►
From YouTube: HTM on Hardware (Part 6)
Description
Jeff talks about HTM hardware implementations.
You should have some background in HTM before understanding these videos. See http://numenta.org/htm-school/ for more videos explaining basic HTM theory.
Music: "Holy Roller" by YACHT
(used with permission from Free Music Archive)
A
This
is
a
kind
of
switching
topics
here,
I
tell
you
what
there
were
a
lot
of
people
is
still
interested
in
HTM
hardware,
hardware
architectures
that
could
support
HTM
from
what
I
understand
at
a
basic
level,
I
mean
no
matter
how
much
we
extend
the
theory,
there's
a
certain
number
of
requirements
that
we
could
apply
right
now
to
hardware
architecture
that
could
support
HTM
in
the
future.
And
what
can
you
talk
about
any
movement?
That's
going
on
in
that
space
right
now,
I
know.
B
So,
there's
a
there's,
a
large
community
of
people
who
are
trying
to
build
normal
market
heart
right
and
these
efforts
have
been
going
on
for
decades,
I
mean
the
problem
is
the
plasticity.
Let's
what
most
people
are
interested?
Why
what's
motivating
most
people
to
build
their
own
market
partner
and
what's
motivating
most
funding,
for
it
is
they're
trying
to
come
up
with
very
low
power
computing
systems
doctor
that
that's
the
number
one
driver,
because
the
brain
is
extremely
the
all-powerful.
B
B
B
We
can't
just
build
a
regular
computer
hold
power.
We've
already
thought
that
so,
let's
try
to
build
a
brain
like
computer
cuz,
we
know
brains
can
run
at
low
power
sure
this
is
where
that
movements
coming
from
now
they
said:
okay.
Well,
let's
start
building
they're
all
networks
in
these.
This
neuromorphic
hardware.
B
Need
the
president
was
there
and
the
neuron
models
are
not
biologically
realistic
at
all.
Now
those
could
be
very
useful
for
deep
learning
type
of
models
because
deep
learning
models
were
built
on
any
computer
they
need,
while
they
they
I,
don't
know,
I,
don't
know.
If
that's
true
or
not
some
of
them,
they
can
do
a
lot
of
software,
but
but
you
could
use
a
point
neuron
model
in
memory
card
where
to
do
deep,
learning
and
there's
some
companies
are
formed
to
do
that.
B
There
are
other
people
really
interesting,
building
real
intelligent
machines
and
that's
not
as
interesting
to
them,
and
so
it
would
been
interesting.
I'll
just
give
it
one
particular
point:
I,
a
couple
years
ago,
I
met
Carl,
Heine's
Meyer
who's
in
charge
of
the
hardware
of
the
neuromorphic
computing
part
of
the
human
brain
project
in
Europe.
So
that's
the
European
wide
effort
to
understand
how.
B
Charge
of
the
hardware
part
of
that-
and
he
had
they
have
a
this
hike
and
chipset,
which
was
really
clever
and
really
smart
and
design.
They
have
big
systems
now
you
can
use,
but
it
was
built
on
the
point
your
honor
so
two
years
ago
was
it
cons
and
I
were
having.
We
both
spoke
at
the
same
conference
back-to-back
and
he
came
up.
B
B
B
Part
of
the
model
is
plasticity,
but
he
wasn't
even
think
about
yet
he
was
this
thing.
Like
my
neurons,
don't
have
active,
dendrite,
you're,
saying
people
say
real
arms.
Do
no
one's
told
me
why
I
need
them?
You
first,
you
know
you're
one,
the
first
people
who
ever
told
me
why
I'm
a
realistic
neural
model
would
need
them
right.
So
what's
something
HTM
Theory
has
provided.
We
have
really
probably
only
legitimate
model
of
neurons
that
require
an
apical
and
distal
dendrites
and
thousands
of
in
it
yeah.
B
So
we
started
working
with
carlines
and
some
of
the
people
in
his
in
his
research
group,
we
had
a
visitor
Sebastian.
You
remember
who
worked
on
porting
HTM
to
the
hi
can
check
that
River
paper.
We've
got
a
paper
guide
on
our
website
says
it
was
not
easy.
It
wasn't
designed
for
this.
It
was
like
you
know:
okay,
we're
gonna.
Try
to
you
know
make
this
thing
work,
be
a
good
luck,
yeah,
so
easy.
B
So
we
you
know,
Sebastian
came
and
we
really
tried
to
try
to
get
how
the
high
can
ship
from
support
a
ship.
It's
really
hard.
It's
I
mean
very
clever
engineering
on
both
the
Sebastian
subitize
part
yeah,
and
but
here's
the
good
news
about
it.
I
just
saw
Colin
Hines
a
couple
weeks
ago
at
this
neural
information
computing
elements
conference
as
a
nice,
the
nice
conference
in
San
Jose,
and
he
told
me
that
their
next
generation
they're,
including
active
dendrites
him
and
it.
B
A
B
B
Some
of
our
content,
but
we're
going
to
add
new
things.
So
you
know
we
may
not.
So
it's
just
supporting
the
neurons
yes,
I
would
argue.
History
says
suggests
that
we'll
need
at
least
one
or
two
more
rounds
of
tweaking
to
get
this
right
explains
getting
better,
probably
break
yeah.
These
ones
get
any
better,
but
we
should
be
able
to
do
something
like
we
should
be
able
no
it'll
be
useful.
Definitely
I,
just
think,
don't
think
I'll
be
done
right.
Yeah,
I'm,
certainly
glad
they're
working.