►
Description
"Principles of Hierarchical Temporal Memory (HTM): Foundations of Machine Intelligence"
The Q & A Session that followed this presentation can be found here: https://youtu.be/EU2Vm-VlfEk
Jeff Hawkins, Co-Founder, Numenta
Numenta Workshop Oct 2014 Redwood City CA
A
A
A
The
second
thing
is,
we
want
to
take
that
knowledge
and
turn
it
into
technology
that
can
be
applied,
and
this
is
the
practical
side
of
this.
So
it's
sort
of
a
science
side
and
a
practical
side.
Now
it's
it's
surprising
that,
but
not
everyone
thinks
this
is
a
good
idea,
and
I
mean
that
not
everyone
thinks
this
is
the
right
way
to
go
to
build
intelligent
machines
or
build
intelligent
technology.
A
So
I
want
to
just
give
you
a
little
bit
of
motivation
why
we
think
understanding
how
the
cortex
works
is
really
important
and
by
the
way
I
should
just
point
out.
The
cortex,
if
you're
not
familiar
with,
is
about
75
of
the
volume
of
your
brain.
It's
where
all
high
level
thought
occurs
and
when
I
talk
about
the
cortex,
I'm
actually
talking
about
a
few
other
structures
with
it,
but
I'm
not
going
to
get
into
that
level
detail
today,
but
we're
talking
about
part
of
the
brain
about
three
quarters
of
it.
A
Okay,
so
let's
talk
about.
Why
should
we
study
brands
to
do
this?
Why
should
we
study
the
neocortex
to
do
this?
Why
would
machine
intelligence
be
based
on
cortical
principles?
A
couple
of
things
you
may
not
be
aware
of
you
might
be.
If
you
follow
this
field
much,
the
cortex
uses
a
very
common
algorithm
for
almost
everything
it
does.
A
When
you
start
thinking
like
okay,
you
know
how
what
how
these
things
are
common,
but
it's
a
it's
a
common
learning
algorithm
and
that
suggests
that
we
can
understand
that
common
learning
algorithm.
We
can
apply
it
to
lots
of
different
things.
Now
it
turns
out
that
our
brains,
human
brains,
are
particularly
good
at
a
lot
of
things,
we're
amazingly
adaptable.
A
We
have
languages,
we
have
science,
we
have
arts
and
engineering
all
the
things
we
do.
These
are
all
a
product
of
the
neocortex
and
we
do
not
have
separate
areas
of
separate
things
to
do
these.
It's
still
the
same
algorithm.
So
these
core
algorithms,
which
were
originally
evolved
to
understanding
low-level
sensory
data
and
building
a
model
of
the
world,
can
be
applied
to
very,
very
deep
problems
that
we
deal
with
what
we
think
about
as
intelligence
in
our
in
our
species.
A
A
Principles
for
machine
intelligence
are
network
effects
things
we've
seen
in
the
past
in
other
areas,
so
people
are
going
to
want
to
work
on
the
most
flexible
solutions
and
we're
going
to
the
most
resources
are
going
to
be
put
into
that,
so
we're
going
to
be
naturally
moving
towards
a
more
universal
solution
and
there's
nothing
more
universal
than
our
brain,
because
we
know
that
nothing
else,
that's
even
close
to
it.
So
these
are
the
sort
of
motivations
we
use
for
studying
the
cortex
and
as
an
example
for
us.
A
Okay,
here's
my
the
agenda
for
my
talk.
I'm
going
to
start
with
some
cortical
facts
things
we
know
about
the
brain,
I'm
going
to
then
go
into
cortical
theory
or
hierarchical
temporal
memory.
The
hdm
theory
I'm
going
to
give
you
a
research
roadmap.
Tell
you
what
we've
done
where
we're
going
what's
next?
What's
after
that,
I'm
going
to
give
you
an
applications
roadmap,
what
kind
of
things
we
can
build
today
in
the
future
and
then
I'll
end
with
a
few
thoughts
on
machine
intelligence.
A
Now,
as
as
craig
mentioned
in
the
introduction,
some
of
this
is
going
to
get
pretty
deep.
So
I
starts
out
easy
and
it
gets
pretty
deep
and
then
it
gets
easy
again,
as
craig
mentioned,
you
don't
really
need
to
worry
about
it.
If
you
don't
need
to
understand
all
the
stuff
that
I'm
going
to
talk
about
it's
and
and
and
you
can
sort
of
just
try
to
if
it
gets
a
little
hard,
you
can
sort
of
pick
up
the
pieces
you
can.
A
So
it's
kind
of
important
to
do
that,
it's
a
little
bit
like.
If
you
want
to
understand
how
a
computer
works.
You
know
if
you
want
to
use
a
computer,
you
don't
really
need
to
know
all
the
details
of
what
you
know,
how
cache
memory
works
and
stack
pointers
and
all
this
kind
of
stuff.
But
someone
had
to
know
that
initially
we're
going
to
get
kind
of
that
kind
of
level.
You
don't
really
need
to
understand
all
this
to
use
this
technology,
but
it's
helpful
to
have
this
as
a
background.
A
Okay,
let's
just
jump
right
into
it.
So,
let's
just
start
just
a
very
high
level
of
what
the
cortex
does
it's
an
organ
of
memory,
it
it
learns
and
it
interfaces
to
the
world
through
a
bunch
of
sensory
organs.
So
we
all
know
about
the
retina
and
the
cochlea
and
the
somatic
senses
there's
quite
a
few
sensory
systems.
You
have
and
the
interesting
thing
about.
A
It,
though,
is
once
you
come
outside
of
that,
once
you
leave
the
retina
or
you
leave
the
cochlea,
it's
just
patterns
of
action,
potentials
or
firings
on
nerve
fibers
and
those
nerve
fibers
are
identical
matter
which
what
they
represent.
So
there's
no
difference
between
a
pattern.
That's
coming
in
from
the
optic
nerve,
there
is
from
the
somatic
sensory
nerves
and
it's
the
brain
does
not
really
deal
with
light
and
sound
and
touch
the
brain,
especially
the
cortex.
A
Basically
deals
with
patterns
and
the
reason
the
world
seems
different,
like
vision,
seems
different
than
hearing
is
because
of
how
the
model
that
the
brain
makes
from
this.
So
the
cortex
is
a
it
takes
in
this
fast
changing
sensory
data.
It's
changing
all
the
time.
Think
about
my
speech.
It's
changing
the
order
of
milliseconds,
it's
flowing
into
your
brain
right
now,
and
it
builds
a
model
of
the
world.
It
builds
a
predictive
model
and
that
predictive
model
is
basically
says.
A
A
Most
of
the
changes
that
are
occurring
on
your
sensory
organs
are
coming
from
your
own
behavior,
not
from
the
world
itself.
For
example,
most
of
the
changes
are
occurring
on
your
eyes
right
now
or
because
you're
moving
your
eyes
and
you're
moving
them
several
times
a
second
you're
not
aware
of
it,
but
it's
constantly
changing.
You
have
a
very
fast
changing
data
stream
coming
because
you
move
your
eyes
as
you
walk
through
a
building.
A
As
you
turn
as
you
touch
things,
it's
all
about
how
you
interact
with
the
world,
so
most
of
the
changes
are
coming
are
from
your
own
sensory
of
your
own
activity,
your
own
behavior.
