►
From YouTube: HTM Hackers' Hangout - Sep 7, 2018
Description
Today I'll give a brief update on the two papers we are working on and take community questions.
HTM Hackers’ Hangout is a live monthly Google Hangout held for our online community. Anyone is free to join in the discussion either by connecting directly to the hangout or commenting on the YouTube video during the live stream.
If you have something specific you’d like to discuss, or if you just want to learn more about the HTM Community, please join HTM Forum at https://discourse.numenta.org. We have active discussions about HTM theory, research, implementations, and applications.
More info on all these topics at http://numenta.org.
A
Hello
and
welcome
to
HTML
cars
hanging
out
its
September
7th
I'm
Matt
Taylor
from
the
Mendte
thanks
for
watching
it's
been
a
couple
months
since
I've
done
one
of
these
because
it's
been
busy
I've
had
a
vacation
in
the
mean
time,
so
that
was
part
of
it
too.
So,
thanks
for
joining
us,
hi
Paul
Lam
I've
got
one
of
our
community
members
online
with
us.
I've
been
any
time.
Paul
I'm
gonna
go
over
my
agenda
and
then
we
can
chat
about
anything
one.
A
So
I
don't
have
anything
to
show
today,
I'm
just
going
to
give
a
few
for
the
community
updates,
sorry
about
the
noise,
I
changed
offices
and
there's
like
a
Main
Street
right
down
there,
and
sometimes
ambulances
go
by.
Sometimes
people
drive
really
loud
cars
anyway.
Community
has
been
great.
If
you
aren't
a
member
of
HTM
forum
go
sign
up
because
there's
tons
of
discussion
there
if
you're,
watching
this
and
you're,
not
on
HTM
forum,
that's
like
the
place
to
be
to
talk
about
HTM
of
conversations
about
really
interesting
things.
A
We're
doing
a
lot
of
discussion
right
now
about
the
thalamus,
which
is,
which
is
one
of
the
things
that
we're
focused
on
research,
is
trying
to
understand
the
thalamus
and
what
role
it
plays
in
in
cortical
processing
and
that
cortical
circuitry
that
we're
always
talking
about.
How
does
this
thalamus
play
that
role,
especially
between
levels
of
the
hierarchy
in
between
the
cortex
and
these
senses,
that
sort
of
stuff
lots
of
discussion
about
that
on
the
forums
about
the
thalamus?
And
that
is
like
one
of
our
big
research
areas.
A
We
have
two
papers
coming
out
soon.
This
fall
they're,
both
almost
done.
I've
read
one
of
them
and
the
other
one
I'm
working
on
it.
So
the
first
one
and
I'm
gonna
I'm
gonna,
do
a
post
on
this
on
the
forum
that
kind
of
goes
over
our
main
papers.
I
think
it
gives
a
little
summary
of
each
papers
and
why
it's
important
and
how
it
builds
on
the
previous
one.
A
So
the
next
one
working
on
we're
sort
of
internally
calling
columns
plus
right,
but
it
takes
where
that
one
let
leaves
off
and
gives
software
simulation
and
lots
of
details
about
it.
So
we
talk
specifically
about
what
displacements
are,
how
they
can
be
applied
to
a
set
of
grid
cell
modules
to
update
them
to
into
a
new
location
so
that
we
so
we'll
have
simulations
of
that.
I'll
have
lots
of
visualizations
of
that
soon.
A
There
are
some
great
graphics
in
this
paper
that
I'm
going
to
adapt
into
the
next
HTM
school
or
two
episodes
of
that.
So
you'll
see
this
paper
converted
into
video.
Soon,
that's
one
of
the
things
I'm
working
on,
but
the
papers
is
about
locations
in
the
neocortex
and
sensorimotor
object,
recognition
with
cortical
grid
cells.
So
the
big
idea
is
how
we
think
grid
cells
work
in
the
cortex
for
object,
recognition
and
we'll
give
concrete
simulations
and
examples
of
that.
A
So
that
will
be
in
the
columns
plus
paper
and
if
you
like,
math
they'll,
be
some
math,
so
you'll
be
happy
there
and
including
simulations
I
haven't
gotten
to
the
simulations
yet
before
this
goes
out.
I'm
going
to
be
running
all
these
simulations
and
and
testing
them
out
and
everything
so
but
I
haven't
gotten
quite
that
deep,
yet
I'm,
still
working
on
understanding
all
of
the
the
graphics
and
the
math
math
was
the
hard
part
for
me,
of
course.
A
B
A
Not
the
the
columns
plus
paper
I,
just
described
as
very
detailed
lots
of
like
I,
said
lots
of
math.
There's
code
go
along
with
it.
