►
From YouTube: Jeff Hawkins on Minicolumns & Spatial Pooling
Description
Jeff Hawkins describes minicolumns and Spatial Pooling. Discussion at https://discourse.numenta.org/t/jeff-on-minicolumns-spatial-pooling/6401.
C
C
D
B
E
B
Where
the
Sun,
but
the
way
I'd
like
to
think
about
this,
is
that
easy,
the
inhibitory
cells
are
really
what
implement
the
algorithm
to
the
brain
Marcus.
You
can
jump,
please
jump
in,
have
anything,
that's
what
I'm
saying
or
get
ready
if
I
get
anything
wrong,
so
these
sort
of
these
sort
of
like
well,
you
know:
how
does
this
thing
do
something
besides
just
a
bunch
of
neurons
right,
so
the
general
belief
is
across
the
memory
is
stored
in
here.
Then
these
synapses,
in
not
all
of
it.
B
So
the
mix
together
everywhere
in
the
neocortex
and
so
the
channels
have
talked
about
columns,
and
so
this
is
also
a
confusing
topic.
If
you
look
at
a
neocortex
in
a
human
man,
I
think
we
hear
you
typing
I
think
you
need
a
new,
inhuman
and
your
cortex
is
about
two
and
a
half
millimeters
thick.
Your
power
is
bit
right
and
one
of
the
things
you
see.
B
If
you
look
under
this
on
your
microscope,
of
course,
you
see
lots
and
lots
of
neurons,
but
in
some
animals
some
mammals,
it's
more
obvious
and
others
that
one
level
of
organization
is
this
thing
called
a
mini
column
and
a
mini
column
is
really
skinny
it's
it's
literally.
It
can
be
anywhere
from
50,
maybe
to.
B
B
Some
times
so
these
are
these
are
physical
things
that
that
seem
to
exist
and
and
won
the
debate.
It'll
be--it'll
they
do
anything
because
something
some
animals
are
harder.
The
city
and
other
animals
are
very
visible,
but
that
doesn't
tell
you
anything,
because
you
could
have
the
exact
same
connectivity
and
not
be
visibly
separated,
and
you
do
the
same
thing
and
they're
all
so
we
know
that
they're
derived
from
growth,
so
when
an
animal's
brain
starts
developing
was
a
progenitor
cell
or
a
region
that
sends
these
neurons
all
off.
C
B
From
again
it's
a
subset
of
all
the
cells
would
be
a
subset
of
the
excitatory
cells,
typically
in
most
of
the
cells
that
couldn't
figure
out
what
they
did
so,
but
the
ones
they
click
right
opens.
You
know
the
ones
they
do
respond
to
this
training.
You
might
say:
oh,
these
always
find
in
the
computer.
Cortex
season
always
poncho
a
vertical
line
he's
responding
to
the
next
one
or
like
this,
and
these
respond
to
the
next
one.
B
I
know
it's
frames.
If
you
get
older
everything,
the
brain
strengthened,
you
know,
but
I
think
is
a
misimpression
that
is
pretty
much
defined
at
birth.
You
know
and
there's
this
over
perfusion
of
connections
which
disappear,
but
it's
not
something
like
it
grows
as
you
grow
in
life.
It's
not
like
it's
getting
bigger
as
you
learn
things.
It's.
B
B
Basic
idea
is
that
all
the
different
types
of
cells
are
you're
going
to
find
you
ever
find
interesting
event.
That
was
an
idea
I'm
going
to
been
proven,
so
he
basically,
this
is
a
unit
of
computation
that
seems
to
incorporate
everything
and
get
it
all.
These
things
somehow
are
related
in
one
way
or
another.
B
So
that's
the
mini
column.
It's
a
physical
thing.
It's
there
at
during
development
of
the
brain
when
you're
in
utero.
It
also
seems
that
a
functional
role,
which
is
known
ability
of
the
sign
to
control,
to
no
one
understood
why
it's
like
that.
We
have
a
theory
of
whether
it
is
and,
and
they
vary
in
size-
is
based
on
animals.
