►
From YouTube: What the Brain says about Machine Intelligence
Description
"What the Brain says about Machine Intelligence"
Jeff Hawkins
Co-founder, Numenta
21 Nov 2014
A
Hi,
I'm
jeff
hawkins
and
my
talk
today
is
about
machine
intelligence,
and
the
title
of
the
talk
is
what
the
brain
says
about
machine
intelligence.
So
what
nomento?
That's
what
we
do
we're
working
on
machine
intelligence
and
we
study
the
brain,
and
this
talk
is
about
how
the
brain
informs
us,
how
we
might
go
about
building
intelligent
machines.
A
Let
me
start
with
an
analogy:
this
goes
back
to
the
1940s
and
in
the
1940s
we
were
witnessing
the
birth
of
programmable
computers,
and
so
at
that
time
they
were
just
starting
to
build
computers,
and
it
was
very
confusing
time.
There
was
a
lot
of
there
wasn't
a
lot
of
agreement
about
how
to
go
about
doing
this
so
different
approaches.
A
Some
people
are
building
dedicated
machines
that
were
designed
to
solve
a
particular
problem,
and
they
generally
did
a
better
job
than
people
who
are
designing
more
universal
machines,
but
they
can
be
applied
to
many
problems.
Some
people
built
analog
computers
because
they
were
better
at
certain
problems
and
other
people
built
digital
computers
because
they
were
better
at
certain
problems.
There
was
a
debate
between
whether
digital
computers
should
be
decimal
or
binary.
A
There
was
debates
about
how
the
programming
should
work
and
what
kind
of
memory
architectures
there
should
be,
and
so
this
is
a
very
confusing
time,
but,
as
you
left
the
1940s
and
we
entered
the
1950s,
we
settled
on
a
single
dominant
paradigm
for
computing,
the
one
we
still
use
today
and
so
that
dominant
paradigm.
We
have
universal
machines,
they're
not
dedicated
they're
digital
they're
binary,
they
have
a
memory
based
programming,
they
have
a
two-tier
memory
system
and
so
on,
and
this
is
what
we
built
all
of
the
computing
for
the
last
70
years.
Now.
A
Why
did
we
settle
on
one
paradigm?
Why
did
one
paradigm
win
one
because
of
network
effects?
Network
effects
are,
is
a
term
that
refers
to
when,
when
you
have
a
leader
in
some
area,
more
people
invest
in
that
leading
solution
and
it
gains
momentum
and
it
moves
away
from
the
other
solutions,
and
so
we
tend
to
end
up
with
a
single
solution
or
single
type
of
paradigm
in
technology.
A
Now,
why
did
this
particular
paradigm
win?
And
the
answer
is
because
it
was
the
most
flexible
and
the
most
scalable
it
isn't.
Computers
are
not
the
best
solution
for
all
problems.
You
can
almost
always
design
a
better
solution
to
a
particular
problem
than
programming
computer,
but
they
are
they're
universal,
so
we
can
apply
them
in
to
many
many
different
problems
extremely
flexible.
A
They
also
can
scale
from
the
very
small
embedded
in
you
know,
an
appliance
to
room
size,
computers
and
it's
these
attributes
that
led
us
to
saw
to
choose
one
dominant
paradigm
and
the
one
we
did
now
today
in
the
in
the
2010s
we're
in
a
very
similar
period.
We
are
witnessing
the
birth
of
machine
intelligence
right
now
and
it's
a
confusing
time
just
like
it
was
in
the
1940s,
there's
different
ideas
how
this
is
going
to
play
out.
A
These
are
some
of
the
differences.
Some
people
are
working
on
a
problem.
They
come
up
with
a
specific
solution
that
says
this
is
the
best
solution
for
this
problem.
Other
people
work
on
more
universal
solutions
which
are
easier
to
apply,
but
may
not
be
the
best
solution.
There
are
different
approaches.
There
are
mathematical
approaches,
there's
memory
based
learning
approaches,
different
sort
of
fundamental
ways
to
think
about
learning,
there's
different
ways
of
training,
batch
versus
online
learning
or
labeled
versus
behavioral
based
learning.
A
So
it's
kind
of
a
confusing
point
in
time
right
now,
but
as
we
leave
this
decade,
in
fact,
even
before
the
2020,
I
believe
will
have
settled
on
one
dominant
paradigm
and
part
of
this
talk
is
to
argue
for
what
that
dominant
paradigm
is,
and
we
are
basing
our
argument
essentially
on
the
fact
that
the
brain
gives
us
an
example
for
this.
So
what
is
that
dominant
paradigm?
It's
a
universal
algorithm
for
machine
intelligence,
it's
not
specific
for
different
problems,
but
a
universal
one.
It's
memory
based
it's
online
learning,
which
means
it
learns
continuously.
A
It's
behavioral,
based
learning
and
so
on.
Now
I'm
going
to
talk
about
these
details
in
my
talk,
but
the
question
is:
why
are
we
going
to
settle
on
one
paradigm,
the
same
as
before
network
effects
once
you
start
getting
momentum
behind
something
it'll,
it'll
snowball,
why
this
particular
paradigm
same
as
before
it
is
the
most
flexible.
A
We
know
that
brains
can
solve
all
kinds
of
problems
and
in
tremendous
flexibility
in
the
human
brain
in
other
brains,
and
it
is
scalable.
We
know
from
nature
that
you
can
build
small
brains
and
large
brains.
So
how
do
we
know
this
is
going
to
happen?
We
have
a
proof
case
in
biology,
the
brain,
the
neocortex,
and
we
have
made
great
progress
in
understanding
how
it
works.
A
So
this
is
not
something
that's
going
to
happen
10
years
from
now.
Something
is
happening
right
now
and
the
rest
of
my
talk
is
going
to
go
in
detail.
What
this
is
about.
Okay,
our
company's
mission
is
essentially
who
have
two.
We
one
of
the
first
ones
discover
the
operating
principles
of
the
neocortex.
Now,
just
remind
you,
the
near
cortex
is
is
a
big
wrinkly
thing
on
top
of
your
brain,
it
is
about
75
of
the
volume
of
your
human
brain
and
it's
the
locus
of
all
intelligence
language,
hearing,
vision
and
so
on.
A
A
Now
our
goal
is
not
to
recreate
a
human
or
anything
anything
like
a
human
or
recreate
anything
like
a
particular
brain.
It's
basically
to
say
how
do
brains
work,
how's,
the
neurocryptics
work.
Let's
build
machines
that
work
on
those
principles.
It
is
not
to
pass
the
turing
test
or
to
build
something.
That's
human-like
all
right.
