►
From YouTube: Numenta Office Hour - June 7, 2016
Description
Come chat live with Numenta founder Jeff Hawkins, VP of Research Subutai Ahmad, and members of the Numenta engineering team.
A
B
A
C
A
D
A
If
you
want
to
join
otherwise
there's
a
Q&A
interface,
you
can
type
in
your
question
as
a
comment
and
I
will
keep
track
of
those,
so
you
so
will
feel
those
as
they
come
across,
but
for
now
I
think,
let's
I
don't
like
to
start
this
off
with
with
you,
differentiating
between
the
different
types
of
columns
that
there's
been
a
bit
of
confusion
about
in
this
discussion
when
we
take
me
column
and
use
a
portable
column.
You
know
that
there's
some
naming
problems
here.
You
can
talk
about
how
you
think
of
who's.
Looking
at.
E
Me
man
Jeff,
yes,
well,
okay,
I
think
the
mini
column
is
pretty
uncontroversial.
That's
a
physical
thing
that
know
exists.
I
starts
at
the
development
of
the
brain
and
somewhere
around
100-120
cells
in
this
very,
very
skinny
column.
It
with
us
falling
somewhere
between
30
and
50
microns
wide,
and
so
that's
been
known
for
a
long
time
and
also
there
sometimes
a
vision,
they're
actually
visible
under
the
microscope,
sometimes
or
not
there.
There's
a
lot
of
evidence
suggests
that
they're
partly
defined
by
certainly
so
immature
cells
so
on.
E
But
now
we've
used
a
meeting
count
for
a
long
time.
We
use
that
in
our
temple
the
algorithm
kind
of
the
core.
They
come
a
memory
algorithm
and
we
have
proposed
a
functional
need
for
many
columns.
That's
nothing's
changed
there
and
the
Newark
were
still
keeping
move.
Actually
many
columns
and
all
that
in
the
in
the
temporal
memory
couldn't.
E
Yeah
I'd
always
have
some
number
of
them
and
we
typically
use
2048.
That's
a
good
number.
You
can
get
down
with
little
as
a
thousand,
but
then
you
start
losing
some
of
your
str
properties.
So
we
pick
2048
and
that
block
of
many
columns
is
the
way
we
use
it
in
the
temple
memory.
In
some
sense,
they
they're
all
the
output
of
the
SP,
the
spatial
pooler
and
they're
further
they're,
not
they're,
not
differentiated
other
than
in
in
terms
of
topology
they're.
Just
like
a
unit.
E
That's
how
we
always
worked
that
that
is
not
exactly
correct
from
a
biological
point
of
view,
because
the
cortex
is
sort
of
a
continuum
of
these
things,
but
it
works
when
you
just
think
about
like
okay.
What's
the
inhibitory
spread
between
many
comps
now
in
the
new
stuff,
we're
basically
doing
the
same
thing
and
I
were.
We
are
essentially
defining
sort
of
the
minimal
amount
of
cortex.
E
We
can
use
to
do
spatial
temporal
motor
infants
and
one
way
to
think
about
it
that
this
been
helpful
for
us
or,
for
me
at
least,
is
instead
of
thinking
of
a
continuous
sensor
like
the
retina
or
the
continuous
space
on
your
skin.
You
can
just
think
about
Oh
what
am
I
just
think
about
the
tips
of
my
fingers
and
those
those
are
not
represented.
E
Those
are
so
somewhat
separate,
really
represented
in
the
cortex
near
each
other,
but
divided,
and
we
can
just
think
about
having
the
minimal
size
another
things
coming
up
on
the
screen
there
is
that
something
I'm
supposed
to
be:
no
okay,
it's
like
right
in
my
face
so
and
it's
really
not
much
different.
We
did
with
the
temple
memory
algorithm,
but
we
reach
a
little
bit
more
we're
trying
to
use
multiple
on
it
once
now.
E
Well,
memory
we
just
had
one
and
we
could
do
sequence
members
just
one
of
these
things,
but
now
we
have
multiple
ones.
So
we
have
to
start
thinking
about
them
like
one
representative
in
parts
of
the
different
tips
of
different
fingers
or
different
parts
of
the
retina,
and
so
now
what
do
we
call
that
in
biology?
Actually
there's
as
I
said,
there's
a
continuum.
It's
not
like
the
really
rigidly
divided
some
places.
E
We
only
just
had
one,
and
so
now
we're
going
to
have
multiple,
and
so
we've
talked
about
using
the
term
hypercom
or
cortical
com
or
cortical
module,
they're,
all
kind
of
referring
to
the
same
thing.
It's
basically
enough
of
cortical
tissue.
A
couple
thousand
many
columns
where
there
are
all
sort
of
under
the
same
auspices
of
one
spatial,
puller
and
so
mutual
inhibitory
and
I
think
Fergal
suggested
using
the
term
cortical
column.
That's
one.