So
we
say
the
cortex
builds
or
learns
a
sensory
motor
model
of
the
world.
It
learns
how
the
world
behaves
largely
when
we
interact
with
it,
but
also
how
it
behaves
on
its
own.
So
this
is
what
the
goal
of
our
system
here
is
to
build.
A
sensory
motor
model
of
the
world
and
from
that
we
can
generate
behaviors.
A
Okay,
let's
just
jump
into
some
real
cortical
facts.
Here's
a
little
picture
of
a
you
know,
cortex!
Next
to
it
I
show
a
rack
neocortex
just
to
let
you
know
that
I'm
talking
about
the
same
thing.
It
doesn't
really
matter
what
species
we're
talking
about.
If
it's
a
mammal
it
has
the
neocortex
and
the
properties
I'm
going
to
talk
about
now
are
universal
across
species.
A
A
To
give
you
a
sense
for
what
the
this
is
a
good
model
for
human,
near
cortex,
it's
about
the
right
size
and
about
the
right
thickness
and
maybe
somewhere
around
60
billion
neurons
in
the
sheet-
and
this
is
what's
in
your
head
and
it's
what
in
my
head
right
now
and
this
sheet
of
cells
is
listening
in
your
head
and
mind-
is
generating
speech
now
what's
interesting
about
it?
Is
it's
remarkably
uniform?
A
You
can
find
differences
here
and
there,
but
it's
incredibly
uniform
both
anatomically,
so
you
can
look
at
different
species
and
different
areas
in
your
cortex.
It
looks
virtually
the
same
and
it's
functionally
very,
very
uniform,
meaning
you
can
literally,
and
people
have
done
this
experiment,
take
an
optic
nerve
and
an
auditory
nerve
and
switch
them
on
an
animal
and
the
part
of
the
cortex
that
was
auditory
becomes
up.
You
know,
visual
and
the
visual
becomes
auditory.
A
If
you
delve
down
next
level
of
structure,
you'll
see
that
it's
organized
as
a
hierarchy.
Now,
why
is
it
organized
in
a
hierarchy?
Because,
even
though
it's
a
sheet
of
cells,
different
areas
in
the
sheep
connect
to
other
areas?
And
if
you
follow
that
map
you
get
a
hierarchy
and
humans
have
a
very
deep
and
big
hierarchy.
Other
mammals
have
a
smaller
one
if
you
dive
down
next
level,
if
you
look
at
a
slice
through
that
two
and
a
half,
millimeters
and
you'll
find
the
next
structure.
You'll
see
is
layers,
there's
layers
of
cells.
A
How
many
layers
depends
on
who's,
counting
but
they're.
Basically,
four
layers
of
cells-
two
three
is
one
believe
it
or
not,
and
then
there's
four
five
and
six
four
layers
of
cells.
If
you
dive
down
further
still
you'll
see
that
there's
neurons
in
there
and
the
neurons
have
an
organizational
property
in
these
something
called
mini
columns,
they're
organizing
these
very
many
miniature
columns.
Many
columns
exist,
there's
a
debate
within
the
neuroscience
world,
whether
they're
functionally
relevant.
A
Now
these
neurons
have
thousands
of
synapses
on
them
anywhere
between
three
and
ten
thousand
synthetics
on
each
synapses
connections
on
each
cell
and
what's
interesting
about
it,
is
that
only
a
small
percentage
of
about
ten
percent
are
close
to
the
cell
body.
This
is
what
most
people
think
about
when
they
think
about
a
neuron,
most
artificial,
neural
networks.
A
They
think
about
these
synapses
that
get
summed
in
the
cell
body,
but
90
of
these
synapses
are
far
away
and
and
for
many
years
people
couldn't
understand
what
they're
there
for,
because,
if
you
activate
one
of
those
distal
synapses,
it
seems
to
have
no
effect
at
all.
So
people
say
what
is
these
thousands
of
synapses
doing
out
here?
We
now
know:
what's
going
on,
you
can
jump
down
further
and
we
now
know
this
is
something
maybe
in
the
last
15
years
or
so.
A
It's
become
very
clear
that
if
you
look
at
one
of
these
little
branches
on
these
dendrites
off
near
not
near
the
cell
body,
far
away
from
the
cell
body,
they're
active
processing
elements
and
if
you
have
a
set
of
synapses
that
become
active
relatively
close
period
in
time
and
relatively
close
in
space
to
meaning
they're
near
each
other,
it
can
generate
what
they
call
a
dendritic
action
potential
which
travels
to
the
cell
body
and
depolarizes.
The
cell
has
a
large
effect
on
the
cell.
It
doesn't
make
the
cell
fire,
but
it
depolarizes
it.
A
So
now
we
have
all
these
thousands
of
synapses
out
there
that
are
doing
some
sort
of,
like
coincidence,
detector
and
then
finally,
people
think
that
learning
in
the
brain
we
used
to
think
and
many
people
still
do
it's
all
about
changing
the
synaptic
weight.
That
is
the
strength
of
these
connections.
Well,
that
happens
to
some
extent.
We
now
know
that
synapses
are
formed.
New
synapses
are
formed
all
the
time
and
are
being
lost
all
the
time,
and
this
is
a
much
more
powerful
type
of
learning.
It's
called
synaptogenesis,
and
so
it's
not
like.
A
Oh
I'm,
just
incrementing
a
little
weight
here.
I
can
form
a
completely
new
connection
and
this
is
really
real,
actually,
where
most
learning
occurs
in
the
brain.
Okay.
So
that's
some
cortical
facts
now,
what's
the
theory
behind
this?
Well,
we
have
a
overall
theory
for
this.
We
call
htm
hierarchical
temporal
memory.
It's
it's
pretty
straightforward.
Pretty
simple
essentially
says
we
have
a
hierarchy
of
identical
regions,
meaning
they're
all
doing
something
very
similar.
So
that's
pretty
much
fact.
They
are
learning
something.
So
it's
memory,
and
now
but
here's
the
thing.
A
We
believe
that
all
these
regions
are
primarily
memory
of
of
time-based
patterns,
it's
memory
of
sequences
or
temporal
transitions
and
what
it's
like.
It's
like,
you're
learning
melodies
in
each
one
of
these,
and
what
happens
is
if
you
can
make
proper
predictions
if
it's
a
predictive
model,
predictive
memory,
if
you
can
predict,
what's
going
to
happen
and
you
form
a
stable
representation
for
it
and
you
end
up
with
representations
being
more
stable
as
you
send
the
hierarchy,
it's
like
you're
learning,
names
of
sequences
and
the
names
of
the
sequence
of
the
sequences
and
so
on.
A
And
similarly,
when
you
have
a
high
level,
stable
representation,
it
can
unfold
into
very
long
complex
speeches
sequences
like
my
speech
right,
so
I
have
some
very
high
level
concepts,
I'm
thinking
about
and
then
I'm
just
playing
back
memories
that
I've
recorded
earlier.
I
I've
said
these
words
before,
and
I've
said
these
ideas
before
and
I'm
just
playing
back
recorded
sequences,
and
it
turns
out
to
be
a
very
fast
level
pattern.
