►
From YouTube: Numenta Research Meeting - Dec 9, 2019
Description
Jeff Hawkins will lead this research meeting. Topic is dimensionality in grid cells and the thousand brains theory of intelligence.
Discussion and photos of the white board at https://discourse.numenta.org/t/numenta-research-meeting-dec-9-2019/6898
A
A
But
it's
a
potentially
very,
very
interesting
idea.
So
I
go
through
one
two,
three
four
I'm
going
to
start
off
with
some
basic
background,
mister
mind
some
stuff,
but
now
have
supposed.
He
proposed
a
mini
comp
with
even
a
replication.
My
comp
with
mini
con
and
I
thought
he
came
from
that.
There's
a
section
of
cortex,
oh
she's,
progenitor
cells
that
that
during
the
moment
they
grow
and
they
send
out
a
series
of
neurons
that
populate
every
finger
of
a
column
and
is
about
120
cells.
A
That's
reported
each
of
those
mini
columns
and
they
contain
all
optional
types
of
neurons.
So
that
is
that
there
are
multiple
types
of
many
columns
that
each
mini
common
is
a
complete
set
of
cell
types.
If
you
just
do
the
basic
math,
it
would
be
about.
150
million
of
these
mini
compliment
your
cortex
and
then
need
to
find
these.
That's
the
replicated
a
replication
that
the
function,
one
minute
is
a
larger
column,
and
that
is
not
a
physical
thing
that
may
comes
of
physical,
but
the
function
is
sort
of
a
thing.
That's
build
up.
A
A
So
there's
multiple
in
this
larger,
calm
everything
everybody
comment
like
a
commis
potentially
connected
anything
down
here
when
you
go
to
the
next
larger
column,
that's
representing
a
different
area
on
the
sensory
right,
and
so
he
showed
these
papers
that,
as
you
as
you,
move
a
cloth
that
needs
meaning
through
the
cortex
you.
You
know
it's
like
all
right
here
that
it
jumps
all
over
here.
We
reviewed
those
papers.
Okay,
I'm,
going
to
talk
about
the
miniconjou
today
is
key
proposals:
we've
never
eat
for
golf
partial
legislation.
A
B
A
About
50
percent,
not
everyone
may
become
chair,
receptor
field
properties,
and
so,
as
you
go
across
all
the
layers,
there's
a
commonality
there.
Now
we
think
about
it.
Actually
often
I've
been
in
the
snake
thing
about
all
things:
spatial
orientation,
what
they
actually
found,
but
they
reported
that
mostly
cells
are
due
rationally
century
the
vast
majority,
the
cells
become
our
connection,
steps
of
the
Carol
which
the
center
movement
is
occurring
and
their
greater
than
the
response
in
that
regard,
I've
always.
A
A
Define
receptive
field,
the
unbelievers
is
smaller
than
the
lower
legs.
You
look
at
what
what
the
cell
and
upper
layers
respond
to?
They
can
be
shorter
and
smaller
items
and
plural,
a
so
much
bigger.
That's
another
confirmation
that
was
pulled
out.
So
that's
in
the
background
information.
Now
we
will
come
over
here.
We
propose
the
thousand
theories
of.
C
A
The
largest
nominees
were
called,
we've
learned,
a
comma,
but
then
call
the
minimum.
Every
column
must
have
a
multi-dimensional
metric
space.
We
talked
about
that
live
here,
the
only
examples
we
know
about
education.
We
sell
my
cortex.
Therefore
we
conclude
that
there
must
be
multiple
bits
on
modules
in
a
column,
one.
C
A
So
modulus
of
efficient
eating
multiple
months
to
define
a
space.
So
that's
that's
that's
one
of
the
the
premises
of
the
thousand
great
intelligence.
It
also
has
to
be
a
multi
dimensional
orientation
match
the
multi
dimensional
metric
space,
so
we
have
to
know
which
way
you're
a
person
which
way
you're
facing
of
your
finger,
which
which
orientation
and
so
on
and
different.
C
A
Shelter
represent
orientation
relative
to
the
space
crystals,
so
this
is
important
parts
of
this
basic
stuff.
That's
in
the
questions,
there's
a
lot
of
things
we
don't
know.
These
are
some
other
puzzling
questions.
We
have,
for
example,
where
are
the
big
coal
miners
in
a
calm
we
think
they're
in
layer,
six
or,
but
we
actually
need
many.
Many.
A
Each
grain
six
we've
had
the
same
problem
in
the
entorhinal
cortex.
We
looked
at
drying
partition.
The
research
on
that
and
it'll
show.
We
need
negative
ones.
Actually,
both
one
block
you
could
to
make
it
work
and
it's
going
to
create
aa
person.
