►
From YouTube: Numenta Research Meeting - Apr 22, 2019
Description
Broadcasted live on Twitch -- Watch live at https://www.twitch.tv/rhyolight_
B
Marvy
found
this
article
that
he
was
reading
and
he
presented
it
last
week
we
discussed
it,
I
think
Thursday
and
Friday
yeah
and
there's
like
interesting
ideas
in
it
and
so
I
think
it's
good
to
go
and
try
to
review
it
again,
really
quickly
what
it
is
and
and
then
we
continue
from
there.
So
you
want
to
do
that.
B
A
A
Yeah,
so
this
I
saw
it
a
while
ago,
but
last
week,
I.
Finally
like
understood,
update
rules
and
like
how
to
update
orientation
under
the
theory,
and
it's
actually
really
nice
in
it,
and
it
it
poses
head
Direction
cells
differently.
Then
then,
we've
often
talked
about
them,
so
a
head,
Direction
selves
I
mean
as
an
animal
if
the
animals
on
a
like
on
a
flat
surface,
unlike
if
they're,
if
they're,
basically,
if
they're
coplanar
with
gravity
or
coplanar,
with
like
the
perpendicular
to
gravity
directions,
cells
of
course
behave
in
like
this
way.
A
A
Just
do
this
last
week,
I'll
just
use
that
picture
where
one
way
it
worked
is
like.
If
the
animals
facing
like
to
decide
what
direction
Dukes
house
to,
is
they
just
like
projected
down
to
the
to
the
like
the
earth
plane
and
represent
that
angle
of
a
and
as
the
animal
rotates
around
here,
this
rotates
around
here,
but
one
complication
with
is
once
once
the
animal
is
not
coplanar
with
earth.
I'll
call
it
the
diversity.
A
B
B
Just
like
we
imagined
I
could
get
impaler's
Mirko,
and
so
that's
one
way
of
doing
it,
and
this
paper
proposes
a
very,
very
different
way
of
doing
it,
and
that
is
that
is
the
boy
doing
mu
cevaz
last
time
is
that
you
just
have
2d
orientation
cells,
but
there
is
another
thing
called
a
reference
vector
which,
in
this
case
of
the
wrap,
would
be
private.
So
if
you
know
the
orientation
of
gravity
to
your
body-
and
you
know
your
2d
orientation-
that
sufficient
Sarno
your
3d
orientation
yeah.
A
A
B
In
many
cases,
this
gravity
wouldn't
it
well,
we
got
into
the
case
of
that.
The
head,
Direction
orientation
circle
was
that
was
the
line
with
the
you
say,
we
don't
they're
still
yeah
like
it's
like
if
you're
climbing
up
in
a
while
I
could
join
mine
straight
in
the
face.
If
the
rat
is
like
climbing
up
the
states
and
rotating
around
like
this
it'll.
C
A
C
B
B
A
Other
set
of
possible
changes
in
orientation
is
to
rotate
this
plane
in
various
directions
and
the
update
rule
is
okay,
as
this
plane
the
plane
that
the
rad
is
on
rotates
relative
to
the
gravity
vector
the
reference
vector
if
it
rotates
into
the
vector
like
if
it
rotates
toward
the
head.
Direction
cells,
don't
update
that
has
no
effect
on
the
head,
Direction
cells.
So
if
the
rat
walks.
B
A
A
So
if
the
rat-like
sidles
around
this
corner,
that's
a
90
degree
rotation
that
causes
a
head.
Direction
cells
to
rotate
by
90
degrees,
technically
I
think
maybe
that'll
be
counterclockwise,
but
it
would
rotate
by
90
degrees-
and
this,
like
these,
just
very
simple
update
rules
of
just
update
the
head.
A
Direction
cells
linearly,
both
rotations
on
the
plane
and
rotations
of
the
plane
about
the
gravity
axis,
give
you
this
path:
integration,
property
with
orientation
where,
where
any
path
to
any
point
on
this
cube
or
any
point
on
the
sphere,
okay,
something
I
don't
know
yet
is
how
the
upside
down
is
handled.
