►
From YouTube: Numenta Research Meeting, Sept 4 2019 (partial version)
Description
We experienced an internet outage in the office during this video, a complete video is here: https://youtu.be/Hw6rA4Hq24k
A
A
No,
no
okay,
so
I
have
a
really
half-baked
idea
today
and
so
I
just
want
to
walk
through
it
to
the
goal
today
is
to
get
to
this
half-baked
I
didn't
want
to
call
that
Pappas's
such
a
well
on
my
part.
Is
this
last
idea
and
I:
don't
even
understand
it
myself
so,
but
I
think
this
was
just
trying
to
describe
it,
and
maybe
I
get
some
feedback
about
how
to
think
about
this.
A
So
it
is
related
to
it,
because
I
was
thinking
about
different
types
of
representations
and
the
coracle
positive.
That's
the
office
on
your
meanings
of
them,
but
not
specifically
this
stuff.
So
I'ma
start
with
it
like
a
review
of
things,
but
for
some
of
you
guys
won't
be
able
to
review
and,
like
so
many
new
things,
so
I've
just
seen
about
a
single
block
of
cortex,
just
generic
cortex
column.
A
If
you
will-
and
we
can
say-
okay,
what's
the
goal
of
a
block
of
cortex
a
column
for
taxi
the
way
I
talk
about
the
goal
that
thing
is
is
to
learn
in
a
model
of
the
world.
It
gets
some
input
from
some
place,
it
doesn't
if
its
genetic
or
it
doesn't,
that
input
represents
there's
no
idea
what
it
represents.
It
doesn't
know
it's
modality.
A
Also,
it
mostly
changes
through
movements,
so
we
learn
a
model
the
world
through
movement,
as
some
inputs
change
if
something's
moving
and
it's
trying
to
figure
out
the
structure
of
that
thing,
we
know
that
it's
a
predictive
model,
meaning
it
predicts
whether
it
was
the
next
input
and
therefore
it
must
have
some
copy
of
the
motor
command.
That's
the
Ephrem
EC
is
efference
copy,
that's
a
nice
term
that
was
copied
or
command,
so
it
must
know
something
about
the
movement.
Otherwise
it
could
predict.
What's
going
to
happen,
we.
D
A
And
it's
not
clear
where
that
shows
up
here
is
very
clear,
with
the
input
show
that
the
input
comes
from
something
and
goes
to
the
thalamus
in
this
context
of
a
fork
of
all
the
places,
but
this
green
being
the
border
efference
copy
is
kind
of
a
mystery
when
he
spent
a
lot
of
time.
Talking
about
that,
we
do
know
a
lot
about
the
output
of
this
thing
and
the
goal
the
goal.
A
This
comes
not
just
for
a
model
because,
what's
the
point
of
just
learning
model,
the
purpose
of
the
end
is
to
use
that
model
to
create
behaviors,
and
so
once
you
have
a
model
of
the
world,
not
only
can
predict
the
future,
we
could
say:
oh,
how
do
I
get
to
a
particular
state
of
that
model?
How
do
I
get
to
the
particular
position
in
space
and
actually
I'm
a
distinction
between
location
and
state,
because
things
are
stakes?
I
would
turn
around
my
phone.
My
cell
phone
as
a
state
I
want
to
get
some
state.
E
D
A
A
A
five
themselves
come
out
and
they
go
something
like
that
either
because
in
the
primary
sensory
cortex
now
that
would
imply
the
input
from
the
retina
is
a
motor
commands
and
we've
actually
so
I'm
thinking
about
that
and
couldn't
possibly
be
viewed
as
a
motor
command
and
so
I'm
getting
closer
to
that
problem
here.
But
I
didn't
want
to
dress
it
over
here
here.
There's.
A
A
A
D
A
C
D
A
A
lot
that
we,
this
thing
has
to
learn
a
model
of
space
that
this
thing
is
looking
at.
It
doesn't
know
what
it
doesn't
know.
What
doesn't
what
the
world
is
like
bunch
of
bits
coming
in,
and
here's
my
motor
fans
and
they're
changing
and
I
have
to
figure
out
what
is
the
underlying
structure
of
this
thing
that
I'm,
observing
and
a
model
of
it
and
we've
talked
a
lot
about
what
that
model
looks
like
the
model
can
be
built
on
locations
and
I'm,
not
going
to
go
further
than
that
today.
