►
From YouTube: A Framework for Intelligence (paper review) -- The Thousand Brains Theory of Intelligence
Description
Yes, we're reviewing our own paper. :P Two newer Numenta hires are going to review our latest theoretical neuroscience paper. This is more for the benefit of the new hires to completely understand the Thousand Brains Theory of Intelligence.
D
B
C
F
B
D
E
My
hierarchy
is
present
in
the
neocortex
and
then,
whereas
grid
cells
are
present
in
the
complex,
both
you
know
I'm
just
going
to
go
into
just
a
very
like
high
of
high-level
idea
of
it,
but
he
of
this
notion
of
like
regions
in
the
New
York
cortex
that
sort
of
pass
up
information,
and
that's
like
the
flow
of
information.
As
you
sort
of
passed
it
passed
up
to
higher
and
higher
levels.
The
feature
complexity
gets
higher.
E
E
F
B
B
C
F
E
Okay,
so
that's
the
motivation
of
the
motivation
of
a
lot
of
dimensions.
Work
is
understanding
like
what's
actually
happening
to
your
expertise
like
the
nose.
99%
is
those
connections.
What
are
they
actually
doing?
Why
are
they
important
so
at
the
core
of
this
or
sort
of
a
the
main
kind
of
center
for
computational
unit
has
been
reviewed
by
Jeff.
So
in
a
previous
video
and
it's
kind
of
overview
series,
so
I'm
go
and
to
the
death
that
he
did,
but
just
to
kind
of
give
a
broad
idea
that
this
had
behaved
like
a
cortical
column.
E
So
I
drew
it
here
in
this,
like
big
square
being
the
cortex
and
your
cortex,
smaller
square,
being
the
cortical
column
roughly
about
a
millimeter
by
millimeter.
It's
not
physically
Demark,
but
it
has
this.
But
the
notion
is
that,
within
this
millimeter
by
millimeter,
those
neurons
have
the
same
receptive
field.
So
I
take
that
as
like,
if
you
think
about
the
but
from
like
the
part
of
the
finger,
maybe
that
would
be
like
one
cortical
column
and
maybe
like
the.
E
So
so,
with
that,
we
sort
of
want
to
understand
what
these
columns
are
doing
and
thinking
about
like
just
the
high-level
idea
of
the
paper
is
that
within
these
columns
originally,
it's
just
thought
that
grid
cells
exists
in
the
became
a
complex,
maybe
parts
of
the
neocortex.
But
this
paper
is
asserting
that
sort
of
quit
so
like
neurons,
exist
in
every
single
column
and
essentially
the
those
columns
are
learning
over
location
based
representations.
C
C
It's
using
what
we
call
a
forest
cooling
course:
cooling
I
had
these
defenses
squares
here,
city
of
different
blogs,
ABC
and
I
can
see
Forex
block
if
they
can
sing
the
block
or
it's
not
in
the
blood,
and
course
coatings
can
allow
me
to
have
better
representation,
more
precise
representation
in
this
case,
if
I'm
closer
to
the
center
and
a
more
master
stylist
or
a
less
precise
representation
in
the
in
the
outskirts
of
a
space.
So.
C
C
So
here
we
had
2
^
3,
so
we
have
8
possible
ways
of
representing
our
and
the
gasoline,
so
it
just
I
can
I
want
this
box.
Let's
write
it
like
joining
four
pieces
and
then
I
have
5
to
the
power
of
3,
so
Vaden
can
be
in
one
of
these
four
bucks
or
not
in
the
book
at
all.
So
I
have
five
options
per
watt.
The
first
we
have
125
days
of
purpose
entities
instead.
C
C
C
C
C
F
B
B
B
C
G
B
D
C
B
B
B
A
C
And
I
can
explain
itself
so,
okay,
so
thanks
I
should
be
ones
well,
the
locations
were
oneself
are
the
locations
where
one
cell
fires
right
and
then
square
would
be
another
one,
and
so
it
will
be
another.
So
if
only
expires,
you
could
be
many
of
those
location
classes
and
then
X
n
square
fires.
It
could
be
in
any
location
with
X
in
space
of
attacks
and
circle
fire
so
that
sense,
amplifier
theater,
like
what
they're
close
together
yeah.
C
A
F
F
D
B
F
F
F
F
F
F
A
F
F
B
F
B
B
G
Anything
help
you
understand
it
to
the
other
way.
Maybe
now
these
were
doing,
if
you
imagine
how
to
do
just
regular
numbers,
there
are
ways
they're
kind
of
happy,
but
there
are
ways
and
there
it
can
be.
It
can
be
a
useful
exercise,
but-
and
it
does
not
people
that
understand
it
know
that
you
can
go
and
and
woman
all
of
us
using
numbers
if
you
want
to
like,
if
you
just
use
3d
space
within
origin
and
just
usually
I,
don't
know,
use.