The
frameworks
paper
is
going
to
be
more
of
a
high-level
paper
and,
describing
you
know
the
system
as
a
as
a
framework
and
the
idea
I
think,
and
definitely
we'll
talk
about
this
when
it
comes
out.
I'm,
basically
talking
to
you,
the
community,
like
this
small
conversations
and
generally
people
out
there
in
the
public
and
on
Twitter,
don't
don't
watch
these
hangouts,
but
for
you
guys,
you
know
this
frameworks.
A
A
We
know
so
one
part
over.
Here's
got
to
be
doing
something
like
this.
Another
person
got
to
be
something
like
this
and
I
got
to
be
connected
somehow
like
this
right.
So
the
frameworks
is
putting
it
filling
in
the
things
that
we
know
about
the
brain
and
and
and
making
sure
that
there's
room
for
the
other
things
and
we
know
we're
happening,
but
maybe
we
don't
know
how
or
we
don't
know
how
it's
connected.
A
We
know
something
over
you're
like
this
is
happening
so
putting
it
together,
a
framework
of
cortical
product,
not
just
cortical
processing
you
like
brain
processing.
You
know
cuz,
like
they
said
we're
talking
about
thousands
now.
You've
talked
we've
studied
a
lot
about
hippocampus
and
around
cortex
that
contributes
to
how
we
think
things
are
working
in
the
neocortex.
Obviously,
so
the
frameworks
is
going
to
be
much
more
high-level,
I,
don't
think
it's
going
to
be
the
code
associated
with
it,
but
I
will
probably
still
like
do
another.
Do
an
HTM
video
about
this
frameworks
idea.
A
So
those
are.
The
two
papers
are
working
on
columns
plus
lots
of
detail
about
locations
in
the
new
cortex
and
object.
Recognition
includes
math
includes
code
and
the
frameworks,
one
which
is
much
more
high-level,
so
those
will
come
out
is
probably
similar
times
within
the
next
month
or
two
Jeff's
got
a
bunch
of
speaking
engagements.
This
fall
we're
trying
to
time
the
publication
of
these
papers
so
that
you
know,
if
you
can
talk
about
them
at
these
speaking
engagements.
A
We're
hoping
that
there's
a
couple
of
journalists
that
come
out
with
the
article
or
two
about
Numenta
is
sort
of
at
the
same
time,
because
I
really
like
this
column's
paper,
I,
think
you
guys
in
the
community
that
have
been
working
on
object,
recognition
and
how
it
could
work
are
gonna
like
this
I'm
going
to
be
able
to
dig
into
it
and
actually
do
some
things
with
it.
Oh
I
haven't
seen
the
code
yet
I
know.
Marcus
has
and
Scott
and
I've
been
really
working
on
that
code,
getting
it
cleaned
up
and
Louis
who's.
A
A
So
that's
my
update,
it's
all
the
short
and
sweet,
only
seven
minutes
or
so,
but
we're
working
on
papers.
That's
the
main
thing
soon.
As
those
get
out,
we
should
have
a
nice
flurry
of
activity.
This
fall
with
these
publications
coming
out
and
some
toxic
Jeff
is
giving
and
hopefully
a
couple
of
articles.
So
that's
what's
going
on
at
momenta.
Sorry,
I!
Don't
have
anyone
on
the
research
team
right
now,
Jeff's
out
of
the
office
today
take
a
break
sometimes,
and
but
they
have
been
really
busy
and
I
really
am
liking.
A
This
stuff
that
they're
they're
putting
out
now
now
that
I
see
the
papers
going
and
then
editing
them
and
I'm
getting
to
read
them.
It's
good,
good
stuff,
so
happy
to
see
that
now
it's
time
for
you
guys,
Paul
I,
know
you're
online.
You
want
to
talk
about
anything
and,
oh,
let
me
see.
There's
somebody
chatting
thanks
for
your
work,
no
problem
all
right.
Any
questions
on
chat,
I'll,
throw
it
on
chat.
If
you
don't
want
to
join
anything
from
you,
Paul
I
am
no.
A
B
B
A
Great
thanks
for
your
efforts
and
thanks
for
being
a
important
part
of
the
community
pop
I
love
your
posts
and
your
responses
to
people,
especially
helping
out
new
people
that
are
asking
questions.
I,
don't
have
time
all
the
time
to
come
in
and
answer
them
and
you've
been
great.
You
saved
me
a
lot
of
time.
So
thank
you.
Thank
you.
A
C
Hey,
hey,
Matt,
good,
to
see
you
yeah
I
just
got
this
emails
like
oh
they're,
doing
one
now
I
like
your
long
hair,
yes,
not
too
relevant
to
most
things:
yeah
yeah,
it's
cool
I
know
you
always
have
a
cool
hair.