That's
why
our
next
question
is
what
is
the
dimensionality
if
it's
a
logical
to
think
of
them
as
two
dimensions
and
then
throughout
the
cortex?
And
maybe
that's
true?
Maybe
it's
not.
We
thought.
We've
got
a
whole
bunch
of
papers
about
how
some
animals
might
not
be
the
three
dimensions.
A
It's
not
clear
what
it
is
in
the
cortex.
It
could
be
two
dimensions.
We
don't
know
that
we
don't
know
where
the
orientation
cells
and
your
cortex
and,
more
importantly,
the
orientation
and
the
green
cells
have
to
work
together.
You're
like
how
do
you
update
results
depend
on
your
orientation.
If
you
want
to
modify
one.
C
A
A
Being
motor
drive,
but
we
concluded
that
updated
by
sensation
as
well.
This
was
the
idea
that
you
should
be
playing
a
video
game
and
as
you're
moving
through
the
video
game
or
not
you're
walking,
but
your
sense
of
location
changes,
and
so
that
has
to
be
based
on
flow
information
on
screen.
So
talk
about
something
about
this,
but
when.
B
A
Is
a
column
it's
getting
some
sort
of
motor
input
which
I'm
showing
here
and
I
wonder
if
it
can
move
the.
We
know
like
a
movie
itself,
but
I
always
thought
that
that
had
to
be
learned
that
account
doesn't
know
what
motor
behaviors
cooking.
It
has
to
figure
that
out
so
I'm
gonna
work
on
the
sensation,
part
of
this
and
then
and.
A
So
you
do
a
lot
of
very
puzzling
problem
and
I'm
innocent.
Just
a
solution
is
an
idea.
Maybe
a
proposal
I'm
not
ready
to
call
it.
It's
just
an
idea,
but
it's
a
very
exciting
idea
for
me
and
I'm,
not
sure.
If
you're
trying
about
the
basic
idea,
I
just
have
to
answer
these
questions
and
the
basic
idea.
That
goes
as
follows:
each
mini
column
represents
from
one
dimensional
solar
system
in
one
dimension
in
the
space.
A
That's
a
model,
so
I
said
that
maybe
column
is
an
over
complete
basic
set
of
dimensions,
so
I
might
have
100
mini
column
dimensions
in
physical
space,
I'm,
going
to
disperse,
like
one
dimensional
vectors
and
over
complete
fashion,
and
that's
how
it's
that
word
well,
a
sudden
mini,
confident
we're
at
any
number
of
dimensions,
meaning
it
could
argue
that
you
can
discover
the
dimensionality
spaces.
The
two-dimensional
space
will
end
up
in
common
in
three
dimensions
around
one
dimensions
that
are
plain.
If
it's
big
mentions
it
be
mentioned.
C
A
Figure
out,
based
on
the
sensory
input,
let
the
dimensionality
of
the
spaces,
so
you
can
imagine
these
little
bit
cell
models
would
point
you
to
the
directions.
If
you've
got
a
question
on
this,
let
me
know
that's.
This
is
a
key
thing
here
and
I'm
gonna
say
that
each
mini
column,
therefore
one
need
we'd,
sell
module
and
it's
gonna
be
a
later
stage
now
I'm
going
to
show
that
over
here
in
a
moment,
I
have
no
physical
evidence
for
this.
Okay,
this
is
not
making
this
off
I
would
do
think.
A
We
have
a
lot
of
evidence
in
grid
cells
or
in
later
sticks.
I
won't
do
that
right
now,
but
I'm
gonna
argue
the
following
I'm.
Actually,
these
are
mini
comp
and
there's
a
bunch
of
cells,
making
a
major
succession
like
that
create
twenty
of
hundreds.
It's
obviously
this
could
be
twenty
like
that
and
then
I
there's
a
bout
of
activity
in
that
mini-com.
A
Own
that
you
can
think
of
the
cells
here
as
in
one
dimension,
and
there
might
be
a
couple
of
months
so
that
you
can
wrap
around
usually
I.
Don't
really
know
the
point
of
that.
Each
of
these,
in
many
columns
have
a
bomb
that
represents
they
would
sell
like
repeated
location,
England,
one
of
the
mansion
and
and
I
wanted
to
meet
out
my.
A
Survived
hundreds
of
these
guys
I've
got
a
subsample
from
them
almost
randomly
on
the
character.
You
get
a
good
representation
location.
Now
these
bumps
have
to
move
when
you
move
right
s
all
ideas.
If
I'm
moving
my
finger
over
an
optimized
point,
I'm
walking
through
space
he's
half
the
moment,
they
don't
all
move
at
the
same
rate,
absolute.
