►
From YouTube: Thousand Brains Theory & Hierarchy (Episode 16)
Description
The Thousand Brains Theory of Intelligence says that each cortical column learns models of complete objects. There are many models... models for vision, models for touch. Cortical columns vote within regions and across the brain to reach a group consensus on object representations. Of all the models in our brains, we are only really aware of the consensus model, which is why we have a singular perception of reality.
Hierarchy plays a different role than we currently model today's machine learning networks. Watch this HTM School to understand why.
A
A
These
regions
of
neocortex
all
contain
thousands
of
copies
of
this
unit
and
are
connected
together
in
a
messy
and
loosely
hierarchical
way.
Let
me
take
you,
through
the
classic
model
of
hierarchy
in
the
brain,
how
most
researchers,
especially
in
machine
learning,
see
hierarchy
today
in
the
classic
model
of
hierarchy.
Input
comes
from
the
senses
and
flows
into
regions
in
the
bottom
level
of
the
hierarchy.
These
regions
identify
low-level
features
like
edges.
These
representations
are
input
to
the
next
level
of
the
higher
as
we
ascend
the
hierarchy.
A
Input
from
the
lower
levels
is
used
to
compose
larger
shapes
or
object
parts
again
these
representations
converge
in
a
higher
level
and
at
the
top
of
the
hierarchy,
complete
objects
are
constructed
from
these
converging
features.
Deep
learning
depends
heavily
on
this
type
of
classical
hierarchy,
as
you
can
see
in
this
graphic
from
a
popular
deep
learning
book,
some
widely
used
deep
learning,
architectures,
like
residual
networks,
use
the
tactic
of
adding
lots
of
levels
to
the
hierarchy.
A
This
is
the
way
deep
learning
hierarchy
works,
but
it's
not
the
way
the
brain's
hierarchy
works,
the
hierarchies
and
our
brains
are
not
deep.
They
are
wide
and
really
tangled
together.
Let's
look
at
this
from
the
thousand
brains.
Theory
point
of
view
where
we
assume
that
every
cortical
column
in
every
region
at
every
level
of
hierarchy
is
essentially
performing
the
same
computation.
A.
A
Crucial
aspect
of
this
theory
is
the
importance
of
movement.
Our
brains
build
a
sensory
motor
model
of
the
world,
so
in
many
ways
all
objects
are
encoded
within
the
context
of
movement.
If
you
reach
your
hand
inside
a
box
and
trying
to
recognize
an
object,
you'll
move
your
hand
over
the
object
in
order
to
define
its
shape
with
respect
to
your
sensors.
A
Your
eyes
do
the
same
thing
and
you're
largely
unaware.
As
you
look
at
an
object,
rapid
eye
movements
called
saccade
move
your
field
of
view
all
over
the
scene
jumping
from
place
to
place,
searching
your
visual
perspective,
all
the
while
resolving
exactly
what
is
being
sensed
across
the
sensory
hierarchy,
which
is
voting
in
parallel.
A
You
can
experiment
with
this
idea
easily
just
get
a
straw
and
try
looking
through
it.
Static
image
like
this
could
be
a
horse
tail
or
something
completely
different.
The
straw
limits
your
sensory
input
to
mostly
the
high
detail,
but
small
field
of
view
in
the
bottom
of
the
hierarchy.
So,
as
you
move
your
straw,
you
must
be
building
up
a
complete
representation
of
the
object
only
in
the
bottom
of
the
hierarchy.
A
A
Different
levels
of
hierarchy
have
different
fields
of
view
of
the
sensory
input
at
the
lower
levels.
The
field
of
view
of
a
single
cortical
column
is
small,
but
very
detailed
as
you
ascend
the
hierarchy.
The
size
of
a
cortical
columns
field
of
view
increases,
but
the
level
of
detail
decreases.
The
highest
levels
have
the
broadest
field
of
view,
but
the
least
detail.
A
Every
cortical
column
in
every
region
in
this
hierarchy
is
modeling
objects,
and
even
though
there
are
some
hierarchical
connections,
the
vast
majority
of
them
are
non
hierarchical.
That
means,
if
one
cortical
column
has
a
clue
what
the
object
is.
It
can
vote
with
its
neighbors
to
come
to
a
group
conclusion.
This
allows
a
local
clue
to
spread
through
the
neocortex
and
the
hierarchy
as
an
object
is
identified,
so
any
level
of
the
hierarchy
can
potentially
inform
any
other
levels
as
columns
vote,
they
resolve
towards
an
agreed
object,
representation.
A
Let's
think
about
this
in
terms
of
one
object
being
sense:
visually
at
different
distances
at
the
lowest
level
of
the
visual
hierarchy.
The
field
of
view
of
one
cortical
column
is
about
the
size
of
your
thumb
at
arm's
length.
Although
the
field
of
view
is
small,
the
resolution
of
this
sensory
input
is
really
good,
so
we
can
easily
make
out
a
horse
standing
on
the
horizon,
but
what
about
at
the
top
of
the
hierarchy?
The
sensory
input
we
get
here
is
very
broad,
but
the
details
are
missing.
Is
that
blob
a
horse
or
an
elephant?
A
The
lower
levels
of
the
hierarchy
might
not
be
able
to
identify
the
object,
but
perhaps
the
higher
levels
could
help
them
out.
In
summary,
the
thousand
brains
theory
of
intelligence
says
that
each
cortical
column
learns
models
of
complete
objects.
There
are
many
models,
models
for
vision,
models
for
touch,
cortical,
columns,
vote
within
regions
and
across
the
brain
to
reach
a
group
consensus
on
object,
representations
of
all
the
models
in
our
brains.
We
are
only
aware
of
the
consensus
model,
which
is
why
we
have
a
singular
perception
of
reality.
A
Hierarchy
plays
a
different
role
than
we
currently
model.
In
today's
machine
learning
networks.
The
hierarchy
is
serving
other
purposes
like
object,
composition,
we're
still
trying
to
understand
the
details,
but
we
know
that
a
lot
more
is
happening
within
just
one
level
of
the
hierarchy
than
we
ever
thought.
A
So
I'm
not
sure
when
I'll
have
another
video,
but
please
subscribe
to
this
YouTube
channel
where
we
stream
live
new
mental
research
meetings
weekly,
you
can
keep
up
with
our
very
latest
research
and
even
chat
with
me
in
our
online
community.
If
you're
interested
in
that,
you
can
meet
with
other
people
interested
in
HTM
and
biologically
inspired
intelligence
on
HTM
forum,
our
community
discussion
board.
Here
you
can
find
Co
implementations,
neuroscience
paper
discussions
and
good
old-fashioned
arguments
about
the
meaning
of
life
itself
and
the
purpose
of
humanity.