I,
don't
I,
don't
have
that
intuition,
yet
I
don't
know
how
well
handles
there
are
some.
A
So
but
yeah
I
mean
I,
could
just
state
like
the
confusion
about
the
upside
down,
is
like
I'm,
imagine
walking,
okay,
the
rat
tilting.
What's
this
way
keeps
walking
straight?
What's
this
way
keeps
walking
straight
Direction
cells
don't
update
because
it's
tilting
into
the
gravity
axis
it
walks
around.
The
bottom
keeps
walking
straight.
That
would
suggest
that
that
that
that
these
are
the
two
direction
north
directions,
so
that
I
could
make
a
case
that
okay
on
the
bottom,
this
one
should
be
active
when
you're
playing
this
well
or
I
could
say.
B
A
A
So
if
the
rat
is
like
walking
like
I'm
like
here,
then
I'm
here
and
I'm
here,
if
the
rat
swap
is
walking
down
this
direction,
it's
silence
down
to
the
bottom,
and
now
it's
facing
that
way
and
now
I've
just
made
an
argument
that
the
cells
should
be
active
when
there
are
facing
this
way
at
the
bottom.
So
I
commit
I
can
make
arguments
for
both
directions
and
I.
Don't
know
how
to
how
the.
B
B
B
B
B
A
A
Logical
just
first
thing
to
reach
or
first
first
thing
that
comes
to
mind
and
I
had
no
idea
how
to
draw
a
sphere
on
a
whiteboard,
but
basically
not
a
2d,
Rainey's
idea.
Thomas.
Imagine
a
ring.
That's
like
the
surface
of
a
sphere
or
like
there's
a
sound
here.
That's
out
here,
there's
a
year
etc,
and
this
represents
like
the
direction
of
the
of
the
gravity
vector
well.
How's
that
ring
you
just
seen
is
a
spiced
outer
ring
it's
this
sort
of
like
a
ring
in
that,
like.
A
So
I
mean
this
is
like
a
2d
located.
This
is
a
point
on
the
sphere,
so
as
2d-
and
this
is
the
third
ya
dimension,
you
could
also
have
I
mean
like
this
is
the
this
is
the
2d
version
or
you
could
do
like
the
latitude
and
longitude?
Essentially,
those
are.
Those
are
a
little
weird
moogly,
because
it's
like
it's
like
a
ring
and
a
half
a
latitude
and
a
longitude,
you
don't
yeah
yeah.
A
B
A
I
said
posted
a
like
and
then
otherwise
like
a
third
option,
that
I
don't
know
if
I'm
the
third
option,
that
I'm
pondering,
if
it's
worth
talking
about
something
that
comes
up
in
robotics,
a
lot,
a
way
that
that
that
3d
orientations
get
represented
in
robotics
as
through
quaternions,
which
are
these
four
dimensional
numbers
that
make
everything
work
really
nicely.
I'll
come
back
to
you.
A
If
that's
worth
our
attention
right
now,
it
may
not
be
that
it
may
be
the
case
that
we
don't
need
to
think
about
those
other
than
when
we're
when
we're
figuring
out
how
these
cells
all
need
to
be
interconnected,
but
it
may
not
be
necessary
to
talk
about
those,
but
anyway,
the
point
is
there
is
another
org
and
I'll
just
a
quarter.
Nians
is
another
thing,
but
I
don't
know.
If
we'll
actually
need
to
talk
about
that,
so
there
are
a
few
options.
This
is
what
I'm
thinking
about
right
now
is
exactly
like.
B
Each
others,
along
with
motor
commands,
so
that
they
okay,
where
does
this
money,
come
from
well
structure?
Doing
that
yeah
they
just
you
know,
I
keep
pointing
to
things
like
okay,
I'm,
trying
to
structure.
We
know
about
the
court
in
cortical
column
and
that
may
not
be
right,
but
I
want
to
try
it
well.