We've
considered
I
like
it,
because
it's
CC
that's
easy.
One
of
the
problems
in
this
long
answer.
I
hope,
that's.
E
Okay,
one
of
the
problems
is
a
long
history
of
this
nomenclature,
and
so
you
can
find
it
as
a
paper
I
like
which
I've
read
been
other
ones.
That
basically
say
the
cortical
column
doesn't
exist,
it's
nothing!
It's
not
real,
and
so
by
using
that
term,
you
sort
of
invoke
the
skepticism
of
some
people
so,
but
we
haven't
decided
the
term
for
yet,
but
it's
basically
just
multiple
ones.
So
the
things
we've
been
doing
before
and
recognition
that,
in
reality
and
core
textures
are
continuing
with
these
things.
E
F
E
A
David
race
chatting
some
questions
which
you're
going
to
do
with
your
it
says
is:
is
there
evidence
in
the
biology
that
the
transform
that
we're
speaking,
I
was
actually
occurring
and
is
didn't
known,
how
the
transformers
I
students
in
biology
and
do
alluded
to
this
in
the
video,
the
soul
of
learned
procedure?
Is
heart
disease?
Your
outer?
How
we
do
it
is
everyone
going
to
know
what
we're
talking
about
here,
yeah
I'm,
assuming
that
everyone
has
watched
the
video
that
I've
linked
everywhere?
A
E
First
of
all,
this,
this
sort
of
transformation
has
been
known
to
needed.
It's
understood
in
the
non
neuroscience
literature.
It's
it's
a
fundamental
premise
of
Robotics.
If
anybody
who's
doing
a
you
know,
body
just
base
transformation.
So
it's
not
a
new
idea
at
all
and
and
I
think
it's
been
known
to
be
needed
in
biology.
It's
sort
of
an
obvious.
The
thing
that
has
to
happen
I,
don't
think
I'm
not
aware
of
anyone.
Who's
actually
just
speculated
on
neural
mechanisms.
For
that
may
be.
There
happen
you
a
shaking
the
title
b
line,
senior.
E
E
A
says:
no,
I'm
not
aware
of
it
either,
but
it's
kind
of
like.
If
I
mean
I
can't
say
nothing
is
totally
hunting
obvious,
but
it's
pretty
clear
that
this
has
to
happen
and
and
it's
complicated,
it's
not
easy.
This
is
not
a
simple
thing
to
do,
but
it
really
has
to
happen
and
it's
known
that
it
has
to
happen
outside
of
nerves
on
it.
Just
don't
ask
my
inside
of
neuroscience,
it's
known
to
happen.
It
has
to
happen,
but
no
one's
really
proposed.
E
As
far
as
we
know
many
mechanisms
for
this,
but
given
we
can
define
what
it
is
and
what
it
has
to
do,
then
you
know
it
has
to
occur
in
the
neural
tissue.
Also,
the
current
believe
pretty
strongly
is
that
has
to
occur
everywhere
in
the
cortex.
It's
a
fundamental
property
of
all
all
sensory
cortex.
At
least
you
know
primary
or
secondary
and
tertiary
central
cortex.
So
it's
not
like
some
little
thing.
That's
occurring
over
here.
It's
pretty
fundamental
thing
as
that
everywhere.
E
So
that's
that's
pretty
much
yeah
and
this
papers,
which
I
read
a
paper
here,
I,
don't
I
referred
to.
In
that
talk
I
gave,
but
there
was
a
paper
here,
was
doing
psychophysics
that
I
was
talking
about
where
people
study
the
time
it
takes
to
do
these
mental
rotations
or
these
transformations,
and
so
there's
a
lot
of
psychophysics
evidence.
E
And
now,
when
you
try
to
solve
this
problem
of
this
transformation,
you
need
a
lot
of
neural
machinery
to
do
it,
and
so
all
of
a
sudden,
you
can
start
laying
this
into
some
of
the
stuff
going
on
in
layer,
6
and
layer
4,
and
maybe
the
thalamus
that
we
hadn't
have
a
reason
for
before.
So
it's
we
don't
have
any
answer
to
it.
Yet,
but
all
of
a
sudden
we
have
a
set
of
requirements
of
a
functional
requirements.
A
E
H
E
F
Let
me
know.thank
an
add-in
is
even
though
they
haven't
found
necessarily
direct
evidence
for
transforms.
There
are
different
frames
of
reference
that
are
known
as
in
the
brain,
so
there's
clearly
lots
of
areas
where
your
sense,
sensing
things
that
are
relative
to
the
sensory
frame
of
reference.
Whatever
you
know,
retina
is
just
relative
to
wherever
we
are
I
read.
You
know
you
know
pixels
on
the
eye
and
then
in
premotor
cortex.
F
There
are
frames
of
reference
that
our
body
centric,
so
it
represents
a
position,
that's
independent
of
the
position
of
your
body,
but
a
position
of
your
limbs.