B
A
B
A
A
A
B
A
C
A
That's
the
frustration,
fuller
and,
and
it's
going
to
take,
take
all
these
different
possible
directional
senses
bits
and
it's
gonna
divide
them
up
into
a
nice
set
they
so
spans
in
space
like
the
station
truly
does.
But
these
the
outputs
of
this
are
going
to
essentially
be
directions.
They're
going
to
be
essentially
1b
directions
in
face,
meaning
I'm,
going
to
look
at
all
these
things
and
try
to
apply
all
these
different
directional
risk
and
if
I
turns
out
to
be
the
only
the
two
dimensions
and
like
this.
A
Function,
I
will
then
define
many
columns
to
you.
They
don't
want
dimensional
attractions
and
the
activities
of
these
cells,
meaning
how
they
are
at
any
point.
Time
is
exactly
this,
so
if
I
this
the
uncensored
movement,
then
these
cells
will
become
active,
integrated
fashion
and
that
radioactivity
would
be
the
signal
and
we
need
to
know
how
to
remove
these
cells.
A
Happens:
I'm
just
saying
with
that
busy
signal
like
one
this
is
envision.
Of
course,
we
can
I
think
where
this
might
work
envision
during
the
treatment
vision.
I
have
these
directly
sensitive,
that's
coming
if
I
have
ones
that
are
converging
or
coming
apart.
I
move
in
three
dimensions
like
this,
and
they
don't
like
this
for
that
I'm
hoping
to
be
this
way,
so
I
think
that
would
work
I,
don't
like
immediate
option
on
this
one,
but
it
was
a
lesson
so
we're
racing
up,
say:
okay!
Well,
this
is
advised
if
I
say
this
might
regret.
A
A
A
couple
other
things
before
I,
get
to
my
thing:
little
confused
about
every
every
column
has
the
motor
output
those
we
player,
five
B
cells
which
are
motor
around,
but
so
we
should
learn
any
mini
column.
What
this
this
is,
it
might
generate
a
tremor
that
is
vital
805.
We
start
here
and
become
active.
A
A
A
Vacuum
there's
ticks
and
therefore
that
would
still
work
here.
This
has
to
be
something
different,
though
right
right
so
I'm
saying
here
this
must
be
spatial
Hatter.
Thank
you.
I
was
recognizing
a
directional
pattern
here.
I
need
to
say
whether
the
spatial
pattern
that
I
wanna
associate
with
that
point
locations
with
then
examine
them.
Those.
A
A
B
I
did
say
that
yeah,
but
a
one
sort
of
interesting,
not
gonna,
buy
that
I
think
this
is
true
with
bipolar
cells.
You
have
some
wearing
a
homeland
that
are
wearing
the
cell
lobbies
up
there,
two
three
and
they
go
down
and
if
you
have
another
set,
went
to
cell
bodies.
I
declare
six
going
because
one
plane
makes
you
like
Tuesday
Monroe's
here
is.
We
have
two
basic.
A
A
Oh,
and
so
it's
very
angry
at
this
point
in
time,
but
I
say
my
sense:
spaceship
hadn't
appear
in
a
sense,
directional
patterns
down
here,
I
think
I
can
show
you
know
the
entire
educational
organization,
all
that
stuff
down
here,
and
it
will
look
like
what
we
see
to
detect.
We
see
these.
We
see
responses
like
this
punch,
many
columns,
but
we
think
of
I
notice.
It
orientation.
We
know.
A
So
if
we
do
not
over
here,
it's
okay,
so
so
we're
gonna
drizzle
modules
on
comm
I'm,
marking
that
this
is
just
a
hypothesis
that
there
in
each
individual
community
calm,
we
need
many
of
them.
We
won't
have
hundreds
of
them
in
any
mini-com.
What
does
it
mention?
I
think
this
whole
module,
so
that's
one
D
and
and
where
were
they
were
impatient
cells
on
murder?
The
orientation
is
actually
encoded
in
the
many
calls
himself.
A
Input
there,
how
did
they
were
in
taste,
ourselves
and
good
cells
work
together
to
be
more
to
move,
acted,
big
cells
correctly?
I
was
learning
this
treatment
based
on
orientation.
Well,
it's
very
simple
in
station.
Just
have
the
scalar
that
scalar
this
says
when
I
activity
in
this
this
cell
here
within
this
activity,.
A
A
B
A
A
A
B
A
B
A
B
A
Yeah
I
am
so
many
know
what
I
do
slowly.
First
of
all,
remember
we
need
lots
of
many
columns
because
we
need
also
a
bit
illogical
to
just
all
the
difference
back
to
university
in
large
spaces.