Okay,
that's
the
basic
idea
of
htm.
The
question
now
we
want
to
ask
is
well
how
does
this
exactly
work
in
detail?
A
What
do
the
regions
do?
What
are
the
cellular
layers
are
doing?
How
do
the
neurons
work
etc?
And
this
is
what
we
spend
most
of
our
time
studying,
so,
let's
just
jump
into
it.
Further.
Here's
a
slice
of
that
two
and
a
half
millimeters
of
cortex,
and
we
can
see
the
four
layers
there,
there's
roughly
two
feed
forward
layers
and
two
feedback
layers.
A
So
the
two
layers,
two
three
and
four
are
feet
forward,
and
the
five
and
six
are
feedback,
and
what
we
believe
is
each
layer
of
cells
is
implementing
a
type
of
sequence
memory.
Each
one
is
implementing
actually
a
variation
on
the
same
type
of
sequence,
memory.
The
two
feed
forward
layers
are
doing
inference
or
pattern
recognition,
and
I'm
going
to
go
in
details
about
how
those
we
believe
those
work
and
the
other
ones
are
more
of
a
feedback
layer.
Five
is
the
layer,
the
layer
that
has
cells
that
generate
motor
behavior.
A
So
my
speech
is
being
generated
by
cells
in
layer
five
in
parts
of
my
cortex
and
then
layer
six.
Is
it
has
to
do
with
attention
and
hierarchy?
So
those
are
the
basic
idea
there
and
we
think
again.
Every
one
of
these
layers
is
doing
something
similar,
but
it's
doing
it's
a
variation
and
it's
being
applied
to
different
problems
by
what
it's
connected
to.
So
you
can
take
this
sort
of
generic
sequence
memory
and
then
turn
it
to
different
uses,
because
you
know
motor
behavior
is
sequence,
memory
and
inferences
sequence,
memory
and
so
on.
A
So
let's
just
jump
into
layers
four
and
two
three,
the
two
feed
forward
inference
layers,
and
these
are
the
ones
we
understand
the
best,
a
very
classic
neuroscience.
If
you
read
any
papers
in
neuroscience,
you'll
see
this
often
that
the
input
to
a
particular
region
first
arrives
at
layer
four.
This
is
the
the
basic
idea
and
then
it
projects
to
layer,
three
and
then
layer,
three
can
projects
down
to
the
next
layer
up
the
hierarchy.
So
this
is
this.
Is
the
basic
feed
forward
pathway?
A
A
Now
everyone
thinks
about
the
input
coming
into
the
brain
as
being
sensory
data
like
at
the
primary
visual
cortex
or
primary
arthritic
cortex,
it's
information
coming
from
the
eyes
or
the
ears,
but
there's
another
thing:
that's
coming
into
er,
which
people
don't
remember,
or
they
don't
know
which
is
you
get
a
copy
of
motor
commands?
So
the
cortex
is
not
just
sensing
the
world
it
actually
gets
a
copy
of.
Whoever
else
is
making
the
behavior
in
the
body.
A
What
we
think
is
going
on
in
layer
four
is
is
what
we
call
sensory
motor
inference.
We
learned
sensory
motor
sequences.
This
best
example
and
I'll
use.
A
fair
amount
in
this
talk
is
when
you're.
Looking
at
an
image
or
you're
looking
at
me
and
as
your
eyes
are
moving,
you
can
constantly
completely
change
the
input
to
your
to
to
your
brain
completely.
It
doesn't
feel
that
way
the
world
feels
stable
to
you,
but
every
time
your
eye
is
moving,
the
entire
innervation
is
different
and
this
is
not
a
high
order
sequence.
A
This
is
not
a
sequence
that
repeats
itself.
You
can't
predict
just
by
the
order
of
what
patterns
come
in.
What's
going
to
be
next,
however,
if
you
do
have
the
copy
of
the
motor
command,
you
can
do
that
if
you
say
well,
here's
what
I'm
seeing
I'm
about
to
move
over
here.
I
can
predict,
what's
going
to
happen
next
and
we
believe
that's
what's
going
on
in
layer
4..
A
This
is
a
predictive
memory.
If
the
system
can
predict
correctly
what's
going
to
occur
next,
we
want
to
form
a
stable
representation
in
the
next
layer,
in
this
case
layer,
3,
and
if,
if
it
can't
it
says,
look
I'm
not
able
to
model
this
change.
It
passes
those
changes
through
it
says.
Look
I
can't
handle
this.
The
next
guy
is
going
to
get
these
changes.
What
layer
three
does
is
a
it's
a
it's
a
more
of
a
pure
auto-associative
sequence
memory.
A
We
call
it
a
high
order,
sequence
memory
and,
and
a
good
example
of
what
high
order
means
is.
I
just
give
you
two
a
very
simple
example
here.
Imagine
I
have
two
sequences,
a
b
c
and
d
and
x,
b,
c
and
y
and
notice
that
if
I
want
to
show
you
after
I
train
on
those
sequences,
if
I
show
you
abc,
you
should
predict
d
and
if
I
show
you
x
bc,
you
should
predict
y
now
I
can't
just
use
the
previous
state.
I
can't
just
say
well
c
what
should
I
predict?
A
I
can't
tell
you
that
I
have
to
go
back
in
time.
This
makes
it
a
high
order,
sequence
and
most
of
the
world
is
like
that.
Language
is
like
that
when
you
walk
around
a
building
is
like
this
most
of
the
world
manifests
itself
as
high
order
sequences.
So
these
are
the
two
basic
types
of
patterns
you
can.
Even
you
can
see
in
the
world.
These
are
all
these
are
like
universal
instant
steps.
A
If
you
think
about
it
deeply,
there
isn't
much
else
the
brain
can
work
on,
it
can
say:
look
I
can
try
to
make
a
predictive
model
based
on
my
own
behavior
or
I
can
try
to
make
a
predictive
model
based
on
some
sort
of
high
order
sequences
I
can
observe.
If
I
can't
do
that,
then
I
can't
do
it
it's
it's.
Then
it's
like
random.
These
are
universally
applied
to
every
sensory
modality,
they're,
nothing
specific
about
vision,
hearing
or
touch
it's
a
very.
These
are
very
deep
concepts.
A
It
also,
if
you
know
anything
about
neuroscience.
I
won't
give
you
the
evidence
for
this,
but
these
steps
completely
explain
the
type
of
receptor
field
properties
we
see
in
layer,
four
and
layer,
three
and
layer,
two.
These
concepts
lead
to
that.
So
we're
pretty
confident
this
is
what's
going
on
now
we
wanna
jump
in
further.
I
wanna
jump
down
to
like
exactly
what's
going
on
one
of
these
layers
and
then,
when
we
get
to
the
bottom
of
all
that
then
we'll
come
back
up
again.
A
A
So,
let's
just
talk
about
the
the
biological
neuron,
as
I
mentioned
earlier,
there
is
about
10
percent
of
the
synapses
are
close
to
the
cell
body.
These
these
receive
feed
forward
input.
This
is
where
the
input
that's
like
the
sensory
intervals
or
the
feed
forward
patterns
come
to.
It
adds
linearly
in
the
cell
body
approximately,
and
this
is
what
generates
the
spikes
in
the
cells,
the
other
two
regions,
what
we
call
the
basal
dendrites
in
the
bottom
and
the
apical
dense
rights
on
the
top.