D
B
B
B
The
same
structure,
and
so
maybe
they're
both
like
little
modules
or
module
of
things
where
you
get
a
unique
representation,
and
so
maybe
one
way
cigarette,
maybe
though
maybe
one
of
them
is
one
is
like
the
2d
gravity
vector
and
the
other
is
3e
linear,
vector
of
itself
and
then
maybe
orientation
slices
through
all
thing,
thankee
thankee
I
keep
coming
back.
An
idealist
is
trying
to
be
crazy.
B
They've
got
this
column,
you
got
these
mini
columns
that
seem
to
have
some
sort
of
like
you
know,
orientation,
look
to
them
and
we
don't
know
why
they
goes
all
the
columns
of
all
the
layers.
We
don't
know
why
they
go
through
all
the
layers,
there's
some
cells
and
all
the
layers
that
look
like
orientation
every
lady.
Not
all
the
cells
is
the
subset
of
themselves,
but
in
all
the
layers,
and
then
we
have
these
two
things
which
look
like
you
know
like
they'd,
be
good
places,
for
you
know
like
a
grid
cell
motto.
B
B
B
D
B
B
But
I
also
have
to
have
the
same
mechanism
we're
going
to
work
for
conceptual
objects
where
it's
you
know,
high-level
flaws,
so
there's
got
to
be
more
of
a
pure
mechanism
for
doing
this,
and
now
I
was
thinking
like
well.
You
know
mention
I
do
I.
Do
I
establish
a
vector
or
preferred
vector
for
this
cop,
because
you
know
I
mean
I
pick
the
long
dimension.
B
B
C
C
In
and
make
predictions,
and
if
you
got
that
would
give
it
the
right
input
it
like
gravity
and
so
on.
It
will
learn
this
particular
mechanism.
If
you
give
it,
some
other
system
would
learn
to
use
that
to
do
the
transformation
at
the
end
of
it.
You
want
a
generic
system
that
can
learn
transformations,
and
this
is
an
example
of
a
simple
transfer
me
much
more
complex
transformation
for
you
to
do
for
doing.
Mathematics.
B
B
C
B
B
This
is
your
input
and
and
then
we
have
this
other
one
that
beats,
and
so
then
there's
this
the
here
so
say
six
be
six
day
and
there's
five
here,
and
we
know
that
this
there's
this
other
input
that
comes
in
right
around
the
border
between
five
and
six
and
it
seems
to
be
everywhere
and
and
so
in
the
past
you
might
remember
I
wish
I
was
trying
to
think
of
what.
Why
did
you
have?
What
purpose
and
I
I
speculate
once
before
that
this
might
be?
Establishing
orientation?
I
talked
about
that
it
might
be.
B
Establishing
orientation
and
I
could
I
need
to
know
my
orientation
but
make
any
sense
of
it
was.
It
was
a
tunable
idea,
like
I,
have
essentially
input
but
I
need
to
know
my
orientation
and
that's
that
I
always
talked
about
maybe
like
there
was
bigger
cubes
I
said
this
might
have
a
broader,
a
broader
receptive
field
so
that
you
know
larger
area
and
then
help
you
define
your
orientation.
That
was
a
very
poorly
defined
idea,
but
if
I
am
or
specific
version,
that
would
be
saying:
oh
no,
it's
not
real
imitation.
B
B
B
So
I
know
I
just
don't
mean
this
out
again
because
again,
I'm
gonna
stick
trying
to
fit
this
world
into
this
world.
I'm
thinking
like
well,
if
I
can
rely
camera
line,
vestibular
I
can't
I
mean
oh
yeah
I,
don't
what
is
going
to
be
sensory,
something
sensory
that
I
and
I
don't
think
it's
gonna
be
this
input,
because
this
is
the
thing
I'm
trying
to
predict
so
there'll
be
another
place.
B
A
A
B
B
D
B
We
have
a
comic
operating
on
its
own,
or
maybe
not
maybe
you
know,
there's
also
the
fact
that
the
later
five
has
a
long
range.
There's
long
range
connections
in
later
five
and
remember,
there's
two
sets
of
one
reads
connection
and
stay
free
and
they
are
five
and
maybe
one
way
to
think
about
those
long
range
connections
would
like
by
devoting
on
rafters.