A
My
topics
in
my
talk
today,
I'm
going
to
give
you
start
off
with
some
cortical
facts
and
some
details
about
how
the
what
do
we
know
about
the
brain,
the
neurocortex
I'll,
give
you
a
high-level
description
of
the
theory
of
how
it
works.
I'll,
tell
you
about
our
research
roadmap,
so
you
know
where
we
are
in
this
process
and
then
we'll
talk
about
applications.
A
A
Your
eyes,
your
ears,
your
skin,
and,
although
we
think
of
those
as
light
sound
and
touch
once
you
get
beyond
the
side
of
the
senses,
it's
just
patterns,
there's
nerve
fibers
that
are
carrying
information
from
the
retina
and
the
cochlea
in
the
somatic
senses
into
the
neocortex,
and
those
neurons
are
identical.
They
don't
there's
no
distinction
between
the
ones
that
representing
light
and
sound
and
touch,
and
from
the
neocortex's
point
of
view,
there
is
no
light,
sound
and
touch
it
is
just
patterns,
it
is
a
pattern
system
and
it
treats
them
turns
out.
A
Amazingly,
it
treats
them
all
the
same
way.
So
this
influence
comes
streaming
into
neocortex,
rapidly
changing
over
time
and
the
neocortex
has
to
build
a
model
of
the
world
from
that
fast.
Changing
sensory
data,
that's
what
it
does
when
you're
born
your
neocortex
knows
nothing
about
the
world,
but
through
exposure
it
learns
how
the
world
behaves
and
it
builds
a
model
of
the
world.
This
model
is
a
predictive
model,
meaning
it's
constantly
predicting.
What's
going
to
happen
next,
you're,
not
even
consciously
aware
of
this.
A
That
is
constantly
predicting
what
you
can
see
here
and
feel
it's
a
predictive
model.
It
can
also
tell
when
things
are
different
anomalies,
and
it
also
generates
all
your
high
level.
Behavior
subsets
of
my
speech
right
now
is
coming
from
my
neocortex
now,
because
the
neocortex
generates
your
all
this
behavior.
When
you,
when
you
act
in
the
world,
you
actually
move
your
sensors
in
the
world
and
you
end
up
changing
the
inputs
that
are
coming
into
the
sensor.
A
So
actually
most
of
the
changes
that
are
occurring
on
your
sensory
organs
are
coming
from
the
fact
that
you're
moving
you're
moving
your
eyes,
you're
moving
your
head,
you're
turning
your
body,
you're
making,
sounds,
etc.
So
what
we
say
is
the
neocortex
learns
a
sensory
motor
model
of
the
world.
It
learns
how
the
sensory
data
changes
when
you
act
upon
the
world
and
from
that
we
can
do
goal-oriented,
behavior
and
all
the
things
that
humans
do.
So
we
want
to
know
how
it
does
this
in
detail.
A
So,
let's
just
dive
into
the
cortex
a
little
bit
more
in
detail,
we'll
start
with
a
picture
here.
This
is
a
human
ear,
cortex
and
right
next
to
it
is
a
rating
of
cortex.
I
show
the
rat
neural
cortex,
because
every
principle
I'm
going
to
talk
about
today
applies
to
all
neurocortex,
regardless
of
the
animal.
The
only
difference
really
between
a
rat
in
your
cortex
and
a
human
near
quartz
is
primarily
the
size
of
it
and
we'll
talk
about
that
in
a
second.
A
So,
no
matter
what
neocortex
you're
looking
at
it's
always
a
sheet
of
cells,
and
it's
about
about
two
and
a
half
millimeters
thick.
It's
a
pretty
thin
sheet
of
tissue
of
neurons
and
the
difference
again
between
a
human
and
a
rad
is
just
how
big
that
sheet
is
the
area
of
it.
Now
it's
remarkably
uniform,
you
know
in
two
ways:
one
is:
if
you
look
at
the
detail
in
this
sheet
of
cells
and
there's
a
lot
of
detail
there,
we'll
talk
about
in
a
second,
it's
remarkably
uniform.
A
You
can
take
the
auditory
nerve
and
the
visual
nerve
and
swap
them
and
the
auditory
cortex
becomes
visual
and
the
visual
cortex
becomes
auditory.
Okay,
this
sheet
of
cells
is
actually
organized
hierarchically.
That
is
the
different
regions
in
it
are
projected
to
each
other,
and
if
you
take
that
map
you'll
see
that
there's
a
hierarchy
of
regions-
and
so
information
comes
in
at
the
bottom
of
the
hierarchy
and
moves
up,
and
then
it
also
flows
back
down.
But
all
those
regions
look
the
same
and
doing
something
similar.
A
If
we
zoom
in
further
into
a
particular
layer,
any
layer
anywhere
any
near
cortex
you'll
see
that
the
first
level
of
detail
under
the
microscope
is
layers
of
cells.
There
are
multiple
layers
of
cells.
Typically,
four
depends
on
how
you
count
them:
we'll
label
them
layers,
two
three:
four:
five
and
six
four
layers
of
cells.
A
Now,
if
you
zoom
in
further
you'll
see
that
the
neurons
themselves,
the
cells
are
organizing,
these
things
called
mini
columns
and
they're
they're,
really
these
very
tightly
little
very
close
and
very
skinny
little
columns
of
cells
with
about
100
to
120
that
span
across
all
the
layers.
A
If
you
zoom
in
further
still
you'll
see
the
actual
neurons-
and
this
is
a
picture
of
a
classic
cortical
neuron
and
the
neuron
is
a
characterized
by
this
big
sort
of
tree-like
branching
structure
called
the
dendrites
on
it
and
they
have
all
these
connections.
You've
heard
about
called
synapses
so
anywhere
between
three
and
ten
thousand
synapses.
Ten
thousand
is
probably
closer
to
the
typical
number
of
connections
on
every
single
cell.
A
Now
it
turns
out
that
most
of
those
connections
are
close
far
from
the
cell
body
on
these
branches.
Those
are
called
the
distal
connections
and
only
about
ten
percent
are
close
to
the
cell
body
called
proximal
ones.
If
you
zoom
in
further-
and
you
make
a
picture,
look
at
a
picture
of
the
of
an
actual
dendrite
one
of
those
little
branches
that
coming
off
here,
you
can
actually
see
all
the
synapses
alone
and
they're
very
tiny.
Those
are
little
spines
all
arranged
along.
Those
are
the
acts
of
connections.
A
Now
we
now
know
something
we
didn't
know
even
20
years
ago
that
these
branches
are
active
they're,
not
just
parts
of
the
cell
body.
They
are.