A
The
structure
here,
but
this
is
very
generic,
because
this
could
be
some
high-level
cortex
doing
language
and
not
getting
sensory
inputs
record.
So
this
is
the
general
idea
and
I
think
you
can
take
general
approach.
The
column
is
that
what
on
this
model
and
from
that
create
behaviors,
that
that
allow
you
to
get
to
something
good
reach,
a
goal
which
is
some
location,
some
state
in
the
world
and
of
course
you
have
lots
of
columns
donors.
We
have
the
input
to
the
column
here,
so
you
get
a
bunch
of
bits
coming
in
combos.
A
A
Some
part
of
the
cortex,
the
dimensionality
of
the
world
that
it
represents,
is
unknown.
For
example,
often
our
sensory
organs
are
two-dimensional
attention,
but
those
are
projections
on
three-dimensional
space,
so
my
vision
of
the
production
on
three
dimensional
space,
so
I
figure,
but
we
know
in
general
we
can't
say
anything
about
I,
don't
know
the
promises
of
the
three-dimensional
space
and
repeated
by
other
objections
or
consumed
with
n
dimensional
space
in
some
n
dimensional
projection
on
it.
Maybe.
A
A
A
Muscles
directly
they're
always
they're,
always
projecting
some
other
area
that
drives
the
muscles
and
that's
from
the
cortex.
We
we
can
say
that
it's
going
to
be
another
sparse,
distributed
representation
because
everything
that
rains
as
far
as
the
duggars
out
representation,
we
can
say
that
I
always
said
I
believe
that
that
these
cells
don't
know
what
the
projected
term
they
don't
know
how
to
do
this.
They
have
to
learn
how
to
control
the
subcortical
thing.
So
this
is
a
generic
learning
system.
A
Ok,
I
haven't
said
nothing
about
how
to
socially
link
to
these
other
cells,
someplace
else
in
making
motor
behavior
so
that
I
can
control
them.
So
that
has
to
be
learned.
This
is
not
like.
Oh
this
cell
goes
to
this
and
it's
logo-
staffing,
I,
just
punch
a
self-projected,
some
other
bunch
of
cells
and
somehow
have
to
learn
to
control
them.
A
So
in
general,
every
region,
the
cortex,
where
I
do
not
be
self
conjunctive,
some
subcortical
molar
area.
That,
then,
does
the
actual
behaviour
and
one
of
the
questions
that
I've
always
worried
about
or
thought
about
is,
and
we
don't
really
have
an
answer
to
this
is
what
is
the
form
of
this
motor
SD
or
how
I
think
about
that
I?
Don't
way
of
thinking
about
the
info
representing
X,
there.
A
Have
to
control
something,
but
what's
how
I
understand
them,
how
it's
like?
What
are
these
bids
mean?
What
are
the
other
five
styles
represent
inspected
me,
so
that
was
like
what
is
the
form
of
the
motor
STR,
but
did
the
bits
these
l5
cells
mean,
can
I
put
some
kind
of
meaning
on
and
then
how
do
they
associate
linked
to
the
sub
part
of
the
motors?
It
said
this
whole
link
here
was
kind
of
a
this
representation
and
link
was
much
poor
form
idea.
A
A
Okay,
so
this
is,
and
now
how
do
I
incorporate
it,
it
would
be
different
and,
and
now
I
have
at
least
a
way
of
thinking
about,
like
five
cells
to
say,
yeah,
there
are
subsets
of
many
columns,
they
represent
what
that
slice
is
doing,
which
one
and
that's
a
good
representation
of
a
motivator
movements
orthogonal
to
whatever
features
on
detecting
it's
kind
of
fuzzy.
My
row
of
mine,
so
I.
A
E
Yeah
I'm
from
a
given
cortical
columns
perspective.
It
just
has
a
little
bit
of
influence
and
so
to
do
prediction,
it
really
needs
to
know
what
actually
happened
and
it's
not
just
stuff
cortical
motor
structures.
It's
also
all
of
the
other
cortical
columns
that
are
controlling
things
yeah,
you
know.
So
this
is
before
or
they'll
call
it
v1
v2,
you
know
all
sorts.
A
A
Maybe
yeah
purpose.