A
G
G
B
You
good
back,
you
know
it's
just
there's
a
whole
list
of
properties
that
I'm
satisfied
and
the
hardest
ones
are
the
one
we
don't
know
it
does
satisfy.
We
don't
know
about
power
yet,
and
so
you
know
it's
almost
impossible
to
come
up
with
something
else
that
could
satisfy
stuff.
You
don't
add
you
don't
know.
Yet
all
we
know
is
that
the
brain
does
work.
F
F
C
C
C
D
B
F
F
For
any
kind
of
metric-
or
you
can't
do
path,
integration
with
just
essentially
drive
so
an
individual
player
might
respond
to.
This
call
me
come
into
my
phone,
but
it
wouldn't
respond
to
this
to
where
these
missiles
always
and
regulars
and
so
metric
space.
Could
you
can
you
measure
calculations
or
integration.
F
F
E
F
D
G
F
E
E
So
I
take
that
sort
of
like
the
main
kind
of
important
parts
of
the
grid
cells.
Being
you
have
a
way
to
disambiguate
your
location,
and
then
it
has
a
very
like
a
lot
of
like
nice
properties
to
it,
which
I
guess
we
don't
like
fully
understand,
one
of
them
being.
It
has
like
a
very
like
high
representation
capacity
you
can
like
represent.
E
E
Okay,
so
with
that
in
mind,
I'm
gonna
kind
of
go
over
the
main
idea
of
the
grid
self
paper
and
like
a
little
bit
Lucas,
who
sort
of
go
over
the
kind
of
computations
that
are
done
sort
of
like
what
they
say.
The
main
idea
is
that
each
column
has
not
only
sort
of
information.
The
sort
of
in
this
higher
will
be
where
the
information
is
just
like.
E
D
E
They
kind
of
say
well,
okay,
maybe
have
like
point
a
my
B
Point
C
and
those
are
like
unique
locations.
Then
you
have
waking
features
that
so
maybe
my
fingertip
may
sort
of
like
represent
the
kind
of
like
point
it's
at
this
edge,
the
sort
of
curvature
beside
me.
This
is
fine
to
the
bottom,
and
so
every
single
column
has
this
sort
of
representation
of
location
and
associated
features
and
the
mechanism
for
driving.
That
is
great
spots,
and
that's
like
phenomenal
idea.
E
So,
but
before
the
paper
there
was
I
guess
some
evidence
that
there
good
cells
in
the
New,
York
cortex,
and
this
is
papers
asserting
that
good
cells
are
everywhere
in
New
York
four
times
every
single
column
has
its
own
sort
of
a
set
of
grid
cells.
I
guess,
therefore,
its
own
set
of
modules
of
grid
cells
that
are
constantly
associating
features
with
their
respective
locations
and
using
efferent
motor
commands
I.
Guess
a
copy
of
this
sort
of
movements
of
your
finger
or
whatever
senses.
E
E
You
need
a
location
within
the
objects.
Reference
frame.
You
need
some
ideas
like
than
the
other
experiment
spring,
and
each
problem
needs
this,
which
I
thought
was
actually
like.
An
interesting
assertion.
So
I
didn't
read
that
paper,
but
I
will
try
to
give
my
own
motivation
for
why
it
could
meet,
and
this
is
the
this
is
where
feedback
I
guess
could
be
helpful
to
motivations
and
then
proscribe
like
one
first.
So
what
I'm
talking
about
it
is
like,
if
you
sort
of
had
some
service.
E
E
So
you're
kind
of
imagining,
like
so
you're
kind
of
putting
yourself
in
the
position
of
the
tip
of
the
finger
and
you're
trying
to
predict
like
what's
going
to
come
next,
and
so
let's
say
you
move
to
the
left
and
all
of
a
sudden,
it's
smooth
sort
of
without
having
any
idea
of
likely
your
location
or
idea
of
like
where
you're
moving
like
either
know.
If
you're
matching
them
to
the
lecture,
you
don't
know
if
you've
got
to
move
to
the
right,
so
it's
not
actually
any
way
for
that
finger,
just
hurt
without
the
location.
E
E
That
it's
touching
it
doesn't
know
where
it
is
on
the
object
or
what
it's
about
on
the
object
sounds
like
one
kind
of
thought:
experiment
that
I
had
with
this
so
I,
don't
know
if
that's
sort
of
I
guess
sort,
if
that's
a
good
way
to
think
about
it,
and
if
there
are
two
min
there
are
more
ways:
I,
better
waste.
It
certainly
think
about
like
why
you
actually
have
had
in-state
location
information,
yeah.
F
F
F
D
E
Well,
you
only
make
sense
for
sort
of
like
lower
level
senses.