A
As
I
mentioned
earlier,
these
are
non-linear.
They
are
they.
They
generate
dendritic
action
potentials
and
they
depolarize
the
cell.
They
don't
make
the
cell
fire,
they
just
put
the
cell
in
a
state,
that's
ready
to
fire,
and
we
call
we
call
that
a
predictive
state.
We
model
this
basic
arrangement.
We
have
a
set
of
synapses
on
our
model,
neurons
that
are
feed
forward,
the
linear
summation,
we
activate
the
cell,
and
then
we
model
the
distal
synapses
as
a
set
of
coincidence
detectors.
A
Essentially
they
say
if
I
see
10,
15
or
20
synapses
active
at
the
same
time,
within
on
a
dendritic
segment,
we
will
generate,
may
put
the
cell
in
a
predictive
state
going
one
level
further.
We
have
to
talk
about
learning
biological
synapses.
These
are
the
connections.
You
can
see
a
little
section
in
the
dens
right
here.
You
can
actually
see
the
synapses
on
there.
A
We
now,
as
I
mentioned
earlier,
think
that
learning
is
mostly
about
forming
new
synapses
and
synapses
themselves
are
very
unreliable
things
they're
very
low,
fidel
fidelity.
They
don't
always
work.
You
know
you
know
they
kind
of
have
the
time
to
work
half
the
time
they
don't,
and
so
anyone
who
has
a
model
of
a
neuron
or
model
of
the
cortex
that
relies
on
high
precision,
even
one
or
two
digits
of
precision,
is
not
a
biologically
accurate
model,
because
their
synapses
aren't
very
good.
The
way
we
model
this
is
different
than
most
people
do.
A
Is
we
model
the
growth
of
a
synapse?
It's
like
it's
an
idea
called
potential
synapses.
You
have
an
axon
of
dendrite
that
are
near
each
other,
but
they
don't
make
a
connection
and
over
training
you
actually
grow
the
spine.
The
connection
between
these
two
this
is
well
documented
and
and
it's
a
growth
we
model.
A
So
we
give
that
a
scalar
value,
a
zero
to
one
value
and
when
we
train
we
sort
of
increase
that
scale
and
it's
like
growing
the
synapse
at
one
point
at
some
point
it
makes
a
connection
to
the
between
the
two
and
there's
a
threshold
for
that,
so
that
when
they,
when
the
permanence
hit
some
threshold
in
this
case
point
four,
we
say
the
synapses
exist
and
before
that
it
didn't
exist,
but
we
give
it.
We
give
it
a
weight
of
one
or
zero.
There's
no
scalar,
it's
just
a
binary
weight.
A
Then,
what's
the
point
of
having
this,
if
you
keep
training
it,
what's
the
point
of
having
the
permanence
go
up
higher,
it
makes
it
harder
to
forget
in
a
brain
what
happens
if
you
keep
repeating
and
training
a
real
synapse?
It
gets
thick
and
it
develops
this
bhutan
and
it's
just
much
harder
to
forget
it.
So
we're
trying
to
we
model
that
that,
as
opposed
to
the
weight
of
it,
okay
now
one
more
detail
and
then
we
can
start
showing
how
this
whole
thing
works.
Together,
I'm
going
to
talk
about
sparse,
distributed
representations.
A
Subatai
is
going
to
talk
in
much
more
detail
about
him
later
he's
going
to
give
you
some
of
the
mathematical
foundations,
I'm
just
going
to
give
you
some
of
the
concepts
right
now.
I
call
this
the
language
of
intelligence.
This
everything
in
the
brain
is
all
about
sparse,
distributed
representations.
You
can't
understand
how
any
of
this
stuff
works,
so
I
have
to
define
what
that
is,
and
so
we're
going
to
do
that.
A
Pretty
simply,
the
simplest
way
to
understand
is
compared
to
the
kind
of
representations
we
use
in
computers,
which
is
called
the
dense
representation
in
a
dense
representation.
You
might
have
a
byte
or
a
word
of
some
number
bits
and
you
use
all
combinations
of
ones
and
zeros
and
an
example.
Being
ascii
coding
is
a
perfect
example,
and
these
bits
have
no
inherent
meaning.
In
fact,
if
I
change
one
of
the
bits
in
ascii
code,
I
get
something
completely
different.
A
It's
the
whole
thing
has
to
be
looked
at
at
once
in
a
kind
of
arbitrary
assignment,
say
you
know
it
doesn't
really
matter
as
long
as
we
all
use
the
same
convention
in
the
brain
and
in
htm
theory
we
use
sparse,
distributed
representations
this
you
have
to
have
at
least
several
thousand
bits
to
do
something
useful.
Now,
when
I
say
a
bit,
you
can
think
about
it
as
a
neuron.
A
It's
a
neurons
either
active
or
it's
not
and
they're
sparse,
because
at
any
point
in
the
brain,
a
time
in
the
brain,
you'll
find
very
few
of
the
synapses.
Being
very
few
of
the
neurons
being
active.
Most
of
the
time
is
only
one
or
two
percent
that
are
active
and
most
and
most
of
the
time
the
rest
are
relatively
inactive.
So
when
I
talk
about
zeros
and
ones,
you
can
think
neurons,
okay,
active
inactive,
so
you
have
many
thousands
of
bits
and
they're
sparse,
because
we
have
a
very
small
percentage
of
number
one.
A
A
So
if
I
was
going
to
represent
a
letter,
I
might
have
a
bit
that
represents
you
know,
is
it
a
a
consonant
or
a
a
vowel,
or
how
is
it
sound
or
how
is
it
drawn,
does
have
a
sender's
descenders
things
like
this
attributes
of
it
and
I'd
pick
the
top
40
attributes
that
match
this
letter.
We
don't
do
that.
A
It
has
to
be
learned,
but
that's
the
basic
idea.
Okay.
So
what
are
the
properties
of
this?
First?
Very
simple
property
is
similarity
if
I
have
two
sparsely
stupid
representations
and
they
both
have
two
percent
of
their
bits
active
if
there's,
if
they
share
a
bit
meaning
if
they
share
a
cell.
If
it
was
in
the
brain
being
active,
then
they're
sharing
some
semantic
meaning,
and
this
does
not
happen
by
chance.
It's
very
very
because
these
are
sparse.
A
The
next
thing
you
want
to
do
we're
building
a
memory
system.
The
brain
is
a
memory
system,
so
the
first
thing
you
have
to
think
about
is
like:
how
do
I
store
a
pattern
and
how
do
I
remember
it
and
how
do
I
recognize
it
when
it
occurs
again?
So
imagine
I
got
these
thousands
of
cells
or
thousands
of
bits.
The
way
we
want
to
store
a
pattern
is
not
remembering
the
whole
pattern.
We
only
have
to
remember
the
locations
of
the
one
bits.
A
If
I
can
just
remember
where
the
one
bits
are,
and
I
can
see
a
new
pattern
come
in
and
say
well,
does
it
match
those?
If
it
does
I'm
good,
I
don't
have
to
look
at
all
the
other
ones.
So
in
this
case
I
might
have
index
of
41
bits.