They
are
active
processing
elements
and
what
they
act,
like
is
a
little
coincidence,
detectors
and
they
see
if
a
bunch
of
synapses
become
active
at
the
same
time
close
together
and
close
in
time,
then
it
actually
has
a
large
effect
on
the
cell
body.
A
Surprisingly,
almost
all
neural
networks,
today,
artificial
neural
networks
and
neural
models
do
not
account
for
all
these
synapses
and
these
active
dendrite
properties,
but
this
is
essential
part
of
neurons
now.
The
other
thing
we've
learned
recently
is
people
for
many.
Many
years
have
thought
that
learning
occurs
by
the
strengthening
and
weakening
of
synapses.
A
We
now
know
that
new
synapses
form
all
the
time
and
they
disappear
all
the
time
quickly
in
the
outer,
sometimes
minutes
or
even
tens
of
seconds,
and
so
learning
that's
a
much
more
powerful
part
of
learning
and
so
learning
actually
is
largely
about
the
forming
of
new
synapses,
not
just
strengthening
and
wicking
existing
ones.
So
this
picture
here
this,
if
you
think
about
this,
this
is
everything
that
we,
this
is
a
big
picture
of
your
brain
and
everything
you
do
everything
a
human
does.
A
Everything
you've
ever
done
in
your
life
is
operating
in
this
structure.
This
is
a
picture
of
what
intelligent
structures
in
biology
look
like,
and
if
we
can
understand
how
all
these
components
work,
we
understand
how
the
neural
cortex
works.
So
we
have
a
theory
about
how
this
all
works.
It's
a
very
high
level
theory.
It
starts.
It's
called
hierarchical
temporal
memory,
hierarchical
temporal
memory,
htm,
it's
some
very
basic
premises
to
it.
First
premise
is
that
it's
the
neocortex
is
a
hierarchy
of
identical
reasons.
That's
pretty
much
a
fact.
A
Okay,
the
next
one
is
each
region
learns
sequences.
Now
here
is
we
start
we're
adding
something
that
most
people
don't
think
about
that.
The
memory-
and
this
is
all
about
memory-
the
memory
in
each
of
these
regions
is
primarily
a
memory
of
time-based
patterns.
Time
patterns
are
changing
over
time.
It's
like
learning
melodies
and
what
happens
is
if
a
region
can
build
a
predictive
model
of
a
sequence.
It
then
creates
a
more
stable
representation
like
the
name
of
the
melody
in
the
next
region
up
and
then
that
learns
sequences
of
sequences
and
so
on.
A
As
you
go
up
the
hierarchy,
you
see
more
stability.
This
is
observed
in
brains.
Similarly,
you
can
take
a
pattern:
that's
fairly
stable
at
the
top
of
the
hierarchy
and
unfold
it
in
time,
and
you
release
the
faster
and
faster
changing
patterns
at
the
bottom.
That's
what
my
speech
is
right
now,
I'm
taking
high
level
concepts
and
unfolding
the
very
fast
changing
patterns
that
make
sounds
okay,
so
that's
the
basic
of
hdm
theory,
but
now
we
a
lot.
That
brings
a
lot
of
questions
exactly
what
are?
What
is
the
region
doing
exactly?
A
What
are
the
different
layers
doing?
What
do
the
cells
do
this
and
how
they
implement
this
how's,
the
memory
actually
work,
and
so
on.
So
let's
dip
into
that
a
little
bit
deeper,
deeper
and
give
you
a
flavor
where
we
are
in
understanding
this?
Okay.
So
here's
a
picture
of
the
slice
of
neocortex
of
our
four
layers
of
cells,
layers,
two
three,
four:
five
and
six.
You
can
see
some
of
the
mini
columns
over
here
on
the
left.
This
is
obviously
a
character
drawing.
This
is
not
a
real
photograph.
A
Now
two
of
these
layers,
the
upper
ones.
Two
three
and
four
are
essentially
feed
forward
layers,
information
going
up
the
hierarchy
and
layers.
Five
and
six
are
essentially
layers
of
information
flowing
down
the
hearkey.
So
when
information
and-
and
what
we
believe
is
going
on
here-
is
that
each
one
of
these
layers
is
implementing
a
type
of
sequence
memory,
each
one
of
them.
So
this
is
this
is
a
very
interesting
idea.
It's
basically
it's
the
same
basic
idea,
repeated
in
different
areas
to
do
different
things.
A
Now,
when
information
comes
in
typically,
this
is
classic
neuroscience.
It
comes
in
and
it
goes
to
first
to
layer
four
and
we
think
of
this
as
sensory
data
like
information
from
your
eyes,
your
ears,
your
skin
things
like
that.
What
most
people
don't
know
or
don't
remember,
is
that
not
only
do
we
get
sensor
data,
but
we
also
get
a
copy
of
your
own
motor
commands.
So,
as
your
body
is
moving,
it
generates
these
neural
firings
that
make
your
muscles
move,
the
literally
those
cells
split
into
and
send
the
same
copy
of.
A
What's
actually
going
to
your
muscles
is
basically
sent
to
the
cortex,
so
the
cortex
gets
to
see
not
only
what
you're
sensing,
but
also
what
behaviors
you're
actually
implementing
right.
Now
we
believe
layer
four.
Is
it
doing
a
type
of
inference
or
you
could
think
of
it
as
pattern
recognition?
But
in
france,
it's
doing
sensory
motor
inference.
It's
trying
to
build
a
predictive
model
of
what's
going
to
come
in
based
on
the
behaviors.
A
You
just
executed
here's
a
picture
on
the
right,
the
simplest
way,
understanding
this
is
think
about
like
a
when
you
look
at
a
face
or
get
an
image,
your
eyes
are
constantly
moving
several
times
a
second
three
to
five
times
a
second,
your
eyes
move
over
different
parts,
and
yet
your
perception
of
the
image
is
stable.
It's
not
moving
around
you,
don't
you're,
not
even
aware
that
your
eyes
are
moving,
and
so
how
is
it
you
create
this
stable
representation?
How
is
it
that
that
you
understand
this
pattern,
while
your
eyes
are
moving?
That's!
A
What's
going
on
layer,
four
now
notice
the
order
in
which
you
move
your
eyes
over
when
you're
looking
at
something
is
not
the
same.
Every
time
it's
not
like
a
repeatable
pattern.
It's
not
like
a
melody.
It
changes
all
the
time.
So
the
only
way
you
can
actually
make
a
prediction
about
what
you're
going
to
see
next,
what
the
brain
is
doing
is
to
know
exactly
what
you're
seeing
now
and
what
motor
behavior
you're
about
to
do.