But
roughly
you
can
say
about
200
million
of
them,
so
mal
castle,
Verner
mountain
castle
is
one
of
the
people
who
really
said
you
know.
This
is
obviously
an
important
functional
thing.
B
We
need
to
understand
what
this
thing
does.
If
you
understand
what
this
thing
does
you
get
200
million
of
similar
there's
one
big
exception
in
primates
in
primates
on
like
so
all,
mammals
have
neocortex,
but
primates
v1
that
the
primary
visual
cortex
is
v1
is
twice
as
many
cells
and
we
talk
about
that's
a
separate
topic,
but
other
animals
were
not.
Primates
have
visual
systems
like
cats,
them
dogs
and
they
don't
have
Isis,
and
then
we
have
these
extra
layers
and
b1.
C
B
Choices
make
cells
with
many
calm
down
or
any
any
area
in
v1.
So
now
get
to
your
question
about
column
when
coming
to
the
column,
you
have
to
be
careful,
they
might
be
talking
about
a
meanie
column,
but
they
shouldn't,
they
should
say
mini
column,
but
there
isn't
another
column
which,
which
is
the
following.
You
see
of
this
like
those
these
properties
here
where
this
is
the
vertical
line
right
and
then
you
move
over
at
the
response
to
different
lines
and
if
you
keep
going.
B
B
You'll
see
anyway,
if
you
take
the
area
which
represents
all
the
different
properties
you
can
get
before
it
repeats
again,
it's
about
roughly
a
millimeter
by
a
millimeter,
and
so
also
like
one
of
these
many
columns
may
process
a
particular
type
of
orientation,
a
particular
ie
reference
right,
our
accomodation
different
color
things.
So
there's
other
properties
that
are
embedded
in
here
and
then
you
go
to
the
next
block
over
it
all
repeats
again,
but
it
took
part
of
that.
B
B
This
is
not
a
physical
thing
in
the
sense
that
you
can't
see
it.
There's
no
demarcation
here.
There's
no
one
says:
oh
here's
the
edge
of
it
and
here's
the
next
one
it's
a
continuous
sheet,
but
the
idea
that
if
you
continue
repeating
in
any
direction
things
repeat,
you
keep
going
in
a
direction
repeat.
So
that's
the
larger
column.
B
It
that's
an
interesting
question,
so
Bernie
now
Castle
he's
one
of
the
few
people
really
study
this,
and
so
a
lot
of
the
species
is
old,
but
he
argued
that
he
argued
that
he
saw
this
type
of
column
organization
and
the
way
he
would
do
that
he
would
say.
For
example,
he
studied
a
lot
of
somatosensory
cortex,
so
he
was
tonight
a
century
cortex
shadow.
Imagine
I
have
like
a
millimeter
blocks
here.
This.
B
He
would
show
that
if
he
put
a
probe
through
here
with
he
argued
you
had,
you
would
have
these
discontinuities
into
the
area
of
the
part
of
the
body
that
was
represented.
So
in
this
section
right
here
is
all
the
cells
in
this
section
would
be
representing.
Let's
say
this
was
your
forearm.
Am
I
representing
some
area
on
your
and
and
when
you
travel
around
even
show
you
another
area
forearm,
and
so
he
was
on.
This
is
processing
one
area.
B
Is
that
makes
the
argument
that
this
sort
of
column,
the
organization
exists
everywhere
everywhere
he's
looked
everywhere,
you
could
measure
some
of
the
higher
areas
of
cortex.
It's
like
they
don't
know
what
to
look
for
right.
You
can
do
this
for
like
auditory
cortex.
You
can
do
this
with
visual
cortex.
You
can
do
this
amount.
B
Actually,
probe
the
animal
and
see
what's
going
on,
but
other
parts,
the
brain
is
very
difficult
to
say,
but
he
made
he
made
a
long
argument
and
a
long
paper
about
all
the
evidence
suggests.
This
sort
of
column
organizations
exist
everywhere.
So
it's
a
nice
idea
and
threateningly.
It's
a
it's
a
great
idea,
so,
whether
it's
true
or
not
remains
to
be
seen,
but
at
the
moment
I
take
it.