But
what,
if
I
couldn't
do
that,
what
if
I
could
only
store
the
location
of
10.,
there's
some
subset
and
I
said
you're
not
allowed
the
story
of
the
location
of
all
the
one
bits
all
the
cells
are
active,
just
a
small
subsample
of
them.
A
Well,
what's
going
to
happen,
a
new
pattern
can
come
in
and
I
might
match
those
10,
but
the
others.
I
don't
know
about
you,
so
I
can
make
an
error
statistically,
it's
extremely
unlikely
for
this
to
occur
and
if
you
do
make
an
error
you're
making
an
error
for
something.
That's
semantically
quite
similar
to
the
thing
you
stored-
and
this
is
the
key
to
how
the
brain
generalizes
it's
it.
It
does
this
now.
If
you
haven't
made
the
connection
yet
I'll
make
it
for
you.
A
When
a
cell
wants
to
recognize
a
large
pattern,
it
only
has
to
form
a
small
number
of
synapses
to
other
cells
nearby
that
are
active,
so
it
might
be
hundreds
or
thousands
of
cells
nearby
that
are
active.
But
as
long
as
it's
sparse
activation,
an
individual
cell
only
has
to
make
connection
to
maybe
10
or
15
of
those
to
know
that
the
entire
pattern
is
there,
and
this
is
what's
going
on
on
dendrites
on
neurons.
A
There's
another
property-
and
this
is
the
last
one
I
get
into,
but
is
also
very
very
important
to
the
theory-
is
that
you
can
form
a
union
of
sparse
distributed
representations.
So
I
let's
say
I
took
10
of
them
and
I
just
ordered
them
together
and
I
now
have
a
new
one,
which
is
the
same
number
of
bits,
but
it's
got
more
one
bits.
It's
got
about,
20
a
little
bit
less.
Maybe
I
can't
undo
this.
I
can't
say:
oh,
what
was
the
original
10.?
I
can't
do
that,
but
I
can
ask.
A
Is
this
new
pattern,
one
of
the
original
10
and
I'm
going
to
claim
that
if
I
say
well,
if
the
new
pattern
ones
are
in
the
same
location
as
the
unions,
ones,
I'm
gonna
have
a
match
and
you
could
think
well.
I
could
make
a
mistake
there.
I
could
be
mixing
matching
from
different
of
these
patterns
that
I've
stored
earlier
the
union
again
extremely
unlikely.
For
that
to
happen.
The
math
shows
it's
almost
astronomically
unlikely
to
happen.
A
But
again,
if
you
do
you're
making
a
mistake
for
one
of
the
things
that
are
semantically
similar
to
ones
you
had
before
this
property
is
used
throughout
the
cortex
and
in
super
tight
talk.
He's
going
to
talk
about
that.
Where
is
this
I'll?
Give
you
one
example
where
it's
used?
Imagine
now
I'm
looking
at
a
neuron
and
I'm
looking
at
all
the
synapses
that
are
near
the
neuron
there's
several
hundred
of
them
and
let's
say
I
store
up
10
synapses
for
every
pattern.
A
I
want
to
recognize
and
I
just
throw
those
synapses
all
together,
so
I
now
have
700
synapses
representing
dozens
and
dozens
of
patterns.
Well,
that
cell
will
be
able
to
uniquely
identify
any
one
of
those
patterns
without
getting
confused,
and
so
we
actually
believe
that
we
have
a
sort
of
different
model
of
a
neuron
than
most
people
think
about
it.
We
think
that
the
neuron
in
our
models,
the
the
feed
forward
connections,
the
proximal
synapses,
can
can
activate
the
cell
from
dozens
of
feed-forward
patterns.
A
The
cell
can
actually
respond
to
many
different
patterns
in
a
feed-forward
case.
It's
not
just
one
thing:
it's
recognizing
and
it
can
respond
to
hundreds
on
the
on
the
more
distal
dendrites
that
is,
it
can
recognize
it
can
predict
its
own
activity
in
hundreds
of
contexts
that
are
unique
and
very
precise
all
right.
So
let's
put
this
all
together
in
in
how
we
get
this
to
work.
How
am
I
going
to
build
a
layer
of
cells
that
learns,
builds
a
predictive
model
and
we're
going
to
use
that
and
to
build?
A
This
is
a
picture
of
a
very
few
number
of
cells,
but
it's
about
two
percent
sparsity,
so
we're
just
showing
you
a
little
subset
and-
and
so
this
is
what
you
might
have
and
a
moment
later,
you
would
have.
This
is
maybe
one
time
a
moment
later.
You
have
a
different
pattern,
and
so
this
kind
of
thing
goes
back
and
forth
all
the
time.
A
A
A
Oh
I'm
going
to
be
next,
I'm
going
to
be
an
x2,
I'm
predicting
and
the
reason
you
have
here
we
have
more
cells
predicting
than
we
than
the
more
yellow
cells
than
red
cells
is
that
we
we
train
this
on
three
transitions:
a
to
b,
a
to
c
and
a
to
d.
So
if
I
show
it
a
it
predicts,
b,
c
and
d,
the
union
of
these
three
and
that's
what
the
yellow
cells
represent
the
union
of
three
patterns.
Now
this
is
the
beginning
of
sequence
memory.
It
says:
oh,
I
can
give
an
input.
A
I
can
predict
what's
going
to
happen
next,
I
can
make
a
union
of
prediction,
so
I
don't
have
to
have
a
precise
prediction,
but
I'll
know
if
any
one
of
those
three
occurred
and
I'll
know
precisely
if
they
occurred
now.
This
is
what
we
call
a
first
order:
sequence
memory.
It
is
not
able
to
solve
the
problem.
I
proposed
earlier
of
a
b
c
d
versus
x,
b
c
y,
and
we
need
to
solve
that
problem.
We
need
to
be
able
to
say
you
know
what
what
I
put
in
x
is
depend
on.
A
Something
happened
a
long
time
ago,
not
just
something
happened
a
second
ago
or
a
half
a
second
ago.
So
the
way
we're
going
to
solve
this
is
we're
going
to
use
these
mini
columns
and
I'm
going
to
walk
you
through
that.
This
is
about
as
deep
as
the
comp
that
the
top
talk
goes
and
we'll
come
back
and
get
easier
again.
A
So
now
we're
looking
at
a
slice
of
cortex,
I'm
showing
just
a
series
of
little
mini
columns
there,
and
this
is
a
cartoon
drawing
and
there's
six
cells
per
each
one
of
those
mini
columns,
and
in
this
case
I'm
showing
three.
When
you
have
a
feed
forward
input,
what
we
believe
is
going
on
it
actually
activates
the
mini
columns
each
one
of
these,
and
so
it's
a
sparse
representation
of
many
columns.
So
I've
shown
three
being
activated
now.
A
If
nothing
else
had
occurred
and
there
was
no
prediction,
what
will
happen
is
all
the
cells
in
those
columns
will
become
active.
This
is
a
I'm
unexpected
input
and
an
unexpected
input.
No
prediction:
I'm
going
to
activate
all
these
cells,
it's
sort
of,
like
I
don't
know
what's
going
on.
The
alternate
scenario
is
if
some
of
those
cells
were
in
the
predicted
state
shown
here
in
yellow
and
the
same
columnar
activation
occurs.