A
So,
if
I
say
oh
I'm
up
here
and
I
move
down
to
the
down
here,
I'm
going
to
see
a
nose
that
kind
of
thing
that's
what
layer
four
is
doing
anything
that
layer
four
can't
handle
is
passed
on
to
layer,
two
three
and
layer,
two
three
there's
another
type
of
inference,
and
this
is
what
we
call
high
order
inference.
This
is
really
just
like
a
like
a
melody.
A
It's
a
patterns
that
actually
do
repeat
over
time
and
that
you
can
understand
because
they
they
do
repeat
over
time,
and
so
you
can
make
a
prediction
what
note's
going
to
occur
next
in
a
melody,
if
you've
heard
the
previous
some
number
of
notes-
and
you
say-
oh,
I
recognize
this
melody,
and
this
is
where
I
am
I'm
going
to
predict
the
next
note.
My
speech
is
another
example
of
a
high
order
pattern.
A
Now,
layer,
3,
then
becomes
output
and
then
projects
up
to
the
next
higher
region
in
the
cortex
layer,
5
is,
is
where
motor
behavior
is
generated
in
the
cortex.
So
these
cells
there,
they
project
someplace
else
below
the
brain
below
the
excuse
me
below
the
cortex
to
sub-cortical
motor
centers
and
they
generate
behavior
layer.
Five
cells
are
making
my
my
sounds
coming
out
of
my
vocal
tract
right
now
and
and
layer.
Six
is
a
is
an
attentional
mechanism,
this
projects
back
down
to
hierarchy.
A
Now
before
I
go
on,
I
want
to
point
out
well,
I
have
here
in
this
text
here
each
again.
The
point
of
the
slide
is
each
layer
is
doing
a
variation
of
sequence.
Memory,
you
can
think
about
motor
behaviors,
the
type
of
sequences
I'm
playing
back,
and
then
we
have
two
different
types
of
inference
and
there's
a
common
algorithm
going
on
here,
and
the
really
important
thing
here
is
that
these
are
universal
functions.
Everything
I
told
you
about
here
in
a
cortical
region
applies
to
any
kind
of
modality
any
kind
of
sensory
metabolism.
A
It's
true
for
vision,
it's
clear
for
hearing
it's
true
for
touch.
It's
true
for
language.
It's
true
for
science,
everything,
the
brain
does.
The
cortex
does
can
fit
into
this.
These
are
very
powerful
universal
ideas
and
that's
what
we
want
to
find
in
the
cortex,
because
it's
a
universal
learning,
algorithm.
A
Okay,
let's
just
talk
a
little
bit
about.
How
does
this
sequence
memory
work?
And
you
know
I
say
the
sequence
remember
going
on
in
these
layers?
Well,
exactly
how
does
that
work?
We
have
a
really
really
good
idea
how
this
is
going
on
we're
pretty
certain
we
have
this
figured
out.
I
won't
go
into
great
detail
here,
but
I'll
give
you
some
of
the
flavors
for
it.
We
call
this
htm
temporal
memory,
it's
a
pretty
plain
term.
A
Just
so,
it's
a
memory
of
time-based
sequences
and
here's
a
picture
of
one
of
our
our
simulations.
Those
little
cubes
are
represent.
The
neurons,
we're
modeling
and
the
colors
represent
neurons
that
are
active
and
so
we're
actually
modeling.
In
this
case
one
small
layer
of
cells
we're
not
modeling
the
whole
quadratics,
just
one
layer
of
cells
at
a
time
here,
because
this
is
the
sequence
memory
now.
A
Let
me
just
tell
you
the
attributes
of
this
and
we
won't
tell
you
how
to
det
how
it
works,
although
everything
that
I'm
describing
here
is
documented
online
and
the
code
is
available
online,
and
so
this
is
not
a
mystery.
You
can
read
about
it,
but
what
does
this
thing
does?
Well,
they
learn
sequences.
A
It
recognizes
and
recalls
sequences
as
new
patterns
come
in.
It
says:
hey.
Is
this
something
I've
seen
before
and
it's
constantly
predicting
in
fact
it's
making
multiple
predictions
at
the
same
time.
So
as
any
input
comes
and
says,
you
know,
these
are
all
the
things
that
might
occur
next
now
it
does
all
three
of
these
things
simultaneously,
not
one
after
the
other
constantly
saying,
as
the
data
comes
in
it's
saying,
I'm
learning
I'm
recognizing
I'm
predicting
I'm
learning
recognizing,
particularly
over
and
over
again
we've
built
a
lot
of
these
we've
tested
it.
A
We
understand
it
mathematically.
This
is
a
well
characterized
system
and
I
can
just
give
you
some
of
its
attributes.
It's
extremely
high
capacity,
so
even
a
small
section
of
cells-
maybe
you
know,
100
000
neurons-
can
learn
many
millions
of
transitions
over
time.
It
is
a
distributed
system
and
it
uses
local
learning
rules.
So
there's
nobody
in
charge
of
the
whole
thing
and
one
of
the
advantages
of
this.
It
makes
the
system
extremely
fault
tolerant.
A
You
can
lose
neurons
and
synapses
and
columns
it's
just
it's
amazing
the
robot
bust
it
degrades
gracefully
in
almost
all
situations.
There
are
no
sensitive
parameters,
it's
not
hard
to
get
this
stuff
work
and
it
actually
generalizes.
It
can
take
new
sequences
and
new
patterns
and
say
these
are
similar
to
ones
I've
seen
before
semantically
and
make
predictions
about
new
things
in
case
you're
wondering
this
is
not
a
typical,
artificial,
neural
network.
You
know
there's
a
lot
of
talk
these
days
about
deep
learning
and
other
artificial
neural
networks.
A
This
is
very,
very
different
from
that,
and
just
give
you
three
attributes
just
to
give
you
a
clue
as
to
how
it's
different
than
other
type
of
artificial
neural
networks.
One
is.
It
pays
a
lot
of
adherence
to
cortical
anatomy.
We
do
this
because
we
need
to
not
because
we
just
want
to
so.
We
have
this
idea
of
many
columns
and
inhibitory
cells
and
various
connectivity
patterns
that
you
see
in
real
brains
that
almost
nobody
else
models.
A
The
second
thing-
and
this
is
worth
an
entire
talk
on
itself-
which
you
can
see
online-
it's
built
on
a
type
of
representation,
called
sparse,
distributed
representations
and
just
to
give
you
what
that
is,
if
you
think
about
the
neurons
as
being
one
being
a
bit,
and
you
can
say
when
that
renurin
is
active,
it's
a
one
and
a
neuron's
inactive
to
zero.
A
At
any
point
in
time,
most
of
the
neurons
are
inactive
and
only
a
small,
very
small
percentage
in
their
runs
are
active,
and
so
we
can
represent
that
by
a
whole
like
a
vector
of
a
whole
series
of
mostly
zeros
and
few
ones.