Basically,
that's
the
right
thing
to
go
until
we
know
otherwise.
E
B
Well,
that's
consistent
with
this
idea
that
would
be
considered
cops
so
in
the
mouse
cortex
yeah
they
have
their
whiskers
are
a
very
active
sensing
devices
and
not
just
hairs
right.
They
sweep
them
back
and
forth
and
they
have
the
various
that
they
can
see
with
their
whiskers
and
some
say
if
they
can
learn,
recognize
objects
and
number
things.
B
Are
it's
a
it's
like
our
fingers
in
some
sense,
so
it's
an
active
sense
and
what
they've
discovered
and
then
looking
down
on
them
on
the
rat
or
mouse
in
your
cortex
you'll
see
the
whiskers
is
a
path
of
whiskers
on
their
face
and
those
whiskers
are
represented
literally
Matt.
One
two
one
knows
his
columns:
the
bigger
columns
mac
to
each
whisker.
B
In
this
case,
it's
an
unusual,
you
know
I,
think
I've
interpreted
that
it's
like
well,
your
retina
is
a
continuous
sheet.
Your
skin
is
itching
you
see
here.
The
whiskers
are
not
continuous,
there's
this
unique
things
and
each
one
has
its
own
column.
So
this
is
very
consistent
with
non
council's
basic
proposal
for
columns
everywhere
and
just
in
this
case,
they're
very
visible,
so
you
could
study
them
and
look
at
them,
and
so
these
have
been.
These
have
been
very
modest,
huge
amount
known
about
the.
C
D
B
You
see
these
literally
if
you
can
stain
or
look
at
the
neurons
right.
If
you
have
a
way
of
looking
at
it,
you'll
see
that
the
cell
bodies
would
eventually
connected
here
with
a
little
gap
in
between
them
and
more
so,
as
here,
neither
literally
connected
each
retina
each
whisker,
so
they're
visually
visible
barrel.
That's
what
makes
them
unique,
we're
in
visual
cortex
and
sensory
cortex.
B
You
don't
see
the
columns,
you
see
the
many
columns,
but
you
don't
see
the
columns
themselves
so
again,
I
think
that's
a
uniqueness
due
to
the
fact
that
they,
the
hairs,
are
distinct
and
therefore
they
cannot
this
way,
but
the
photoreceptors
on
the
retina
or
this
on
your
skin
are
not
separated
at
all
in
any
way.
But
anyway,
these
are
neither
actually
should
one
think
of
these
are
either
on
the
small
side.
I
believe
these
were
about
half
a
millimeter
each,
but
they
were
still
qualified
as
a
column.
B
Well,
I'm
bigger
a
little
actually
is
more
than
three
dimensions
to
because
you've
got
depth,
you
got
ocular
dominance
and
you've
got
color.
You
got
a
bunch
of
stuff,
they
not
so
the
general
rule
is,
it
seems
like
what
cortex
does
or
is
bringing
the
revolution
is
not
it
takes
multiple
dimensions
and
you
have
multiple
dimensions
in
your
skin.
You
have
you
have
various
property
of
hot
and
cold
sensors.
You
have
sharp
mention,
sensors,
you
have
vibration,
sensors,
you
have
pressure
sensors,
all
these
things
are
matter
and
they
all
sort
of
overlay.
B
On
top
of
this,
when
you
look
at
a
column
in
the
cortex
I,
whenever
I'm
talking
about
how
the
visual
hyper
column
of
us
known
is
cleared
in
the
air,
you
have
all
those
dimensions,
mat
and
the
imagine
these
kind
of
weird
little
overlapping
wavy
patterns.
But
the
point
was
that
within
one
square
millimeter,
whatever
property
you
have
is
going
to
repeat
so
that's
the
property
of
orientations
can
repeat
if
it's
the
property.
B
A
B
B
B
C
B
They're
not
visible
because
again
you
don't
have
to
be
visible,
have
the
same
functional
structure
right
if
the
neurons
are
all
connected
exactly
the
same
way,
and
you
just
shift
them
around
a
little
bit.