Those
cells
will
fire
first
and
inhibit
everyone
else,
and
I
get
a
very
sparse
representation.
A
The
same
columns,
but
now
it's
a
sparse
set
of
cells.
This
is
a
very
unique
representation
for
this
particular
transition.
If
I
I'll
walk
you
through
an
example
here
for
the
abcd
and
xbcy,
so
here
I'm
going
to
show
you
in
this
sort
of
cartoon
drawing
here's
our
sequence,
a
b
c
d
notice.
I've
shown
three
columns
active
in
each
of
those
representations.
A
This
is
before
training,
there's
no
expectation.
So
all
the
cells
fire,
here's,
the
here's,
the
sequence
for
x,
b,
c
y
notice
x-
is
different
than
a
different
set
of
columns,
but
the
b
columns
and
the
c
columns
are
identical
and
of
course,
then
I
have
a
difference
at
the
different
for
y
and
d.
At
the
other
end.
A
Now
after
training,
what
occurs
oops
there's
a
little
missing
dots
there.
I'm
not
sure
why
the
a
is
still
the
same
as
before,
and
you
can
ignore
those
missing
dots
there.
I
don't
they
weren't
in
my
presentation.
They
were
just
somehow
I
don't
know
what
happens
here
is
that
we
now
get
something
called
b,
prime,
since
it
was
predicted,
it
had
learned
this,
and
so
it's
the
same
columns,
but
now
individual
cells
in
those
columns.
A
This
is
d
in
the
context
of
a
b
c,
and
I
can
do
the
same
thing
for
x,
b,
c
y
and
you'll
end
up
with
a
different
representation
for
b
b,
double
prime
and
c
double
prime
and
y
double
prime,
and
this
is
basically
how
you
learn:
high
order,
sequences,
how
you
learn,
speeches
and
music
and
so
on.
The
capacity
of
the
system
is
amazing.
A
If
I
have
just
take
an
example,
if
I
had
40
active
columns,
a
very
small
number
of
columns
and
I
have
10
cells
per
column,
then
there
are
10
to
the
40th
ways
to
represent
the
same
input
in
different
contexts.
There's
10
to
the
40th
ways
to
say
this
in
this
context
is
unique
from
this.
In
another
context,
and
it's
when
you
start
thinking
about
it,
you
realize
your
life
is
full
of
this.
This
is
what
you're
doing
all
the
time.
There's
a
beautiful
mechanism
works
extremely
well.
A
We
put
all
this
together
and
we
can
make
this
whole
thing
work,
I'm
not
going
through
all
the
details.
We
end
up
what
we
call
the
htm
temporal
memory.
It's
a
it's
a
sequence
memory
it's
equivalent
to
like
a
cellular
layer.
It
converts
it
into
a
sparse
activation
in
columns
and
it
builds
a
temporal
model
and
it
has
some
really
nice
attributes
one.
It
learns
continuously.
That
is,
there's
no
batch
training
here.
Every
new
input,
it's
constantly
adjusting
its
synaptic
waves
and
it's
constantly
learning
it
can
extend
sequences
and
forget
them,
and
so
on.
A
It's
very
high
capacity.
Even
a
very
small
section
of
this
can
learn
millions
of
transitions.
It
uses
local
learning
rules
which
is
important
when
we
get
to
harder
implementations,
but
there's
no
global
supervisor
going
on
here.
It's
naturally
fault,
tolerant,
every
single
component
of
this
is
can
fail
and
and
every
individual
component
can
fail
and
nothing
bad
will
happen.
It
has
no
sensitive
parameters
and
it
semantically
generalizes.
These
are
all
really
desirable
features
in
our
memory
system.
A
Okay,
and
we
believe
this
is
a
building
block
for
both
the
cortex
and
for
machine
intelligence.
So
essentially,
I
can
do
a
variation
of
this
sequence
memory
in
each
of
the
layers,
and
if
I
can
build
a
region
of
cortex,
then
I
can
take
build
put
that
in
a
hierarchy
and
I'm
on
my
way
to
building
a
neocortical
system.
A
So
here's
our
little
map
of
the
four
layers
again,
you
should
recognize
them
by
now.
Hopefully
right
we
have
spent
most
of
our
time
working
on
what
we
consider
layer.
Two
three.
It
is
the
high
order
inference.
It's
a
high
order,
sequence
memory.
This
was
the
right
place
for
us
to
start
in
some
sense,
the
simplest
one,
the
one
that
we
could
characterize
easiest,
I'm
gonna
give
you
a
little.
I
say,
theory
98.
What
does
that
mean?
This
is
a
very
subjective
number.
It's
how
I
feel
about
it.
A
Okay,
that's
it's
just
intuition
about
how
much
we
understand
about
what's
going
on
here,
but
I
think
be
useful
for
you
to
share
that
with
you.
So
I
think
we
have
a
pretty
good
handle
on
this
there's
a
few
tight
ends.
We
may
have
to
tweak
and
clean
up
and
so
on.
This
has
been
extensively
tested
over
years.
We
put
it
into
commercial
products.
We
know
this
works
really
well
in
commercial
settings.
A
This
is
not
just
some.
You
know
ideas
on
paper.
We
have
taken
that
well,
so
we'll
go
on
to
the
next
thing.
The
next
we've
been
working
on
layer.
Four
here
I'm
pretty
confident
we
have
the
basic
idea.
What's
going
on,
you
take
that
same
sequence,
memory
and
you
feed
in
motor
commands
and
it
builds
a
predictive,
sensory
motor,
predictive
model
and
there's
a
lot.
A
This
is
where
things
really
get
interesting,
because
you
start
adding
behavior
and
robotics
to
the
system.
I
think
we
understand
about
half
of
what's
going
on
there.
I
think
we
have
some
really
good
foundation
principles.
We
haven't
started
implementing
any
of
it
yet
so,
but
we
think
about
a
lot
because
it
all
these
things
are
interplaying
all
the
time
and
then
finally,
on
the
layer,
six,
we
haven't
really
spent
much
time
at
all.
I
have
some
key
components
that
I
know
have
to
be
in
there.
A
I
give
it
about
a
10,
maybe
it's
20
understanding
of
it.
So
that's
where
we
are
in
terms
of
how
we
think
about
our
research.
Now,
since
we
did
layer,
2
3,
and
we
did
this
high
order
inference
engine,
we
then
turned
it
into
a
technology
and
we've
used
it,
so
you
can
think
about
what
can
I
do
with
this?
Well,
the
data
it
works
with
is
streaming
data.
It
has
to
be
data
that
changes
over
time
because
there's
no
behavioral
component
in
this.
A
It's
relying
on
data
coming
in
like
music
to
come
in
our
speech
or
data
streams
off
of
machines.
Things
like
that
streaming
data.
What
you
can
do
with
it
is.
We
can
model
that
data.
We
can
make
predictions,
we
can
detect
anomalies
and
we
can
do
classifications.
We've
shown
that
we
can
do
very
good
jobs
at
all
this
on
streaming
data.
The
applications
are
varied,
I'll,
just
mention
a
few
in
a
moment,
but
basically
you
can
do
predictive
maintenance,
security,
natural
language
processing,
anything
that
has
streaming
data.