I
can
definitely
make
sense
for
things
that
are
like
the
eyes
for
moving
in
the
environment
or
your
hands
like
movin
over
a
coffee
cup
and
70s
I.
Think
about
like
lower
levels
in
the
clinic
or
like
hierarchy.
It
definitely
makes
sense,
but
I'm
also
trying
to
think
about
like
for
the
higher
levels
as
well,
but
as
higher
levels,
we'll
just
sort
of
the
extract
features
at
the
lower
levels
and
they
also
I.
F
F
F
D
F
F
F
And
so
that's
where
your
evolutionary
argument
comes
in
and
the
common
algorithm
harmony
comes
in
and
common
architecture
argument
comes
anything
we
can
then
say
well,
could
location
cast
be
the
first
approach
of
the
purush
you
similar?
Let's
go
into
evidence
which
suggests
that
likely
occur.
Can
we
make
it
work?
That's
morally!
F
F
F
D
F
F
F
F
F
F
E
E
Brief
evolutionary
thing
which
is
just
to
be,
but
this
is
a
sort
of
argument
that
told
me
I,
sort
of
liked
it
but
sort
of
okay.
Now
we
sort
of
suppose
that
the
concepts
have
a
notion
of
location
or
is
it
motivation
that
it's
crystals
and
the
sort
of
evolutionary
arguments
that
the
brain
has
you
know,
spends
many
years
sort
of
evolving
this?
E
This
mechanism,
the
hippocampal
complex
to
identify
a
location-
and
it
has
like
all
these,
like
really
amazing
and
powerful
properties
and
sort
of
kind
of
chances
that
it
somehow
reinvented
its
own
wheel
and
sort
of
implemented,
almost
a
very
identical
thing
in
hit
and
the
neocortex
it
just
seems
like
sort
of
like
far
fetched,
though
it's
sort
of
when
make
sense
that
it
sort
of
readapted
that
the
mechanism
for
functionality
of
your
cortex
I
thought.
That
was
like
an
interesting
perspective
of
motivating.
You
know
why
is
it
grid
cells?
You
know
this
is
how
early.
F
A
C
C
C
C
C
F
C
C
Special
inspector
is
taken
from
point
A
to
B.
Here
we
can
also
take
me
from
one
space
to
another
from
curve
to
handle,
for
example,
so
this
blue
line
will
also
be,
and
special
factors
and
special
fibers
will
be
sort
of
transformations
of
their
torn
in
doing
translation,
and
we
also
discussed
yesterday.
This
could
be
other
sort
of
transformations
and
then
Mario
collects
ventures
white
by
displacement,
actors
fixing.
We
wonder
how
the
brain
works
and.
E
C
F
E
That
was
the
whole
motivation:
okay,
okay,
so
summarizing
exactly
what
is
the
thousand
brains
theory
because
has
sort
of
contextualized
the
title
paper
just
yet
and
sort
of
like
how
this
fits
into
the
traditional
like
heart,
I,
hope
you,
the
neocortex.
So
it's
a
it's
a
thousand
brains
theory,
because
the
idea
is
that,
since
each
column
has
its
own
representation
of
the
location
and
the
Associated
features
of
those
locations,
essentially
each
object
is
now
sorry.
E
Each
column
is
capable
of
creating
the
model
of
that
object
and
in
a
way,
it's
sort
of
a
brain
internet
itself
and
the
sense
that
it's
knowing
these,
it's
a
it's
a
modern
these
objects
and
making
predictions
on
those
objects
and
I.
Don't
know
how
many
columns
there
are
in
a
brain.
But
presumably
it's.
F
Right,
not
every
column
models,
everything
so
I
can
only
certain
sensory
things
and
certain
size
and
I
took
my
finger.
Cortex
wouldn't
know
anything
about
democracy
or
logic.
Look
like
or
what
they
sound
like.
So
I
originally
want
to
start
writing,
100,000
brains,
theory.
It
was
too
complicated,
so
we
just
wanted
a
thousand
brain
theory.
My
excuse
for
using
that
is
that
we're
not
talking
about
the
number
of
columns
in
the
cortex,
but
in
the
number
of
models.
F
A
Have
an
interesting
illustration
of
this:
have
you
think
about
some
objects?
You
only
interact
with
in
certain
ways,
think
of
a
horse.
You've
got
tons
of
models
in
v1,
v2
v3
you've
got
models
where
you
touch
horses
with
your
hands,
but
then
the
rest
of
your
body
is
pretty
empty
unless
you've
ridden
a
horse
in
which
you've
got
models
and
your
thigh
areas.
You
know,
because
you
know
what
a
horse
feels
like
to
sit
upon.