It
turns
out
that
this
type
of
representation,
which
encodes
semantic
meaning,
has
some
amazing
mathematical
properties,
which
makes
the
whole
thing
work
so
well.
A
I
won't
go
into
in
further
other
than
say
that
our
sparse
distributed
representations
are
essential
for
how
the
brain
works
and
they're
going
to
be
essential
for
any
machine
intelligence.
That's
a
given.
Finally,
the
neurons
we
model
here
are
unlike
any
kind
of
simple
neurons.
You
see
in
typical
artificial
neural
networks.
They
have
active
dendrites,
they
have
thousands
of
synapses.
We
have
we
learn
by
synaptic
formation,
we're
really
getting
closer
to
what
real
neurons
are
doing,
and
we
understand
why
it
has
these
features
and
they're
essential.
A
Okay,
I'm
going
to
leave
this
for
now
and
if
you
want
to
learn
more,
you
can
go
to
our
website.
Numenta.Com,
learn,
slash
and
there's
videos
and
papers
and
details
about
everything
here.
Okay,
let
me
just
talk
about
a
research
roadmap.
Well,
how
much
of
this
do
we
understand
and
where
are
we
and
what
are
we
doing
with
it?
So,
let's
go
back
to
our
picture
of
a
slice
of
near
cortex.
A
What
we're
trying
to
do
is
work
our
way
through
this
and
understand
exactly
what's
going
on
in
each
of
these
regions,
these
layers
and
how
they
interact
and
so
on.
We
started
with
layer
two
three,
because
it's
the
easiest
one,
it's
the
simplest
one
architecturally,
it's
the
simplest
one
to
understand,
and
so
we
started
with
that.
I
would
say
we
understand
it
very
well.
I
put
the
theory.
Is
at
98:
there's
always
some
things
you
don't
you
know
you
could
get
wrong,
but
we've
extensively
tested
this.
A
We
use
it
in
commercial
code,
we
know
it
works
very
well,
we
understand
it.
We
have
started
working
on
the
sensory
motor
inference
layer.
This
is
again
understanding
how
to
make
predictions
based
on
your
own
movements,
and
I
would
say
the
theory
is
about
80-
done-
we're
well
over
the
hump
on
this
we're
in
development
right
of
it
right
now.
So
this
is
what
we're
working
on
right
now
at
momentum
when
it
comes
to
motor
sequences.
This
is
how
the
cortex
generates
goal
oriented
behavior.
A
We
haven't
started
working
on
that
in
detail
yet,
but
we
have
a
good
piece
chunk
of
the
theory
done
I'd
say
about
half
of
it.
So
we're
to
my
mind.
We
have
a
lot
of
big
big
pieces
in
place
and
it's
a
matter
of
stitching
them
together
and
then
testing
those
ideas,
and
we
haven't
done
as
much
on
the
layer
six,
which
is
the
tension,
feedback
layer,
it's
a
more
complex
layer.
A
So
that's
that's
a
little
bit
further
behind,
since
we
started
here
with
layer,
two
three,
we
said:
okay.
If
this
really
works,
let's
apply
to
real
world
problems
and
see
if
it
works
on
real
world
problems.
So
that's
what
we
did
we
said.
Well,
what
can
you
do
with
this
kind
of
high
order
inference
memory
using
this
htm
temporal
memory?
Well,
then,
upper
right
here
you
can
see.
Well,
we
can
work
on
streaming
data.
A
We
can
work
on
problems
where
the
data
itself
is
changing
over
time,
things
that
are
coming
in
in
a
streaming
format,
and
we
can
do
prediction.
We
can
do
anomaly
detection.
We
do
go
into
classifications,
there's
lots
of
applications
here
in
predictive
maintenance,
security,
natural
language
processing,
I'm
going
to
show
you
some
of
those
in
a
moment.
A
So
how
does
it
do
so?
We
we
actually
built
this,
and
how
do
you
actually
go
about
building
a
system?
What
does
it
really
work
like?
So
here's
how
you
do
it?
Here's,
how
you
build
a
streaming
data
application
using
htm
theory.
You
have
a
data
stream,
you
run
it
through
something
called
an
encoder.
Now
your
data
stream
is
like
numbers
or
categories
or
data
from
a
database,
something
like
that.
You
run
it
through
an
encoder
which
turns
the
data
into
a
sparse
distributed
representation.
A
That's
like
like
a
it's
like
a
sensory
organ
and
then
we
feed
it
into
the
htm
temple
memory
and
we
get
out
predictions
anomalies
and
classifications.
Now
there
are
lots
of
streaming
data
sources
and
it's
not
hard
to
find
them
every
application,
every
server,
biometrics,
medical
devices,
vehicles,
industrial
equipment,
communication
networks,
etc.
All
these
are
spewing
out
data.
In
fact,
you
know
when
people
talk
about
big
data,
mostly
they're
talking
about
things
like
this
things,
that
you
can
read
off
every
few
minutes,
you're
getting
data
points.
A
A
A
Metrics
human
metrics
financial
data
and
then
some
other
ones
with
like
medical
stuff
and
gps
and
natural
language
I'll
give
you
a
flavor
for
each
of
these,
I'm
going
to
start
with
the
first
one
here:
server
metric
data
we
actually
built
a
product
called
grot
and
it's
it's
available
on
the
amazon,
aws
marketplace,
and
it's
for
monitoring
servers.
So
here's
how
that
work,
you've
got
some
servers
running
on
a
server
farm
and
we
take
metrics
from
those
servers.
Now
you
can
take
multiple
metrics.
We
do
this
automatically
for
you.
A
We
run
each
of
those
through
an
encoder
and
we
build
a
model
for
each
metric.
We
experimented
with
combining
the
metrics
before
putting
them
into
the
model,
but
this
actually
works
better.
So
we
just
built
a
model
for
each
metric
using
the
hgm
and
what
you
get
out
of
it.
The
end
you
get
something
called
an
anomaly
scores.
How
unusual
is
this
data
stream
based
on
previous
history,
based
on
what
it's
been
like
in
the
past?
How
unusual
is
it
going
forward?
A
We
display
that
in
a
dashboard,
so
here's
an
example
of
a
dashboard
where
on
a
mobile
device,
where
you're
showing
over
time
how
anomalous
a
particular
server
is,
and
the
height
of
the
bar
and
the
color
there
tells
you
how
anomalous
it
is,
and
we
order
this
and
sort
it
by
order.
So
you
might
be
monitoring
hundreds
or
thousands
of
servers,
but
we
only
show
you
the
ones
at
the
top
of
the
list.