You're
still
gonna
do
the
same
thing
so
so
with
the
physicalness
of
it
is
not
essential,
but
the
fact
that
we
do
see
them
almost
everywhere,
I
think
we
should
assume
didn't
comps,
insist
everywhere.
That
would
be
the
that
would
be
that's
my
assumption
and
I.
My
assumption,
I
bind
to
map
has
the
argument
that
coms
exist.
Oh,
Richard,.
D
C
B
That
his
in
his
first
big
paper
about
this,
he
said
many
columns
as
a
organizing
principle
and
then
later
he
argued
the
column
for
the
organizing.
You
know
both
and
he
got
a
lot
of
pushback,
initially
he's
initial
papers.
On
this.
He
argued
that
the
column
was
dynamic,
that
it
could,
under
varying
conditions
at
moment-to-moment,
could
change.
Then
he
retracted
that
I
think
his
last
thing
was.
He
says
no
I
think
there
fits
the
Vicks
act
alarming
if
you
remember
anything
different
Marcus,
because
what
I
know
that
that
last
part
you're
my
source,
okay.
B
Represent
because
you
can
take
this
video,
you
take
a
piece
of
cord
X
you
plug
in
something
else
into
it.
It
represents
something
else
so
whether
the
size
of
dynamically,
the
questions
and
including
what
it
represents
is
dynamic,
for
example,
the
side
it
could
be
determined
by
somebody's
inhibitory
cells
that
just
span
of
certain
dimension,
and
they
say
that's
going
to
define
the
width
of
a
column
you
know
like
I
can.
B
B
What
they
do
they
do
all
this
stuff,
but
the
general
rule
would
be
that
if
you
D
innervate,
any
part
of
the
cortex
so
no
longer
receives
an
input
from
something
I
could
cut
the
axons
or
something
like
that.
What
will
generally
happen
is
all
these
I,
don't
say
it's
true
for
the
right
barrel:
cortex,
because
that
could
be
totally
hardwired
by
evolution
right,
but
the
general
principle
would
be
forgiving.
We
did
this
to
your
visual
cortex
or
your
auditory
cortex
or
some
other
project
cortex.
B
D
B
Part
of
the
cortex
now
gets
represented
by
here,
and
everybody
moves
apart,
right,
R.
That
was
one
of
the
nice
things
about
our
spatial
Euler.
It
has
this
property
that,
and
we
tested
this
extensively,
but
a
human.
If
you
have
a
trauma
to
part
of
your
grandmother-in-law
Sephardi
brain
or
you
lost
some
of
the
inputs
to
the
brain
facing
there
were
four
months
heard.
These
things
will
movie
to
compensate
as
much
as
possible.
It
was
a
small
enough
trauma.
You
can
completely
recover
your
abilities.
B
B
If
it's
not
too
big
an
error
because
to
be
there,
you're
closed
so,
but
if
it's
smaller
and
then
everybody
kind
of
read
all
the
Act,
all
the
inputs
they
equalize
out
and
our
spatial
pool,
or
does
that
I
was
reachable
because
of
their
and
winner
take
in
winners
situation
and
there's
a
boosting
situation.
If
some
of
them,
if
some
of
the
columns
and
we'll
talk
about
the
spatial
enrollment
is
some
of
the
columns
are
not
getting
input,
then
they
will
try
to
connect
to
the
other
nearby
and
when
they
connect
those
other
inputs.
B
Nearby
I
was
looking
at
these
sort
of
inputs
and
then
you're
over
here
saying
you
start
stealing
some
of
mine
and
then
I
say:
okay,
you're
selling
to
the
land
and
some
of
my
neighbors
and
everybody
kind
of
shifts
over
a
little
bit
and
working
for
the
nicely.
Now
it
was
beautiful
fact.
As
far
as
I
know,
that
was
a
first
normal
simulation
that
actually
captured
that
property.
You
know.
B
B
Okay,
basically
doing
we're
taking
some
input
space
and
dividing
it
up
right
in
some
sort
of
basis
set
I
and
and
so
when
you
think
about
vision,
the
monthly
simplest
basis
that
would
be
maybe
lying
orientation.