A
These
are
the
way
we've
done.
This
is
we've
taken.
We've
built
a
simple
system.
We
take
some
data
stream.
You
have
to
run
it
through
an
encoder.
The
encoder
turns
that
data
into
a
sparse
representation,
so
we
have
encoders
for
numbers
and
categories
and
dates
and
times
working
with
a
company
called
quarter
clio.
We
have
an
encoder
for
words.
We
have
encoded
for
gps,
coordinates,
we've
done
a
bunch
of
these
once
you've
got
them
into
sdrs.
A
You
can
throw
them
into
a
high
order:
sequence,
memory
and
outcomes,
predictions
and
anomalies
we've
applied
to
a
number
of
problems.
We
have
a
commercial
product
called
grock
which
basically
detects
anomalies
in
servers
and
server-based
applications
on
aws
we've
applied
to
the
same
basic
idea
to
human
behavior.
So
can
we
detect
when
a
human
starts
acting
unusual?
Maybe
a
rogue
trader
who
starts
behaving
differently?
We
can
do
that.
We
can
detect
when
people
start
using
their
computers
differently
in
a
very
significant
way.
We
actually
think
we
can
we.
A
We
know
we
can
detect
anomalies
in
stock
volumes.
We've
done
some
cool
work
in
geospatial
anomalies.
We
can
attack
when
things
get
off
track
or
change
the
direction
or
the
change
at
different
speeds
and
so
on,
and
then
working
with
this
company
cortical
io
we've
shown
that
you
can
do
natural
language
processing
and
it's
very
cool,
and
you
should
see
the
demos
for
quark
io
later
today
and
I
think
chetan's
going
to
talk
about
it
too.
I'm
not
going
to
go
further
into
these
at
the
moment.
A
I
just
want
to
say
that
we've
that's
what
we've
done
with
layer
3.
and
by
the
way,
all
of
these
use
the
exact
same
code
base.
This
is
not.
We
don't
even
have
to
tweak
it.
It's
the
exact
same
code
base
and
we're
getting
close
to
that
universal
algorithm.
So
you
can
as
long
as
you
get
the
data
into
a
sparse
representation,
you're
good
okay.
So
that's
what
we've
done
mostly
now.
What
could
I
do
with
the
sensory
motor
inference
the
stuff
we're
working
on
right
now?
Well,
what
would
it
be
good
for?
A
Well,
first
of
all,
it's
it's
good
for
working
with
static
data,
and
but
it
needs
some
sort
of
simple
behavior.
So,
for
example,
we
could
look
at
a
picture
which
is
static
data,
but
to
train
the
system.
I
have
to
move
the
eyes
over
it
and
I
can
do
it
in
a
fairly
stupid
way
and
still
get
good
results.
A
A
A
There
are
other
things
you
can
do
with
a
very
cool.
You
could
do,
for
example,
classify
networks,
any
structure,
that's
out
there.
I
could
have
a
set
of
some
sort
of
very
complex
computer
network,
and
I
want
to
classify
it.
It
could
be
n-dimensionally
complex,
but
I
want
to
classify
and
say
well
what's
it
like,
and
what
is
it
similar
to?
I
think
this
technology
would
work
for
that.
So
some,
I
think
some
very
clever
applications
are
going
to
come
out
of
this.
A
Finally,
of
course,
if
you
go
to
adding
and
motor
behavior,
then
things
really
get
interesting.
It
doesn't
matter
you
can
have
static
or
streaming
data,
but
the
capability
you're
going
to
get
now
is
not
just
simple
behavior
but
goal
oriented
behavior
where
the
cortex
itself
starts
saying
this
is
the
behavior.
A
I
want
to
achieve
a
particular
result,
a
particular
predicted
result,
and
how
do
I
get
there,
and
this
is
when
we're
going
to
be
able
to
enter
into
robotics
now,
when
I
think
of
robotics,
it
could
be
physical
robots,
but
mostly,
I
think,
about
things
that
are
not
physical
robots.
I
think
about
things
that
are
like
smart
bots
things
that
are
scouring
the
web,
trying
to
figure
out
how,
where
bad
guys
are
things
like
that
or
proactive
defense
anywhere.
A
You
have
some
sort
of
system
where
the
the
intelligent
machine
is
navigating
intelligently
through
some
sort
of
structure,
and
it's
trying
to
achieve
certain
results.
That's
going
to
really
where
the
whole
thing
opens
up
and
I'll
mention
briefly
the
layer
6
thing
it's
not
in
the
same
category.
Essentially,
this
is
necessary
for
hierarchy
and
it's
going
to
be
really
necessary
for
building
very
large
multi-sensory,
multiple
multi-behavioral
modalities,
but
the
other
threes
are
really
the
key
things
here.
So
that's
our
research
roadmap.
It
gives
you
a
sense
of
where
we're
going.
You
might
be
saying.
A
Well,
how
long
is
this
going
to
take?
I
don't
know
we
did.
We
did
it
took
us
a
long
time
to
really
figure
out.
What's
going
on
the
the
top
part
there,
the
layer
three,
but
now
it's
a
lot
quicker.
We
were
able
to
do
the
layer,
four
stuff
and
figure
that
out
much
faster,
and
I
think
it's
accelerating.
So
I
don't
know
how
long
this
is
all
going
to
take
to
play
out,
but
I
certainly
hope
to
be
part
of
it
and
I'm
working
at
it.
Okay.
A
Another
part
of
our
research
roadmap
is
essentially
our
approach
to
doing
research
and
we're
very
open
and
transparent,
and
so,
as
you
probably
know,
all
of
our
algorithms
are
documented.
People
created
multiple
independent
implementations
of
them.
Our
software
is
open
source
under
gpl
version.
3
license.
We
have
a
new
pic
open
source
project.
You
can
find
it
at
nomento.org
tomorrow
and
sunday
we're
having
a
hackathon
down
in
san
jose
there's
active
discussion
groups
for
theory
and
implementation.
A
A
We
have
a
long
going
collaboration
with
ibm
research
in
san
jose.
This
is
the
group
we
probably
haven't
heard
about,
but
they're
interested
in
doing
hardware,
implementations
of
these
algorithms
there's
a
similar
relationship
with
darpa
in
washington
dc.
There's
a
program
called
the
cortical
processor,
which
is
also
based
around
htm
principles,
and
we
work
with
other
small
companies.
I
mentioned
cortical
io
they're,
doing
the
natural
language,
processing
and
you're
welcome
to
contact
me
and
momenta,
and
I'm
we're
all
very
open
and
available
to
talk
about
this
stuff.
A
I
know
where
the
big
picture
what's
going
on,
so
we
got
there's
a
lot
of
confusion
in
this
space
about
all
these
different
approaches,
and
I
may
not
be
able
to
clear
up
that
confusion,
but
at
least
I'll
share
how
how
I
think
about
this,
the
I
think,
there's
sort
of
three
basic
approaches
to
building
intelligent
machines
on
the
left
is
the
people
like
ourselves
who
say
you
need
the
model
of
cortex.
These
are
cortical
modelers.
He
said
all
right.
We
got
a
brain,
it's
smart,
let's
figure
out
what
it
does.
Let's
model
it.