If
you
want
to
talk
about
multiple
columns
here,
because
if
you
think
about
what
would
comes
in
a
hierarchy
tree,
you
want
to
be
before
I
met
come
up
with
something
if
I
think
about
multiple
columns
as
not
hierarchical,
but
parallel
with
different.
You
know
the
thousand
branch
apotheosis
they're
parallel
at
some
point
and
representing
different
slices,
the
size
of
the
input
space
I
might
come
up
with
a
different
thing
and
there's
a
huge.
We
still
don't
understand
how
layer
five
cells,
how
do
they
become
the
input
to
the
next
column,
but.
E
A
So,
let's
be
careful
so
you're
right,
so
a
similar!
This
is
one
of
the
mysteries
of
movements.
So
if
you
single-column
like
remember
saying
what
tipped
my
finger
and
it
could
represent
the
motion
that
finger
in
its
space,
but
then,
when
I
actually
do
something,
this
whole
complex,
arbitrary
changes.
I
was
avoiding
that
question
at
the
moment.
I'm
trying
to
tackle
one
piece
of
this
problem,
just
one
piece
as
a
way
of
maybe
performing
a
new
basis
for
figuring
out
the
bigger
problem
like
if
I
could
positives
was
right
and
I'm.
C
F
C
A
It
be
that
way.
Well,
it
could
be
the
art
of
representing
behaviors
that
could
be
the
representation
of
movement
and
the
whole
system
is
centric
motor,
so
I
have
sensory,
and
now
that
I
have
a
SDR
bird,
it
represents
that
sensory
input
under
this
motor
position
motor
command.
The
hypothesis
gives
me
new
way
to
think
about
later
five
new
way,
thinking
about
various
things,
and
now
maybe
we
can
come
back
later
adjustable
why
delay
of
five
cells
projections
of
found
those,
but
why
do
they
connect
to
this
way?
A
C
E
C
A
E
E
A
The
whole
theory,
the
thousand
brains
they're
in
the
Colin
theory,
says
the
world
is
a
complex
structure
and
you
cannot
observe
it
all
at
once
and
therefore
to
learn
about
a
wall.
You
have
to
move
through
the
world
and
if
what
you're
doing
is
feed-forward
of
stuff
you're,
not
really
a
model
classification
or
pattern
classification,
but
to
learn
the
structure
of
the
world
and
not
move
through
the
world,
you
have
to
move.
A
It's
like
I
can't
learn
a
model
of
Redwood
City.
If,
by
sitting
in
this
room,
it's
impossible,
I
have
to
walk
around
the
city.
So
the
same
thing
with
a
cab.
All
I
see
these
things
flying
past
me
I'm.
How
am
I
gonna
know.
What's
going
on,
it's
like
I
can't
associate
my
behaviors
with
you
know:
I
can't
map
out
my
input
on
some
metric
space.
So
we
need
to
have
that.
It's
I,
just
I
thought
there's
something
would
really.
A
And
it's
basically
sort
of
integrating.
If
we
think
about
those
motion,
sensitive
cells,
not
as
like
the
world
is
moving
on.
It's
almost
the
time.
It's
mostly
because
I'm
moving,
then
those
are
sensory,
motor
representations
and
that
and
then
it's
intriguing
to
think
like.
Why
would
I
want
to
represent
motion
as
orthogonal
movements,
so
the
primary
slices
I've
decided
it's
like
it's
like
it
somehow.
A
D
A
A
C
A
A
A
B
I
drop
one
interpretation
of
please
just
this
picture
of
moving
orthogonal
to
a
plane.
I
just
want
to
know.
Please
do
there's
one
easy
way
to
dry
and
I
I'm
5050
I,
whether
this
is
what
you
mean
so
like
first,
just
like
we're
talking
about
like
a
single
mini
column,
and
you
like
some
set
of
input
bits
so
and
let's
say
this
particular
mini
column,
say
this
feature
just
to
keep
this.
It's
gonna
be
these
two
minutes
it
it
likes
when.
B
C
A
B
A
But
I'm
gonna
think
about
it.