B
Maybe
to
me
this
is
a
great
because
the
market
presented
this
I
said:
oh,
it's
real
high
over
L
just
been
stuck
on
this
idea
that
we're
going
to
represent
3d
orientation
with
a
bunch
of
orientation
modules
and
that
they're
all
being
a
projection
on
the
3d
Orion
space
radiation
studies-
and
this
is
just
otherwise
another
thing
that
came
up
last
week,
which
I
want
to
mention.
So
you
can
take
private
sua
time
is.
D
B
I
was
pointing
to
my
keys
last
week:
that's
not
what
it
looked
like
in
the
we
don't
see
any
place
where
he
would
see
all
these
different
modules
within
a
column
in
invest.
I
can
look
at.
This
is
my.
This
is
my
recreation
of
the
tank
drawing
the
tank
paper
where
in
amount
of
supplies
with
cortical
column,
he
was
showing
in
the
courses
into
one
the
cortex
and.
B
And
it's
not
like,
where
would
I
have
multiple
grid
cell
modules
in
a
single
quarter
of
a
cop,
and
so
the
question
then
becomes
well.
Maybe
there
is,
maybe
you
know,
is
the
same
sort
of
check.
That's
being
done
here.
We
have
this
reference
vector.
Could
something
could
I
have
a
reference
vector
to
this
would
make
this
work?
I
haven't
even
thought
about
that.
Just
like
that's
another,
so
going
on
from
the
one
extreme,
we
have
multiple
grid
cell
models,
but
there
are
different
intersections
of
the
different
projections
of
the
3d
space.
B
Orientation
models:
we
don't
issue
modular
reference
vector
and
maybe
the
same
thing
can
happen
here.
You
know
there
is
no
module,
have
a
coolant
reference
vectors
on
how
it
meets
three
eat.
That's
very,
very
high-level
concept:
I
have
no
idea
how
that
would
work,
but
I
just
think
we
don't
see
multiple
law
of
these
guys.
B
A
I
guess
the
other
thing
that
comes
to
mind.
Something
you've
pointed
out
that
yeah,
so
the
current
state
of
this
of
all
these
cells
is
essentially
a
point
on
it.
Promise
like
the
manifold
of
the
cell
activity
here,
is
like
it's
moving
around
around
this
and
wrapping
around.
However,
as
you
pointed
out
that
there's
also
this
like
this
notion
of
like
a
mist
you'll,
sometimes
sometimes
one
of
these
will
be
higher
firing,
yeah,
and
so
that
can
be
thought
of
is
like.
Suddenly,
this
manifold
of
activity
is
becoming
three
and
really
that
came.
A
Second,
that's
not
necessarily
just
3ds,
because
I'm
enough
variation,
just
like
that
pit
and
then
a
fold
of
activity
could
like
going
some
and
these
other
strange
directions
through
allowing
these
to
be
a
different
magnitudes,
so
that
there's
some
flexibility
there,
but
I
don't
know
what
to
do
with
it
and
I.
Don't
know
how
to
give
it
like
this
path:
integration,
property,
like
how
do
you?
How
do
you
take
these
missed
bumps,
and
these
varying
rates
and
encode
something.
B
What
is
reference
vector
was
like
a
key
thing
in
all
of
us.
What,
if
excuse
me
if
I
already
said
this
I'm
sorry
I
can't
remember
again
going
back
to
the
many
columns
all
right.
Is
it
possible
that
the
many
columns
are
including
reference
vectors
and
that
that's
the
thing
that's
passing
through
everything
and
then,
in
that
regard,
to
my
sister
grams?
This
is
orientation
and
this
is
grid.
So
then
you
have
this
with
expensive
all
like
what
would
I
expect
to
see
here
right.
B
What
I'm
thinking
of
myself,
if
I
change
the
speed
at
which
you
I'm
asking
the
question?
I
just
showed
you
a
picture
and
there
was
just
some
line
it's
morning.
You
know
some
grading
right.