That
are
the
most
unusual
at
any
point
in
time.
A
This
is
continuously
updated
because
the
data
is
continuously
coming
in
the
system
is
continuously
limiting.
We
also
have
a
web
interface,
but
I'll
just
stick
to
the
to
the
mobile
one
for
now,
so
here's
some
some
kind
of
anomalies
at
the
system
attack
and
I'm
going
to
just
start
with
the
easier
ones
and
go
to
the
more
complicated
ones.
A
I
just
need
to
describe
the
screen
to
you
a
little
bit
here
and
what
you're
seeing
is
at
the
top
you're
seeing
the
in
the
white
area
you're,
seeing
the
anomaly
score
for
the
entire
server
over
time
at
the
bottom,
in
the
gray
you're,
seeing
the
anomaly
score
for
a
particular
metric,
maybe
cpu
utilization
or
networking,
something
like
that
and
then
the
black
graph
in
the
middle,
with
the
blue
line.
That
is
the
actual
metric
data.
So
when
you
get
an
anomaly
you
can
go
and
look
and
see
what's
going
on.
A
These
are
the
screenshots
to
show
you
that
so
now,
on
the
left.
Here
you
can
look
to
look
at
the
data
and
you
can
see
the
detected
anomaly.
You
can
see
the
red
tall
bars
there.
This
is
a
pretty
simple
one.
The
data
was
going
along
at
one
level
and
also
jumped
up
to
another
level.
This-
and
you
said,
okay,
it's
a
sudden
change.
It
detected
that
the
next
one
over
is
one
where
there's
more
of
a
gradual
change,
basically
the
same
idea,
but
it
says
you
know
what
that's
too
much
of
a
change.
A
I'm
going
to
detect
the
normal.
The
third
one
over
is
one
where
we
have
a
very
regular
data
stream.
It's
it
has
a
very
repeatable
pattern
and
when
it's
very
repeatable,
the
can
say
even
the
slightest
difference
is
statistically,
very,
very
rare
or
unusual
and
detects
an
army.
So
in
this
case
every
hour
has
this
little
double
double
little
spike
in
activity
and
one
time
one
of
those
spikes
is
a
little
bit
different.
You
can
see
it
on
the
right
there
and
that
and
grock
and
hdm
gets
it
right
away.
A
The
third,
the
last
one
excuse
me
the
fourth
one
on
the
right.
There
is
where
you
have
a
very
unpredictable
data
stream.
It's
noisy
it's
spiking
all
over
the
place.
You
can't
predict
it
completely
accurately,
but
there
is
a
level
of
predictability
and
if
that
level,
predictability
change
the
html
says
it's
still
very
unusual.
So
these
would
not
work
with
thresholds.
You
can't
just
put
some
kind
of
threshold
on
these
things
and
detect
these
here's
where
it
gets
really
interesting,
and
we
love
we've
seen
a
lot
of
these
kind
of
anomalies.
A
A
And
if
you
look
at
the
data,
the
blue
graph,
you
can't
see
why
there's
an
anomaly
there.
It's
not
obvious
to
people
to
say
well,
I
would
have
picked
that
as
an
anomalous
point
in
the
data,
but
the
html
said
I
got
it,
I'm
certain
of
it
and
it
doesn't
report
anomalies
very
often,
it's
very
precise.
So
what
happened
there
was.
A
This
is
a
server
that
has
an
automated
build-up
process
and
an
engineer
went
on
one
day
and
at
this
exact
moment
in
time,
started
the
build
process
manually
very
similar
to
what
happens
automatically
but
slightly
different,
and
not
only
did
the
htm
detect
that
it
detected
in
two
different
metrics
simultaneously
independently
and
it's
a
very
subtle
type
of
thing.
So
this
this
could
be
very
useful
for
security
intrusion,
detection
things
like
that.
This
is
the
kind
of
the
power
that
the
htm
has
that
we
don't
know
of
any
other
capability.
A
The
system
has
this
capability.
Okay,
we
took
that
same
basic
idea
and
we
said:
can
we
apply
to
human
matrix?
You
know
people
sitting
at
computers,
typing
away
and
accessing
files.
Can
we
tell
if
something's
unusual,
going
on
on
a
particular
person's
computer,
and
the
answer
is
we?
Can
we
put
these
like
keystrokes
and
file
access
and
we
can
detect
when
people
do
unusual
things?
In
this
case,
someone
created
a
very
large
zip
file.
They
hadn't
done
before.
If
they
did
that,
every
day
it
wouldn't
be
anonymous.
A
We
then
we're
actually
building
a
product
right
now,
that's
built
on
detecting
anomalies
and
company
data
and
we're
looking
at
both
stock
volume,
data
and
social
media
aspects
of
that
company,
and
we
can
actually
look
at
the
stock
volume
of
a
company
and
we
can
detect
unusual
patterns
in
it
and
we
can
do
the
same
thing
with
twitter.
So
that's
something
we're
in
process
of
doing
right.
Now,
then
here's
some
other
applications
are
quite
different.
A
This
is
one
that
was
done
by
a
researcher
at
berkeley
and
they're,
trying
to
take
eeg
readings
off
a
scalp,
and
you
and
decipher
them
do
classification
on
those
edg
readings
to
control
things
like
prosthetic
arms
or
in
this
case
quadcopters-
and
this
is
a
they
got.
Initial
results
were
very
promising.
I
don't
want
to
overplay
this
because
it
was
just
it
hasn't
been
very
far,
but
it's
the
kind
of
problem
that
the
should
be
good
at
you're.
A
Getting
these
complex
temple
data
streams
coming
off
the
the
sensor,
and
we
should
be
able
to
model
that
and
classify
it
and
so
early
results.
Look
good,
here's
another
one
where
this
is
a
company
in
europe
and
amsterdam
that
is,
building
an
application
to
detect
anomalous
behaviors
in
ships
moving
through
harbors,
so
ships
come
in
and
out
of
harbors
all
the
time.
The
idea
is
here:
we
don't
know
what
it's
supposed
to
be
like,
but
they
just
the
html
models,
these
behaviors
using
gps
data
and
it
detects
one.
A
A
ship
is
behaving
unusual,
whether
it's
going
too
fast
or
too
slow
or
turning
in
the
wrong
place,
or
something
like
that.
Very
good
for
security
applications
and
then
I'll
finally
end
up
on
natural
language.
This
is
work
that
we
did
in
conjunction
with
a
company
called
cortical.
I
o
they're
in
austria
and
cortical
io
said
hey.
They
like
this
theory
that
we're
working
on
and
they
they
develop
a
tool
which
is
really
cool.