If
you
were
not
orthogonal,
if
you
were,
if
you're
coplanar,
then
changes
would
not
affect
those
two
bits,
so
other
bits
can
be
changing,
but
those
two
bits
are
staying
on,
so
you
say:
I'm
still
active,
but
you
don't
think
I'm
ignoring
so,
but
if
they're,
both
if
they're
changing
that
means
I'm
no
longer
in
the
same
plane,
I
mean
some
hotels
right
anything
if
they
both
changed
and
I
must
be
moving
orthogonal
to
the
plane
right.
A
A
Yeah
yeah
so
I'm,
looking
at
50
bits
and
so
30
20
bits
and
say
I
respond
to
this
thing.
Now.
If
the
world
changes
the
other
picture,
changing
them,
I'm
still
seeing
this
I'm
still
active,
then
I
would
say
that
there's
been
no
movement
but
all
of
a
sudden,
all
my
haven't
exchanged.
Well
then,
it's
not
all
my
chains
in
a
job
type
of
movement,
saying
I,
think
I
get
difficulty
when
I
think
about
the
plane
and
where
is
but
we
could
talk
about
bits.
Changing
and
I
would
say.
If
my.
A
B
A
Not
sure
I
don't
know,
maybe
I
shouldn't
use
the
word
plane
I
mean
that's
my
tendency
to
think
about
physical
planes
that
are
slicing.
Something
3-dimension.
If
you
think
of
a
plane
is
a,
is
a
lower
dimensional
projection,
a
higher
dimensional
space,
so
some
dimensions.
If
you
change
in
some
dimensions
it
doesn't
your
might
my
point
the
plane.
What
I
was
thinking
about
this,
but
maybe
I'll
be
using
it
incorrectly.
I
was
thinking
about
this
mini-com
represents
a
plane.
What.
A
The
same
mini
column
responds
exactly
the
same
where
some
of
the
bits
that
it's
not
paying
attention
could
change.
It
doesn't
affect
it,
but
if
the
bit
Center-
and
so
that's
like
those
are
those
are
movements
orthogonal
to
the
plane
and
orthogonal
to
the
representation-
and
he
goes
any
changes
from
these
bits
here-
don't
change
this
mini
column
and
and
then
you
say:
okay
well,
those
are
those
are
orthogonal
movements,
but
but
yeah.
A
A
F
A
F
C
A
Motion
sensitive
complex
cells
can
I,
say
this
motion
at
that
point.
I
could
one
moment
to
give
you
the
line
of
the
left
later,
give
you
the
perfect
line
of
the
right
and
the
same
cells
can
respond.
So
it
doesn't
there's
no
implied
motion
at
that
point.
There
is
applied
equivalence.
Okay
right!
That's
super
nice
thing!
That's
like
your!
If
I
get
shot
your
max
pooling
operation
right,
it's
just
a
temporal
cooling
operate
it
just
simple
template
operation
says
these
are
equivalent.
F
A
F
A
E
E
C
E
E
A
E
A
A
B
A
F
A
B
A
B
A
A
No,
no!
It's
built
into
the
system.
Everything
is
this
spatial
pool
or
a
double
boiler,
or
basically
the
whole
idea
that
you
kind
of
pull
them
out.
I
would
say
them
the
20
bits
on
a
dendritic
on
the
dendritic
Sango
aren't
doing
that.
That's
something
different
they're,
not
pooling,
multiple
as
they
are.
They
essentially
doing
spatial
pooling.
If
you
want
to
consider
it
that
way,
yeah.
C
A
Wasn't
a
fire
temple
for
typically
means
that
that
one
STR
and
another
STR
that
occurs
over
time,
which
are
completely
orthogonal
STRs,
they
don't
know
overlaps
bits,
get
converted
into
the
same
output
and
the
dendritic
segments.
We
don't
have
them
doing
that.
The
dendritic
segments
say
I
just
recognize
a
single
pattern
and,
and
they
don't
they
don't
pull
on
that.
Second
I,
don't
the
same
dungeon
segment
doesn't
moment
later
represent
completely
different
STR.
A
E
A
A
Any
way
we
can-
and
here
are
many
guys
have
a
lot
to
do
really.
The
thing
I
would
need
help
with
is
really
coming
up
with
a
visual
metaphor
and
the
use
of
the
term
for
planes
or
slices
or
you
know,
is
it
a
vector
these
are.
This
is
where
I'm
confuse
I'm
thinking
about
playing
to
something
else,
and
so
maybe
that
would
be
helpful
on
that.