Well,
it
might
just
look
like
it's
just
the
orientation,
but
if
IX
you're
shooting
a
real
picture
of
something
or
a
real
object,
you're
looking
at
an
object
is
changing
in
different
directions.
Would
these
cells
behave
differently,
given
this
model
of
the
rough
inspector.
B
B
So
I
wonder
if
actually
there's
even
more
going
on
here
than
we
knew
and
that
it's
more
like
what
you
just
said.
Maybe
they
seem
to
really
be
changing
in
other
ways
too,
that
we
we
didn't
understand
and
because
you
know,
even
though
our
model
would
we
say
yes,
there's
a
subset
and
this
cell.
Now
it's
going
to
fire
a
specific
location
of
the
object
or
some
specific
location
in
video
or
something
like
that,
they're,
not
reliable.
In
that
regard,
they
seem
to
be
there's
still
weird
wonky.
It's
sort.
B
I
remember
seeing
all
the
points
I'm
seeing
that's
quite
right
to
either
like
doesn't
always
seem
to
do
it
at
the
same
time.
I,
don't
know
why
I'm
just
wondering
where
and
then
this
is
this
appears
to
be
just
a
one-dimensional
encoding.
Then
you
need
a
two
dimensional
encoding
for
the
reference
vector.
B
B
B
It's
almost
exactly
the
same
way,
and
so
with
my
ingredients
would
look
like,
like
they're,
already
corrected
some
plane,
but
as
the
plane
moves
around
the
array
to
the
plane
rotates,
then
they
you
know
so
so
I
would
somehow
not
differently.
You
know
just
the
same
way
like
in
cases
all
I
see
the
point
of
it
being
an
open
space
versus
a
closed
space
back.
B
If
I,
if
I,
had
multiple,
if
I
had
multiple
rings
like
this
and
one
vector
but
I,
had
multiple
grid
cell
on
orientation,
cell
modules
and
each
orientation
cell,
what
we
wanted
to
do
is
the
each
orientation
Salman
do
they
do
were
path.
Integration
means
when
I'm
in
the
same
spot,
on
the
same
orientation
on
that
at
the
same
celebrate.
So
if
I,
even
if
I
had
a
bunch
of
these
and
they
were-
and
they
were
anchored
differently,
so
they
all
had
that
same
behavior,
but
they're
anchored
differently.
B
Then
I
would
have
entered,
am
possible,
unique
orientations
right
so
that
I'm
just
saying
could
I
have
in
a
representation.
Can
I
still
have
a
representation
of
orientation
that
you
need
to
the
object?
Yes,
it'll
be
the
object.
It
won't
be
unique
to
the
location.
Yes,
okay,
that's
that's!
Okay,
at
the
moment,.
B
B
A
B
B
B
I'm
wondering
if
I'm
just
if
I
took
tanks
I
did
here
then
I
wouldn't
even
have
unique
locations
on
III
couldn't
Theory
have
the
same.
Encoding
I
would
have
the
same
encoding
at
different
places.
This
wouldn't
be
unique.
This
is
just
one
module
and
it's
it
doesn't
it's
not
unique,
so
it
would
be
it's.
It's
not
unique.
B
Object
and
it's
not
even
really
unique
to
the
location,
to
location
on
the
object.
It's
for
most
cases.
It
would
be
sufficient,
but
not
at
all.
Then
we
have.
We
have
these
three
things
and
the
question
is
that
these
phrase
things
together
sufficient
to
do
everything
we
need
like,
for
example,
I,
have
some
location,
but
it's
not
unique
to
the
object,
but
I
do
online
orientation
which
is
unique
to
the
object.
So
it's
combine
those
two
its
community
object-
and
this
is.
B
B
B
B
Yeah
so
I'm
thinking
like
multiple
orientation
modules
and
I
got
multiple
good
saw
modules
and
I
have
reference
vectors
slicing
down
through
them
in
the
intersection
these
the
intersection
where
I
have
to
resolve
the
same
thing
has
to
resolve
in
here
and
here,
using
the
same
reference
that
there
is
is
going
to
guarantee
that
these
guys
are
in
cahoots
and
everybody's
everything.