So,
let's
see
I
would
I.
Maybe
I'd
ask
a
quick
thing,
so
I'm
being
considered
for
a
possibly
a
job
or
internship,
where
it's
based
on
monitoring
real-time
video
feeds,
which
is
a
thing.
Obviously
it's
a
grown.
C
You
know
popular
thing
and
they
talked
a
lot
about
like
convolutional
neural
network,
and
you
know
we're
gonna
have
like
a
team
convolutional
neural
network.
I
know
there
was
some
stuff
as
I've
a
Glee
recall.
Maybe
this
was
back
in
the
Zeta
one
days
or
something
where
there
was
some
use
of
HTM,
fury
at
least
sort
of
toward
that
or
I
don't
know
if
there's
because
they
want
to
be
able
to
like
detect.
C
If
there's
a
person-
and
you
know
in
a
certain
place
or
if
there's
you
know
like
a
high-sensitivity
area,
I,
don't
know
if
there's
been
I'll
go
and
recheck
the
check
the
discussion
stuff.
But
if
there's
any
I,
don't
there's
anything
that
you
know
of
in
terms
of
maybe
infusing
HTM
into
that
somehow
it
is
anything
come
to
mind,
I'll,
go
and
recheck
the
discussion
or
reread
that
old
stuff,
but
just
off
the
top
I
don't
know
well,
two
things
come.
A
You
know,
and
everybody
knows
how
to
use
deep
learning
and
not
everybody
knows
how
to
use,
especially
that
old
old
version
of
HTM,
which
didn't
include
the
things
like
STRs,
and
so
we
were
the
hierarchy.
Part
was
working
well
back
then,
but
you
know
the
basics
weren't
biologically
correct.
So
that's
why
we
sort
of
threw
that
out
throughout
the
baby
with
the
bathwater,
so
to
speak,
because
that
application
does
not
really
apply
anymore
to
or
current
technology
stack.
A
That
being
said,
and
the
other
thing
I
think
it
will
eventually
I
mean,
but
but
there
is-
and
we
talked
about
this-
we
talked
about
this
at
the
Manhattan
hackathon
a
few
years
back.
Frank
Kerry
is
an
AI
ki
guy
and
in
Manhattan
he
his
hack
was
like
okay.
Let's
take
a
streaming
video
and
run
convolutional
neural
networks
on
frames
of
the
video
that
can
recognize
objects
within
the
frame
right,
so
they
can
do
feature
detection
yeah.
A
B
A
If
anybody's
ever
done,
but
you
take
that
sort
of
output
and
treat
that
feature
array
like
whether
something
exists
in
the
scene
or
not
as
an
SDR
mm-hmm
and
then
how
it's
basically
deep
learning
do
your
feature
detection,
but
for
every
frame
of
the
video
have
a
new
SDR.
That
tells
you
what's
in
the
frame
like
what
features
exist
in
the
frame
and
then
you
should
be
able
to
over
time,
especially
if
you're
encoding
time.
B
B
Yeah
another
way
that
I've
sort
of
married,
deep
neural
networks
with
with
HTM
is
basically
usually
get
an
output.
Like
you
said
where
you've
got
this,
the
the
percentage
of
this
concept
being
present
is
X
percent.
The
percentage
of
this
is
X
percent
right,
it's
usually
a
it's,
not
ever
a
hundred
percent.
You
always
have
like
you
know.
Ninety
percent
or
two
percent,
or
thirty
six
percent
or
whatever
then
take
the
the
word.
B
A
B
B
More
likely
to
Union
you
need
to
know,
is
it
sort
of
like
if
you
look
at
the
videos
on
cortical
dye,
oh
and
on
how
they
make
fingerprints
for
phrases,
it's
the
same
concept
yeah,
if
the
same,
if
three
SDRs
have
the
same
bit,
then
that
gets
a
count
of
three
and
that's
going
to
have
a
higher
precedence.
You're
gonna
pick
that
one
before
you
pick
one
has
a
count
of
two.
For
example:
that's
smart!
If
you'll
come
up
whatever
spar
sea
level,
you're
looking
for
yeah.
A
And
you
should
get
you
should
get
representations
of
object,
features
that
way
not
just
objects
that
the
kind
of
neural
network,
the
deep
learning
part
identifies
you
get
an
understanding
of
features
of
others.
Don't
like
you
said.
If
there's
a
dog
and
a
cat,
both
in
the
frame.
At
the
same
time,
you
should
have
more
furry
bits
right,
yeah,.