A
It's
it
can
take
a
word
and
create
an
sdr
or
sparse
distributed
representation
from
it
and
that's
what
these
pictures
on
the
right
are
representing.
These
are
those
are
sdrs
of
words
and
those
dots
are
the
one
bits
and
the
white
areas
the
zero
bits.
Now
they
did
this
by
feeding
it's
a
complex
technique.
A
They
did
it,
but
they
basically
train
the
system
on
like
a
corpus
of
text
like
wikipedia
and
once
they've
done,
that
you
can
give
it
any
word,
and
it
gives
you
back
an
sdr
and
it
has
all
the
right
properties
that
sdrs
are
supposed
to
have,
which
I
didn't
get
into
in
this
talk.
But
what
you
can
do
is
something
really
interesting.
You
can
take
the
sdr
for
a
word
like
apple
and
the
sdr,
for
a
word
like
fruit
and
now
apple
could
mean
multiple
things.
A
This
is
the
power
of
these,
these
representations
of
the
brain
jews
and
when
the
answer
you
get
is,
if
you
ask
what
is
apple
minus
fruit,
you
get
a
new
sdr.
The
most
closest
pattern
is
to
computer,
and
so-
and
it
says,
okay,
that's
what
that
is,
and
the
next
one's
down
the
list
would
be
macintosh,
microsoft,
max
lens
and
so
on.
A
We
can
then
take
these
sdrs
and
feed
them
into
the
htm,
a
sequence
of
words,
just
like
a
sequence
of
data
coming
off
a
server,
a
sequence
of
words
going
into
the
hdm,
and
so
we
trained
a
system
on
this.
We
started
there's
a
very
first
simple
test.
We
trained
on
three
word
sentences
that
it
was
an
animal,
either
eats
or
likes
something.
So
elephant
likes
water
and
elephant
eats
grass
and
we
didn't
train
on
a
lot
of
sentences.
It's
just
a
50
or
60.
A
Something
like
that,
and
so
what
what
we
can
then
do
is
we
can
say:
okay
after
we
trained
it
on
this
series
of
words,
the
sentence
is
we
can
ask
it
a
question,
we
can
say:
well
what
does
the
fox
eat
and
the
way
we
do
that
is
we
feed
in
the
word
fox,
the
word
eats
and
hcm
makes
a
prediction
now.
The
word
fox
had
never
been
seen
before
by
the
system.
A
It's
it's
something
that
we
have
an
sdr
for
it,
but
the
html
was
never
trained
on
it
that
that
pattern,
and
but
the
word
fox,
would
share
semantic
similarity
to
other
animals,
and
so
we
can
then
basically
feed
in
a
new
animal
and
say
what
do
we
think
the
fox
eats
and
the
answer
you
get
out.
You
get
an
sdr,
you
look
it
up.
The
ansi
gap
is
rodent,
and
this
is
really
incredible.
It's
it's
showing
a
level
of
generalization.
A
That's
pretty
cool
and
it's
working
on
the
same
principles
that
your
brain
works
on.
This
is
a
very
simple
example,
but
it's
the
same
basic
principles
that
are
going
on
in
your
brain
when
you're
understanding
language,
we
think
there's
a
lot
of
applications
for
this.
Of
course,
the
whole
thing
is
unsupervised,
it
does
semantic
generalization,
it
can
work
across
multiple
languages
and
we
think
there
are
applications
in
search,
sentiment,
analysis
and
so
on.
So
this
is
a
whole
area
of
research.
That's
going
to
be
built
around
htms.
Okay.
Now
I
just
showed
you
quickly.
A
Six
of
these
applications-
and
the
point
I
want
to
make
about
them
here-
is
that
every
one
of
these
was
was
built
using
the
exact
same
code
base,
not
similar,
not
a
reconfiguration,
not
a
recompilation,
the
exact
same
code
base.
We
literally
took
data
from
these
different
sources,
fed
them
through
an
encoder,
turned
them
into
sdr
and
fed
them
within
the
exact
same
htm
model,
and
of
course
it
learned
different
things
in
each
case
and
and
we
got
useful
results
in
all
of
them.
This
is
really
getting
to
the
universal
nature
of
cortical
algorithms.
A
There's
no
other
machine
learning
technology
that
can
could
do
this,
but
brains
in
nature
figured
out
how
to
do
this.
Okay,
let's
go
back
to
our
research
roadmap.
We
start
with.
I
just
talked
about
how
we
modeled
layer,
three,
which
is
this
high
order,
inference
and
the
applications
we
can
build
with
that.
A
We're
now
working
on
the
sensory
motor
inferences
again
trying
to
model
how
the
brain
understands
the
world
when
you
move
in
the
world-
and
we
are
now-
and
we
can
ask
ourselves
we're
in
the
process
of
doing
that,
but
we
can
answer
those.
What
kind
of
applications
could
we
build
using
that
technique
and
there's
lots
of
them?
So
essentially,
this
now
works
on
static
data.
A
Instead
of
streaming
data,
it
would
be
static
data
and
yet
we're
moving
through
the
data
we
actually
move
like
it's
like
you're,
looking
at
a
picture
and
you're
moving
your
eyes
over
it.
Now
it's
static
data,
but
it's
with
active
learning
and
you
can
do
classification
and
you
can
do
prediction
we're
working
on
vision
right
now,
because
it's
a
classic
problem,
but
I
actually
think
the
more
interesting
applications
will
be
non-vision.
There
will
be
things
like
network
classification
or
anywhere
you're,
trying
to
understand
and
and
recognize
or
or
predict
complex,
connected
graph
data.
A
I
think
it's
going
to
be
really
cool
out
here.
Eventually,
we're
going
to
go
on
to
not
too
far
in
the
future,
we're
going
to
go
on
to
working
on
layer,
5
and
motor
sequences,
and
now
this
is
where
it
really
gets
interesting,
because
now
you
can
start
adding
a
goal
or
any
behavior.
So
the
system
not
only
can
understand
what
happens
when
you
behave
but
actually
starts
directing
behavior,
which
is
what
we
think
about
when
we
think
about
human
level.
A
Behavior
and,
of
course,
that
introduces
the
idea
of
robotics
and
and
other
things,
but
I
don't
think
most
of
the
applications
are
going
to
be
in
physical
robotics.
I
think
they're
going
to
be
in
sort
of
more
virtual
worlds,
where
we
have
agents
that
are
moving
through
data
in
an
intelligent
way
or
trying
to
do
proactive,
defense
and
so
on.
The
same
principles
can
be
applied
to
lots
of
different
things:
okay
and
then
finally,
layer
six
allows
you
to
build
bigger
hierarchies
in
multi-century
modalities.
