
it
some
sinusoidal
grading
so
that
they
can
find
out.
Oh,
what
is
the
receptive
field
of
these
cells?
Oh,
this
cell
responds
to
a
lying
edge
at
this
angle,
that
kind
of
thing,
but
then,
when
they
actually
record
these
cells
in
real
life
in
a
live
animal,
doing
real
things
they
could
they
behave
completely
differently.
B
They
seem
to
not
care
about
that.
They
respond
to
other
things,
or
they
only
respond
some
of
the
time
very
right.
And
so
what
is
the
receptive
field
of
that
cell?
Well,
there's
something
we
would
say
that
spell
is
might
be
responding
to
a
particular
feature
out
of
and
under
various
different
contexts,
and
so
there's
you
know,
but
that
cell
could
respond
under
different
contexts.
So
in
some
sense
you
could
say
that
cell
has
many
different
receptive
fields
because
all
the
different
contexts
in
which
it
could
be
active.
B
So
so
anyway,
it's
just
very
confusing,
and
so,
if
you
stick
with
the
classic
definition
of
depth
of
field,
then
then
the
feedback
they
just
say:
oh,
it
must
modify
it
somehow,
but
in
reality
it
just
changes
completely.
What
the
cell,
how
you
interpret
the
cell
and-
and
I
don't
know
how
to
better,
describe.
C
B
It's
so
like,
so
you
could
say
that
all
the
cells
in
this
in
this
little
section
some
of
the
cells,
not
all
of
them,
the
some
the
cells
that
they're
actually
able
to
find
responding,
which
is
less
than
50,
seem
to
have
this
preference.
And
then
that's
been
come
over
with
a
different
preference,
and
in
this
case,
since
there's
no
context
because
they're
giving
a
sinusoidal
grading
with
animals
anesthetized,
then
we
expect
to
see
all
those
cells
respond
to
the
same
orientation.
B
The
way
I
think
about
that
is
that
has
to
be
learned.
It's
not
it's,
not
predetermined
and,
and
not
only
do
all
those
cells
have
to
learn
to
respond.
We
have
to
ask:
how
could
all
those
cells
in
the
mini
column
actually
learn
to
respond
to
the
same
thing,
because
why
would
they
respond
to
the
same
thing
they're
competing
with
each
other
and
the
best
answer
we
have?
It
gets
a
little
bit
complicated
that
there
are
these
inhibitory
cells
called
bipolar
cells,
which,
or
that
have
very,
very
skinny
they're.
B
Very
the
axons
and
dendrites
are
really
skinny,
they
kind
of
fit
within
a
mini
column,
and
they
also
have
this
receptive
field,
so
they
can
say.
Yes,
this
cell
is
going
to
respond
to
an
angle
of
45
degrees,
just
like
all
the
pyramidal
cells
or
the
other
cells
in
the
mini
column,
and
what
we
think
is
going
on
is
that
this
is
a
pretty
guesswork,
I'm
not
going
to
say
that,
but
this
is
the
only
explanation
I
have
so
far
is
that
those
those
inhibitory
cells
actually
are
the
ones
that
learn
the
receptive
field.
B
They
compete
with
each
other.
These
bipolar
cells,
they
then
train
the
excitatory
cells
in
their
columns,
saying
hey,
guys,
we're
all
together
on
this,
we're
all
going
to
learn
45
degrees,
because
that's
alright,
that's
what
I've
learned
and,
and
so
that
has
to
be
learned.
And
if
you
change
the
statistics
of
the
data,
they'll
form
different
representations.