A
I
use
htm
as
an
example,
because
I
think
it's
one
of
the
best
and
most
advanced
theories
in
this
space
on
the
then
there's
the
sort
of
the
artificial
neural
network
world.
The
current
favorite
right
there
today
is
deep
learning
and
that's
getting
a
lot
of
press
a
lot
of
success,
and
then
you
have
the
more
classic
ai
and
lots
of
different
things
in
on
the
all
these
categories,
but
I'll
use
watson,
because
that's
it's
been
in
the
news
recently
and
ibm's,
pushing
it
very
heavily.
A
Okay,
so
they're
basically
built
on
different
premises.
One
premise
on
the
cortical
modeling
is
that
biology
matters
and
we
focus
on
biology.
We
use
the
biology
as
a
set
of
constraints.
We
don't
think
it's
a
nice
guideline.
It's
it
bet.
This
is
the
real
mccoy.
I
need
to
understand
how
the
biology
works,
and
once
I
understand
it,
I
can
deviate
from
it.
But
I
don't
just
we
don't
make
up
stuff
willy-nilly
and
say:
well,
I
think
it
might
work
like
this.
A
Just
try
that
now
we
constantly
go
back
to
the
neuroscience
and
constantly
go
back
to
biology,
say
this:
no,
it
can't
work
that
way.
It
has
to
be
something
like
this.
That's
what
drives
us.
The
artificial
neural
networks
are
really
mathematically,
driven
they're,
not
biologically
driven
at
all,
despite
the
fact
that
they're
called
neural
networks
and
people
say
they're
brain
like
they're,
not
the
neurons
they
use
are
totally
unlike
real
neurons.
The
networks
they
use
are
unlike
real
networks.
A
The
training
programs
they
use
are
unlike
biological
training
paradigms,
but
what
they
do
is
they
have
a
mathematical
foundation.
They
can
prove
that
these
algorithms,
these
networks,
will
converge
or
they'll
produce
the
right
result
and
that's
a
very
powerful
thing.
Sometimes
people
I've
been
told
over
and
over
again
from
some
of
the
people
in
this
camp
like
well.
You
can't
understand
if
htms
work,
because
you
don't
have
a
mathematical
foundation
for
it
and
you
just
won't
know
if
it
works.
A
I
said
well,
I
can
see
why
it's
going
to
work
and
I
build
it
and
it
does
work.
They
say.
That's
not
good
enough.
Well,
I
don't
know,
but
you
know
I
say
we
build
computers
and
there's
no
formula
to
represent
how
a
computer
works,
and
we
seem
to
be
happy
with
those
so
and
then,
of
course,
in
the
ai
world.
It's
basically
an
engineered
solution,
we
say:
okay,
let's
we
have
a
problem,
let's
engineer
a
solution
for
that,
the
data
they
work
on.
It's
a
little
bit
different.
A
We
work
with
spatial
temple
data
right
from
the
get-go.
We
knew
we
knew
that
brains
are
all
about
temporal
data,
spatial
temple
data
and
we're
starting
to
add
simple
behaviors
into
it.
So
I've
given
us
credit
for
that
one.
The
artificial
neural
networks
are
primarily
spatial
problems.
Deep
learning
networks
are
primarily
spatial
classifiers
and
they
they
realize
they
have
to
add
the
temple
component
and
a
lot
of
the
researchers
they
are
talking
about
and
working
on.
A
A
The
capability
is
a
little
bit
different
too.
The
cortical
models
are
basically
predictive
models
and
we
can
do
classification
and
we're
starting
to
think
about
how
to
do
goal.
Oriented
behavior
today,
artificial
neural
networks,
like
deep
learning,
are
really
just
classification
networks
and
they're
really
good
at
it,
but
they're
classification
networks
and
then,
of
course,
if
they're
watching
this
more
like
natural
language
querying
now,
all
of
these
things
are
valuable.
I'm
not
trying
to
put
a
value
judgment
on
them.
They
all
solve
problems,
they're
all
very
useful.
A
This
is
not
saying
one's
better
than
the
other
they're
just
different
approaches,
but
what
I
do
think
is
is
is
is
true,
and
my
next
point
here
is:
are
they
on
the
path
to
true
machine
intelligence?
Are
they
on
the
path
to
what
we
all
think
about
of
intelligent
machines?
And
in
that
case
I
argue
that
the
cortical
modelers
are
the
only
ones
who
are
on
that
path.
We
are
definitely
there.
I've
laid
out
a
path
here.
I've
talked
about
the
components.
A
I've
talked
about
how
all
this
stuff
fits
together
and
a
theory
about
how
the
cortex
works
and
behavior
we
have
a
road
map,
even
if
we
don't
understand
it
all.
We
know
where
we're
going
the
ai
world
no
they're,
not,
and
I
don't
make
that
up
the
guys
who
created
watson
said
so
themselves.
They
said
this
is
not
intelligent
machine.
A
It's
not
going
to
lead
to
intel's
machine,
but
it's
really
cool
for
what
it
does
and
good
for
them,
and
I'm
going
to
argue
in
an
artificial
neural
network
world
they're,
probably
not
on
a
path
to
a
machine
intelligence
if
they
want
to,
and
if
it's
going
to
get
there,
they
have
to
add
time
they
have
to
have
behavior.
They
have
to
add
sort
of
these
broader
concepts.
Sdrs
things
like
that
that
I
talked
about-
and
my
hope
is
that
these
worlds
converge.
A
But
I
do
think
that
the
way
to
get
there
and
from
my
belief
is
to
start
with
the
cortical
models.
So
that's
my
view
of
where
we
are
and
where
we're
going
I'll
just
end
with
a
few
comments,
and
then
we
can
do
some
questions.
I
I
think
we're
at
a
pivotal
time
in
humanity.
This
is
a
really
really
interesting
time
to
live.
We
are
at
a
time
when
we
are
actually
figuring
out
how
the
brain
works
and
there's
nothing.
A
In
my
mind,
that
can
be
more
interesting
than
that
I
mean
we
are
a
species
is
defined
by
our
brain.
That's
really
the
only
thing
that
makes
us
unique
we're
not
really
good
at
anything
else
and
and
and
so
and
everything
we
do.
That's
interesting,
our
knowledge,
our
language,
our
arts,
it's
all
products
of
brains
and
knowledge-
is
these
only
can
be
understood
by
brains
and
the
scientific
process
is
the
process
the
brain
uses
I
mean
so
to
understand
humanity.
A
We
have
to
understand
brains
and
the
idea
that
we
can
now
build
machines
that
work
on
those
same
principles,
to
help
us
discover
things
faster
and
bigger
and
and
and
apply
to
problems
that
we're
not
very
good
at
is
tremendously
exciting.
To
me,
this
is
not
about
building
robots
or
you
know
machines
are
going
to
take
over
the
world.
Nothing
like
that
at
all.
A
It's
really
impossible,
but
I
know
they're
going
to
be
amazingly
cool
and
I
think
this
is
going
to
be
driving
technology
development
for
the
next
hundred
years
and
we
we
have
a
chance
at
least
the
opportunity
to
participate
that
now
and
times
it
seems
very
hard
and
difficult
to
do,
but
I
think
we're
making
really
great
progress
so
again,
thank
you
for
coming
here.
I
appreciate
your
time.