A
That's
when
we
really
want
to
scale
this
up,
just
a
word
about
how
we
go
about
our
research,
we're
very
transparent
everything
we
do
is
documented
and
available
to
inspection
and
open
source.
So
all
our
algorithms
are
documented.
The
documentation
could
be
better,
but
it's
sufficient
because
we
have
many
independent
implementations
around
the
world,
so
other
people
have
built
this
off
and
it
works
for
them
too.
We
have
an
open
source
project
called
new
pic.
A
You
can
find
it
at
nomento.org
and
all
of
our
software
is
open
source
there
and
we
actually
put
even
our
daily
research
code
if
we're
working
on
things
and
experimenting
and
building
hacky
code
as
we
go
along,
we
put
that
up
there
as
well.
There's
active
discussion
groups
for
theories
and
implementations,
and
so
on.
A
We
also
have
collaborations
we
have
a
collaboration
with
ibm's
research
group
in
amaden,
san
jose,
california,
and-
and
we
also,
we
have
a
cooperating
with
darpa
in
washington,
d.c
who's,
trying
to
do
a
project
on
building
hardware
for
cortical,
html
algorithms,
and
then
little
companies
like
cortical.
I
o,
which
I
mentioned
with
the
words
this
is
just
a
a
chart
about
some
of
the
metrics
at
our
open
source
community.
We
created
this
about
15
months
ago
and
it's
been
growing
steadily
since
so
we're
very
pleased
about
this.
A
It's
a
it's
a
good
active
community
and
I
would
encourage
anyone
to
join
in
at
whatever
level
you
can.
I'm
going
to
end
with
this
slide
here.
This
is
sort
of
my
my
the
way.
I
view
the
machine
intelligence
landscape.
What's
going
on
right
now,
and
there
are.
This
is
a
rather
simplistic
perspective,
but
I'll
offer
it.
Anyway.
I
see
there's
three
basic
ways:
people
think
about
machine
intelligence
on
the
left.
A
You
have
people
who
are
interested
in
brains,
cortical
modeling,
that's
what
we
do
and
I
would
argue
that
the
furthest
advanced
technology
in
this
area
is
htms.
So
that's
a
good
example
there,
the
cortical
modeling,
we
say
we're
going
to
understand
how
the
brain
works.
Then
we
have
in
the
middle.
We
have
artificial
neural
networks,
part
of
the
machine
learning
community
and
this
the
premise
there
is
that
it's
really
mathematically
based
it's
not
biological,
based
at
all.
A
In
fact,
artificial
neural
networks
have
almost
nothing
to
do
with
neurons
and
and
brains,
or
anything
like
that,
but
they're
they're,
mathematically,
derived
features
and
and
systems
and
the
the
prominent
one
these
days
is
deep
learning.
So
you
hear
that
in
the
news
a
lot
deep
learning
is
just
a
type
of
artificial
neural
network,
very
mathematically,
oriented
and
so
on.
Then
we
have
the
more
classic
ai.
Where
these
are
this
good
example,
this
would
be
watson
which
is
ibm's
a
recent,
a
really
impressive
machine.
A
So
again,
the
premise
under
these
is
different
for
each
one
in
the
cortical
modeling,
it's
biological
in
the
artificial
neural
networks,
it's
mathematical
and
I
would
say
in
the
ai
world,
it's
sort
of
an
engineered
solution
like
hey,
whatever
works,
to
get
this
to
work,
we're
going
to
do
it
they're
all
valuable,
I'm
not
trying
to
put
value
judgment
on
them.
They
do
different
things.
They
work
on
different
types
of
data,
so
the
htms
we
work
on
spatial
temple
data.
A
That's
I
already
showed
that
here
we
showed
we
can
actually
do
some
language
stuff.
We're
gonna
be
doing
more
than
that
and
we're
starting
to
integrate
behavioral
data,
meaning
the
data
from
your
own
body,
artificial
neural
networks
or
deep
learning
is
primarily
spatial
data.
That's
what
they
do
today,
but
they're
working
on
adding
temporal.
So
I
put
that
in
gray
there,
because
that's
an
extra,
that's
a
dimension,
they're
working
on
and
then
watson
is
really
all
about
language
and
documents.
A
That's
the
that's
the
data
type
it
works
with.
They
do
slightly
different
things.
Overlapping
things.
Htms
are
really
good
at
classification
and
prediction
and
we're
starting
there.
We
have
a
path
to
go
on
behavior
deep
learning,
networks
are
primarily
classification,
networks
and
watson
is
sort
of
a
natural
language
query
type
of
system.
A
I'm
going
to
start
on
the
right
here,
I'm
going
to
say
for
watson
and
ai,
probably
not
now
before
anyone
gets
mad
at
me.
This
is
the
actual
self-assessment
of
the
people
who
built
watson.
They
were
asked
this
question.
Do
you
think
this
is
on
the
path,
machine,
intelligence
and
they
said
no
they'd
have
to
do
a
lot
of
things
to
get
that
to
sort
of
be
a
more
general
purpose,
something
beyond
natural
language.
It
can
be
very,
very
useful
for
what
it
does,
but
is
it
on
a
path
to
building
truly
machine
intelligence?
A
I
probably
not
on
the
deep
learning
side,
I
would
say:
probably
not
there
as
well,
they
have
to.
They
have
to
start
incorporating
time.
They
have
to
have
a
understanding
how
you
can
implement
behavior
attention,
etc.
So
they
have
a
lot
of
things
that
they're
not
doing
today,
but
they
could
get
there
if
they
really
start
taking
those
other
features
seriously,
and
this
is
not
like
a
you
know.
This
is
not
it
right
or
wrong.
It's
just
like.
What
do
you
need
to
do
to
get
the
machine
intelligence?
A
I
would
argue
that
the
htms
are
on
the
path
machine
intelligence.
I've
shown
you
an
outline
of
our
research
agenda,
how
it
fits
with
cortical
algorithms,
the
process
that
you're
going
to
do
from
pure
sensory
to
sensory,
motor
and
motor
behavior
and
so
on,
and
we
have.
We
have
a
roadmap
to
get
there.
A
I
showed
a
road
map
to
get
the
machine
intelligence
here,
and
so
I
believe
that
that
is
going
to
be
the
fastest
path
by
far
to
get
there
and
that's
the
end
of
my
talk,
it's
an
exciting
time
to
be
working
in
this
field.
It
is
it
is
that
crazy
time
when
there's
lots
of
ideas-
and
it's
not
very
clear
to
everybody,
just
like
the
1940s
worth
of
computing,
but
it's
going
to
be
only
a
few
years
now
before
this
settles
out.