I
need
the
purpose
that
is
being
represented.
B
C
C
B
But
if
I'm
trying
to
fruit
apple
have
a
dendrite,
they
said:
okay,
I'm
going
to
respond
when
variable
is
to
environment.
B
is
a
fix,
and
these
are
two
separate
cell
populations
in
different
different
places
and
I
make
sure
this
mess.
It's
just
not
going
to
have
the
equally
sample
both
of
them.
That
is
where
I
get
nervous.
I,
don't
like
that
I'm
just
putting
out
the
best.
That's
one
tool
that
okay.
B
D
B
C
B
B
It's
like
there's
some
undistributing.
This
is
another
bus,
intersecting,
these
guys
and
then
somehow
seems
to
be
doing
computational
parsing
or
something.
What
was
like
it's
hard
to
explain
why
you
do
it
otherwise
I
mean
a
while
there's
a
sample
a
set
of
cell,
like
you
just
have
the
set
of
cells
appear
connections.
No.
This
is
like
forcing
a
computational
structure
on
the
thing.
That's
why
I
keep
coming
back
to
it.
It
may
be
wrong,
but.
B
B
B
B
Like
a
teaching,
but
really
maybe
it's
something
differently,
like
the
reference
maker,
which
kind
of
retrospective
might
look
like
orientation
under
many
situations
and
then
I'm
thinking
like
what
would
they
then?
What
would
my
resale
response
properties
be
up
here?
If
I
sell?
No,
it's
not
orientation
thinking
here
and
then
maybe
we
think
about
that.
More
I
can
come
up
with
some
early
because.
B
B
B
A
A
So
I
mean,
but
this
orientation
good
Paul
I
wasn't
clear
about
that.
So
that
I
mean
this.
This
cell
population
represents
a
3d
orientation
and
it
is,
it
has
a
set
of
update
rules
as
a
procedure
motor
commands,
it
updates
itself
and
just
what's
the.
What
is
the
logical
way
to
break
that
and
that's
all
population
is.
D
E
D
A
B
B
B
So
so
my
point
is
that
I'm,
a
column
I
can
only
see
certain
part
of
the
world.
So
I
have
this
this
larger
perimeter
and
have
a
smaller
prick.
The
lord
of
rimini,
this
input
a
smoke,
and
so
so,
like
I,
can't
see
everything.
I
can't
know
when
the
true
extent
of
the
object
is
maybe
I
can
go,
I
don't
know,
but
the
question
what
would
happen
is
only
I.
This
was
a
somewhat
a
local
computation.
B
That's
my
reference
vector
could
be
a
local
computation,
volu
tons
of
some
broader
area
of
which
I'm
touching
or
something
but
may
not
be.
Global's
adopted.
We've
done
with
that
problem.
I
know
in
the
wild
cortex
we
think
of
grid
cells,
as
always
each
other
in
the
same
orientation
anywhere
in
the
room.
So
in
that
case,
maybe
through
entire
room
has
the
same
thing
but
I'm
a
finger
touching
an
object.
That
may
not
be
the
case.
B
B
Okay,
so
we
want
to
learn
it,
we
got
it
and
not
rely
on
the
stupid
letter.
We
have
a
good
line,
some
sensory
input.
You
might
have
voting
about
it.
We
might
have
voting
when
they
find
the
vector
I
could
be
what's
going
on
the
live
voting
which
we
haven't
done.
You
haven't
really
had
a
need
for
yet
you
know
they
said
why
we're
voting
ahead
of
time.
D
E
All
right,
so
you
guys
got
a
snapshot
of
that
hopefully,
and
that
is
the
end
of
the
research
meeting.
Thanks
for
joining
I'm
gonna
go
ahead
and
cut
my
cut.
My
stream
and
I'll
be
back
at
one
o'clock
to
talk
about
responsible,
AI,
license
and
check
out
my
events
calendar
on
my
twitch
stream.
It's
twitch.tv,
slash,
rhyolite,
underscore
slash
events
to
see
exactly
what
I'm
gonna
be
working
on.
The
rest
of
this
week.
Take
care.