So
if
I
look
at
the
patterns
in
some
section
of
the
retina,
what
are
the
most
common
patterns?
I'm
going
to
see
it
would
be,
it
would
be.
The
most
common
would
be
edges.
Were
you
know
the
portafilter
time
so
and
then
it
says:
okay,
but
let's.
E
B
And
then
we
complex
function
of
untimely
me
calm.
It's
really
complex,
like
you've
got
110
Styles
you've
got
many
many
cell
types.
You've
got
all
these
different
layers
in
here,
so
we
just
said:
why
would
we
just
put
one
simple
there
like
a
layer,
four
we're
just
getting
the
infos
and
within
that
layer?
If
I
said
they
were
like
six
layers
here,
and
maybe
they
would
be
very
roughly
I
could
say
there
might
be
16
cells
in
here.
Just
six.
B
B
B
So
we
just
said:
okay,
we're
just
gonna,
take
a
bunch
of
any
comms
feed
it
some
variables
and
we're
gonna
have
them
compete
with
one
another
and
sort
of
learn
how
that
works.
Now
a
couple
of
the
advantages
by
the
way,
another
compact
more
to
them
special
pullin
among
them,
because
well
you
have
16
cells
in
each
one.
One
of
the
things
we
knew
from
biology
is
have
we
didn't
have
a
recognized
pattern
like
hearing
the
sequence
that
you
or
the
visual
scene
or
something
you
can
recognize?
Well,
you
have
as
much
sparser
activity.
B
This
is
an
observed
property
of
grunts
right,
so
we
know
the
unexpected
in
input
to
get
a
lot
of
activity
and
we
never
expected
it
to
it.
You
have
a
and
so
I
simplify
that
I
said.
Okay,
let's
assume
that
we
have
an
unexpected
input.
All
these
neurons
become
active,
but
again
a
predicted
input
unexpected,
but
then
maybe
fewer,
not
this
for
a
while.
B
B
Right
so
one
of
sixteen
sixteen
way.
So
if
I'm,
representing
some
input
with
20
active
bits,
then
each
one
can
be
represented
is
one
of
sixteen
lady
got.
You
got
twenty
to
the
sixth
or
60
to
the
twentieth,
different
ways
of
representing
attempt
a
huge
number.
You
know
and
going
much
you
know
going
up.
We
did
a
lot
of
our
32
now
once
you're
in
the
good
number
you're
sixteen
large.
B
So
this
idea
of
these
two
representations,
many
common
representation,
which
represents
the
base
input
and
then
individual
self
active
in
those
many
columns
represents
that
and
put
in
a
particular
context,
and
the
capacity
of
that
is
extremely
high.
So
you
can
represent
a
statement
but
in
truly
different,
no
problem,
so
that
was
one
of
the
core
things
you
needed
for
for
sequence.
Remember,
it's
be
able,
take
an
input
represent
and
very
high
ORAC
context,
and
then.
A
B
B
C
B
B
B
With
very
skinny
projections
vertically
up
and
down
like
this,
and
they
offended
all
of
these
cells,
they
define
in
some
sense
of
any
column.
They
are.
The
functional
definition
in
comes
a
name
is
what
we
learned
later
was
that
these
cells
also
have
the
same
orientation
property,
and
they
also
are
learned
that
foundation
property
plucked.
So
this
is
exactly
what
we
needed,
because
he.
B
B
B
Partially
organizer
and
then
there's
the
other
hand,
doing
the
normal
operation
of
the
system,
the
temporal
memory
with
you.
So
these
are
like
this
would
be
like
a
bipolar
cell
or
a
double
portunity
cell.
There
are
also
these
basket
cells
which
stand
out
there
all
over
the
places,
the
most
common
inhibitory
cell
and
they're
very
tightly
coupled
very
pronounced.
Now
what.
B
B
B
The
input
comes
in
all
these
cells,
a
spike
or
two,
but
then
what
we
eat.
What
we
require
after
that
is
that
then
there's
gonna
be
a
winner,
because
we
have
to
take
or
whatever
continuous
learning,
but
initially
they
all
spike.