B
Exactly
or
you
know,
a
man
walks
by
a
2
a.m.
everyday,
then
a
woman
walks
by
a
2
a.m.
another
day,
you're
gonna,
see
it's
gonna,
be
some
you're
going
to
see
a
similar
semantic
similarity,
but
it's
not
going
to
be
exactly
the
same.
So
it's
the
same
kind
of
a
concept
here
where
you
know
you've
taken
the
work
that
cortical
dot
IO
has
done
to
relate
words
together,
and
then
you
know
combine
that
with
what
you
get
on
your
output
of
your
image.
Analysis
do.
A
C
That's
great
I,
don't
suppose
that
that's
that's
like
written
anywhere
or
anything
like
did
you
just
kind
of
like
think
of
that
or
because
yeah
I
really
I.
That
sounds
great
to
me.
I'd
look.
Good
I
have.
B
C
B
C
Idea.
Thank
you.
Thank
you.
That's
I
appreciate
that
and
yeah
very
very
cool
yeah.
If
you
happen
to
feel
inclined
to
do
that,
I'll
totally
I'll
totally
have
a
have
a
look
at
that
because
yeah,
it
seems
like
you
could
use
like
some
kind
of
pre-processing
with
deep
learning
or
something
to
get.
You
know
the
idea
that
takes
strike
like
oh,
that's
a
person
and
then
you
know
involve
the
richness
of
HTM
could
really
could
really
add
on
to
that.
C
So
I'd
love
to
be
able
to
bring
that
as
a
you
know,
as
a
thing
maybe
to
this
to
this
company
or
at
least
understand
how
I
could
so
anyways.
Thank
you
and
yeah
I
would
just
second
what
he
said:
Paul
your
your
your
posts
are
always
really
really
cool.
I,
like
how
you
jump
on
on
helping
out
new
people
and
stuff
I.
Try
to
do
that.
A
bit
and
you're
usually
right
there
so
anyways,
but
but
to
have
you
there,
of
course,
and
Matt
to
the
flagbearer,
so
yeah.
A
A
C
A
A
A
Right
no
problem,
thanks
for
joining
nobody's,
asked
some
questions
in
chat,
so
I
am
going
to
sign
off
thanks
everybody
for
joining
today's
HTM
backers.
Hangout
see
you
in
a
month
see
you
on
the
forums.
If
you
have
any
questions
contact
me
there
or
just
ask
the
question
on
the
forum.
That's
best
best
thing
to
do.
If
you're
worried,
like
I,
don't
know
if
I
should
put
it
up,
just
put
it
on
the
forum
if
I
don't
like
it,
I'll
move
it.
Oh,
oh
there's
someone
says:
oh
hi,
David
David's
online
eye.
A
He
said
hi
David
cognition,
you
know
our
java
guy
and
then
someone
had
a
question
about
any
similarity
with
like
curious.
So
if
I
carry
us
know,
I
guess
I,
don't
know
what
they're
doing
in
my
Carius.
They
don't
share
too
much
about
their
technology,
but
I
can
say
that
daily
George
was
the
founder
of
vicarious
used
to
work
for
Jeff
at
Numenta.
I.
A
Think
that
was
like
a
long
time
ago,
eight
years
ago,
something
like
that,
but
when
he,
but
he
left
to
form
vicarious
they're
using
more
standard
machine
learning
techniques
like
a
Bayesian
inference.
Bayesian
prediction
stuff
like
that:
we
don't
use
any
Bayesian
techniques
at
all.
We
focus
on
neurons
and
how
they're
interacting
so
that's
the
big
difference.
Numenta
is
focused
on
biology.
How
two
neurons
interact.
B
A
B
A
Of
fuzz,
all
that
out,
that's
a
big
black
box
and
you're.
Looking
at
more
of
the
population
effects,
you
know
in
the
box
what's
happening
and
that's
more,
the
approach
vicarious
is
taking.
We
want
to
understand
intelligence.
We
want
to
know
how
it
works.
We
don't
just
want
to
model
and
simulate
it,
which
I
said
I
think
there's
more
of
that
approach.
So
hopefully
that
answers
the
question.
I'm
also
we're
very
transparent,
open
about
everything
all
of
our
code
and
everything
but
curious
is
not
I
think
their
their
technology
is
proprietary.
A
Another
question
on
chat:
he
mentioned
about
recursive
cortical
network
and
there
in
the
recent
AGI,
oh
yeah,
delete,
yeah
I
saw
somebody
said
something
about
on
the
forum
about
that,
but
I
don't
know
anything
about
their
their
current
technology,
but
I,
don't
I,
but,
like
I,
said
what
I
just
said.
So
alright
I'm
gonna
sign
off
you
guys.
What
up
Danielle?
Do
you
guys
later
hit
me
up
on
forums,
bye,
okay,.