We
need
a
lot,
so
we
did
some
sample
and
still
get
reliable
results.
We
need
a
lot,
so
we
could
that
or
not
anyone.
Is
it
done
that?
Maybe
this
is
me
I.
A
A
A
A
There's
other
requirements
we
haven't
talked
about,
but
I
just
housing
is
really
excites
me
about
this.
If
I
have
a
solution,
but
I
get
many
bits
of
my
goals,
how
orientation
works?
How
I
could
derive
that
from
sense
we
can
plug,
which
is
the
key
thing.
That's
ideally
no
spatial
Euler
over
a
directional.
A
And
it
will
solve
all
of
the
problems
a
month.
Basic
equality
problems
into
the
simple
solution
and
I
never
have
a
solution
for
end
of
it.
So
that's
what's
exciting
about
it,
but
we
talk
about
one
more
thing
here.
I've
still
had
this
idea
that
the
theta
rhythm
between
the
fountains
in
the
cortex
could
be
could
be
a
way
of
scale
mix
of
mountain
path.
To
guess
about
scaling
that
this
idea
that
that
the.
B
A
A
A
A
A
A
I'm
suggesting
that
the
way
of
smart
thing
about
this
first
is
really
focus
on
this.
One
I
think
this
is
the
primary
way
we
learn.
The
dimensionality
space
is
from
your
rating
actually
detect
motion.
A
secondary
way
has
to
be
learned,
so
there's
no
motor
command
coming
in,
but
he
said
that
could
be
just
the
faders
single
and
if
I
knew
my
orientation
already,
if
I
already
had
my
orientation
shut
up,
then
all
I
need
is
to
say
how
fast
are
you
moving?
A
B
A
A
A
A
C
This
hey
Can
Can,
where
you
guys
take
the
double
back
for
me,
please,
okay,
all
right,
hey
that
was
exciting.
Okay,
so
I'm
gonna
have
to
rewatch
that
because
I
was
distracted
and
it
seemed
like
really
interesting
stuff.
There's
a
lot
of
theory
here
that
Jeff
was
talking
about
that
I
haven't
heard
before
so
this
is
the
first
time
we've
thought
about
it.
First
time,
I've
thought
about
it.
So
if
you
want
to
discuss
it,
the
forum
link
is
in
the
video
description.
C
If
it's
not
I'm,
gonna
put
it
there
right
now,
thanks
everybody
for
chatting
along
I'm
busy.
As
you
can
see,
my
office
is
a
mess.
I
have
to
get
to
work
now,
but
stay
tuned
for
more
research.
Videos,
probably
another
one.
Next
week
at
least
one
maybe
one.
This
Wednesday
anyway,
Cole,
sparse
representation,
Thurs
derive
a
dictionary
of
atomic
features
that
can
be
linearly
combined
it
to
any
observable
feature
using
a
few
atoms
as
possible.
That's
a
good
good
thought.
C
Sacco
says
great
stream.
Thank
you,
Oh
guys
and
gals
folks
in
general.
Please,
like
the
video
and
subscribe
to
our
YouTube
channel.
If
you
haven't
already
good
number
of
people
watching
thanks
again
for
watching
and
let
people
know
what
we're
doing
here,
this
is
very
unusual
in
the
science
community
to
be
I,
think
live-streaming
of
private
companies,
research
meetings
and
in
the
realm
of
artificial
intelligence,
you're
getting
definitely
bleeding
edge
stuff
here
from
Numenta
as
it's
coming
from
our
founders
mouth.
C
This
is
the
first
I've
heard
of
this
theory,
so
interesting
stuff
about
Oh
John
was
asking:
do
you
think
jeff
has
the
same
intuition
that
understanding
spatial
nests
as
a
basis
vector
trig
the
case
of
leads
extrapolation
to
concept?
Math
absolutely
has
an
intuition.
We've
talked
about
it
many
times.
That's.
That
is
the
idea
that
that's
where
we're
trying
to
get
to
understanding
thoughts
as
things
that
can
be
computed
and
compared
and
etc
in
internally,
within
an
agent's,
an
agent
system.
C
Now
how
how
useful
those
representations
are
outside
of
that
system
is
debatable
and
we've
had
had
those
debates
when
we've
talked
about
agency
and
knowledge
transfer
in
the
past
or
my
twitch
stream
I
think
you
can
find
those
on
our
YouTube
channel.
If
you
search
for
agency
and
knowledge
transfer
as
well
of
what
happened
to
my
MS,
the
chat
up
here
hold
on
a
sec.
I
can't
see
I
can't
see
where
the
chat
is
there.
We
go
okay,.