F
E
Okay,
so
there's
on
the
order
of
magnitude
somewhere
around
thousand
bombs
that
gets
beyond
this
whole
debate.
Yeah.
It's
only
been
okay,
but
it
captures
this
idea
that
there
are
many
models.
Let's
say
that
there
are
many
models
of
the
same
object.
You
know
sort
of
distributed
that
through
these
cortical
columns
and
then
sort
of
like
understanding
what
that
means
in
as
far
as
the
hierarchical
model
it
doesn't,
it
doesn't
contradict
micro
model,
but
it's
sort
of
sort
of
explains
are
just
sort
of
pens
like
what's
actually
through
the.
F
F
D
C
Projects
yesterday,
I,
don't
know
how
these
particular
that
there
are
two
ways
you
could
recognize
a
contract.
I
can,
for
example,
I
know,
I
have
only
this
part
of
a
column
with
associated
with
this
small,
since
birth
back
and
I
can
either
like
move
it
through
time,
and
then
here
I
only
get
like
one
point
here:
not
get
chosen
and
get
free
and
the
other
way
I
can
do.
It
is
I,
have
other
cortical
columns
with
other
sensory
patches.
This
is
the
base
of
the
finger
and.
D
C
I
can
either
go
that
way
or
that
thing.
So,
if
I
move
my
finger,
I'm
going
to
get
more
information
at
that
specific
cortical
column
in
its
kind
of
recognizable
character,
the
other
way
I
can
improve.
My
classification
say
is
I
only
have
like
one
secret
location
that
I
have
it
for
several
different
models
and
then
I
can
get
to
the
same.
C
F
Is
different
nation
ships
by
trying
to
balance
it
for
here
it
has
different
relationships
now
in
its
sharing
with
other
objects,
but
each
one
of
these
models
is
really
how
you
relate
to
this
thing,
so
I
think
it
gets
to
the
deeper
level
of
this
courthouse.
Each
one
of
these
models
is
actually
a
relationship.
F
F
So
the
fact
that
you
have
a
large
array
of
sensors
in
the
eye
fool
people,
the
thinking
of
the
problem-
is
a
spatial
or
a
problem.
Whether
really
the
problem
is
a
motor
or
sequence
problem
and
there,
and
when
you
saw
the
motor
secret
problem,
I
recommend
something
with
a
single
finger
who
understood
so
much
better
when
you
have
a
bunch
of
them.
Okay,
we
just
do
the
shortcut
we're
just
we're.
We
have
all
the
different
models
with
your
movement
model
so
tomorrow.
F
F
E
Okay,
so
going
back
to
the
there's
one
other
corners,
you
can
say
with
the
hierarchy,
which
is
oh
I,
guess
I
was
actually
is
going
to
give
like
just
a
sort
of
simple
example
like
if
you
kind
of
think
about
a
column
and
be
one
passing
up
information
to
convey
to,
and
you
may
think
of
sort
of
the
this
v1
column
being
sort
of
associated
with
one.
You
know
sensory
patch
of
the
of
the
eye
that
sort
of
you
know,
column
says
you
know,
I
see,
I,
see
curvature.
E
It
says
you
know
I
think
it
may
be
the
top
of
a
coffee
cut
for
it's
trying
to
you,
know,
sort
of
predict
or
trying
to
get
a
grasp
of
its
location
and
then
I
guess
in
this
model
those
two
things
are
sort
of
put
together
into
a
model
that
normally
just
includes
information
about
the
curvature,
but
also
the
information
about
the
location
of
thing
on
the
cup
and
therefore
sort
of
a
model
of
the
cup
itself.
And
that's
what's
good
to
pass
on
to
on
TV
and.
E
E
Let's
say
if
we
just
want
to
consider
like
region
to
region
connections.
Apparently
40%
of
all
readin
Creek
region
connections
are
possible
or
exist
in
in
the
neocortex
and
then,
when
you
sort
of
think
about
that,
if
you
sort
of
relying
information
about
the
object
in
these,
you
know
across
regions.
E
It
sort
of
paper
that,
with
that
you
can
sort
of
have
this
idea
of
like
voting
sort
of
Jim,
getting
a
general
consensus
on
what
the
object
is,
and
you
know
maybe
the
I
says
it's
like
part
of
the
top
of
like
you
know
a
bucket
and
you
know
the
hand
says:
maybe
it's
like
the
top
of
a
come
together.
That's
definitely.
F
B
If
there
was
a
pure
hierarchy,
the
higher
levels
would
be
depending
on
you
wouldn't
have
this
yeah,
but
if
you
have
v2,
you
know
getting
really
different
sensory
input,
lauric,
witty
higher
reason.
It's
going
to
make
different
types
of
errors
than
v1
would
and
that's
much
better
from
a
mixture
of
experts
and.