And
then
we
pick
a
winner
or
one
of
those
depolarize.
We
just
go
right
to
that
one
there
and
we
don't
so.
These
mechanisms
support
that.
B
B
B
I
still
think
I
still
go
back
to
I,
don't
see
the
difference.
It's
the
same.
Three
cells
became
active
every
time
in
a
particular
context.
I
have
to
say
under
I
have
to
say
on
the
two
related
context,
somehow
generalize
that
I
would
pick
a
subset
of
the
same
cells.
I
know
that
it's
interesting
question.
It's
also
not
as
simple
as
this.
B
B
E
B
Sparse
vacation
and
it
throws
that
out
the
window
where
that
they
overlap
properties
just
off,
and
should
we
try
to
begin
to
bring
it
back
in
some
way
like
you
just
brought
it
yeah
there's
a
bit
of
perennial
question,
so
this
whole
idea
of
the
semantic
representation
in
the
SDR
and
this
participation
pool
output.
That's
that's
the
basic
of
all
of
the
particles
work.
They
start
like
an
old
semantic
representation,
but
as
Martin's
point
out
soon
as
we
go
to
this,
like
random
selection
in
a
new
we've
lost
it
now,
you
just
have
this.
B
You
know
there's
one
power
and
it
just
means
this
is
the
23rd
in
a
wallop
efficient,
beethoven's,
you
know
4/5
and
that's
all
it
means,
and
it
doesn't
mean
anything
else,
it's
ass
it
more
than
thing
you
need
in
the
world.
I
can't
relate
it
to
something
else.
This
idea
of
lack
of
generalizations
always
bothered
me,
but
I
never
saw
way
out
of
it
and
more
in
my
recent
work
and
what
Mark
has
proposed
with
the
displacements
of
else
I'm
more
hopeful
about
that
I
think
this
placement
cells
inherently
represent.
B
B
B
B
B
B
Pattern
and
goblet
responding
the
first
time
right,
but
now
it's
in
some
temporal
and
spatial
content
and
what
they've
shown
is
that
these
cells
do
not
fire
most
of
the
time
when
you
would
expect
them
the
fire.
So
they
only
be
spark
certain
parts
in
the
video
and
they
fire
fairly
reliably
at
that
point,
so
they
seem
to
say
I'm
gonna
fire
that
particular
context
119
put
but
mostly
other
times
like.
B
And
so
that
data
fits
beautifully
with
all
this
and
I
haven't
seen
anything
I
can
enroll
at
each
other
now,
but
they've
done
this
vision
and
the
general
rules
seem
supply
even
sort
of
the
level
of
when
I
read
those
papers.
I
say
always
defining
about
the
right
now
hi,
you
know,
and
it
is
so
there's
a
lot
of
evidence
that
this
basic
mechanism
is
actually
under
play
paper.
So
people
look
at
me,
how
could
you
know
you
know?
B
B
Think
they
don't
test
that
again,
but
they
say
okay,
what
does
this
cell
respond
to
responds
reliably
under
under
these
very
basic
test
conditions?
The
way
they
determine
they
come
up
with?
What
does
that
sound
like
and
they
require
to
realize?
Okay,
that's
all
your
fires
from
this
condition,
and
then
they
put
it
in
a
temporal
context
of
realistic
context,
both
temporal
and
spatial
context,
and
it
doesn't
fire
you're
liable.
It's
still
only
only
fires
when
they
have
that
input,
but
it
doesn't
do
it
on
the
most
complex.
B
B
Else
so
an
iteration
is
not
generally
fair
to
do
it.
As
a
I
mean
you
can
download
this
work
with
a
pussy-ass
snail,
they
have
a
simple
sea
slug
and
they
hope
they
poke
it
enemy,
traction
skill,
that's
that's
its
behavior,
you
know,
and
then,
after
a
while,
you
keep
poking
in
people.
Then,
after
a
while,
it
doesn't
attract
them
skills,
that's
kind
of
learning
and
you
might
call
that
habituation.
Not
they
call
it
line.