But
it's
specific
to
some.
You
know
location,
that's
relative
to
the
central
body
and
in
the
hippocampus
there
are,
you
know,
play
cells
which
seem
to
be
you
know,
environments,
centered,
not
standard
on
your
body
necessarily,
but
it's
specific.
It's
a
frame
of
reference
sets
orient
focused
on
on
the
environments,
I.
E
F
E
It's
happening
everywhere
happening
very
rapidly,
so
it's
it's
yes,
so
you
mean
it's
not
having
a
high
place
and
coordinate
that
doesn't
require
a
higher.
It's
still
me,
it's
going
to
be
happening
every
region,
so
b1
b2,
s1
s2.
Whatever
you
explain
to
correctly
it's
a
core
aspect
of
all.
You
know
you
look
at
what
every
region
cortex
has
certain
cellular
and
layer
properties
that
are
preserved
and
it's
going
to
be
involved.
It's
going
to
be
one
of
those,
it's
convolve
with
that
by
the
way
of
paying
a
lot
of
time.
E
Looking
at
layer,
6
is
a
great
paper
by
alex
thompson
about
layer,
6,
and
it
goes
into
great
detail
about
lately
or
six
are
doing
but
later
successes.
It
comes
to
things
like
6a
and
6b
and
is
communicating
with
the
where
pathway-
and
this
is
a
really
prime
area
understanding
what's
going
on
because
later
sticks
is
the
primary
input,
the
layer
for
surprising
it's
sixty-five
percent,
they
put
the
layer
4,
so
this
close
there.
This
is
this.
E
F
E
A
I
E
I
I
Is
it
you
know
what
those
are,
but
you
guys
well,
no
I
have
no
clue,
but
what
I
was
this
when
you
were
discussing
this
with
Fergal,
it
is
actually
the
first
I
had
ever
heard
of
them,
and
and
I
was
just
wondering
if
you
could
go
into
a
little
bit
more
detail.
You
know
about
what
they
are
yeah.
D
Individual
system,
when
the
very
information
first
arrives,
you
primarily
rural
context.
We
want
its
paste
split
up
into
two
pathways.
My
is
known
as
the
world
pathway,
which
is
concerned.
Are
they
protecting
the
identity
of
object
like
for
our
recognition?
Is
recognition?
That's
in
the
ventral
stream
of
two
on
the
bottom
of
the
top
ring
and
the
other
pathway
is
known
as
a
weird
pathway.
The
care
is
lies
about
what
is
being
presently.
E
E
We've
also,
we
looked
at
papers
and
we
it's
pretty
well
documented,
also
in
the
auditory
system
and
the
somatosensory
system.
So
it's
a
it's.
A
kind
of
common
feature
of
all
sensory
processing
is
that
the
input
comes
in
from
the
Census
and
it
goes
in
two
parallel
higher
and
one
is
associated
as
you
a
said,
with
sort
of
like
where
objects
are
in
your
body,
space
and
related
to
actions
related
to
your
body
and
the
other
one.
The
one
people
think
about
most
is
like
what
the
objects
are
and
and
actions
related
to
the
object.
E
So
you
have
these
two
parallel
things
going
on
and
they
talked
back
and
forth
to
each
other
continuously.
That's
what
I
mentioned
in
layer
6.
So
you
have
the
sort
of
combination
of
where
something
is
and
what
something
is,
and
it's
funny,
because
those
too
high
the
hierarchies
are
different,
but
the
regions
were
very
very
similar.
They
know
they
look
identical,
so
you
got
the
same
layer
structure
in
there.
E
So
the
basic
idea
is
that
they're
getting
slightly
different
input
and
ones
modeling
like
we
think
one's
modeling,
proprioceptive
sense
and
the
others
modeling
like
more
than
visual
object
anyway.
This
is
not
a
lot.
There's
a
lot
known
about
it,
but
not
a
lot
known
about,
but
it
seems
to
be
pretty
important
and
I.
Think
in
that
talk,
I
talked
about
them.
A
bit
did
I
forget
we're
one
game:
yeah
a.
A
E
E
Blanket
if
I
have
a
region
on
the
what
path
of
the
region
the
web
path
with
their
layer,
6
just
communicate.
Oh
okay,
so
I
mean
the
basic
ideas.
Imagine
I'm!
I
find
moving
something
in
my
body.
Space
like
I'm
moving
my
hand
forward
where
it
moves
on
an
object,
is
independent
of
the
orientation,
the
objects
and
so
on.
So
I
can't
say
that
I'm
moving
my
hand
forward.
E
It's
going
to
be
moving
on
at
any
particular
thing
on
an
object
you
have
to
make
transform
and
it's
simile
if
I
want
to
move
from
one
feature
of
an
object
like
it's.
My
coffee
cup
I
want
to
move
some
on
feature
here,
find
one
move.
My
finger
around
the
circle
the
behavior
I
have
to
do
is
depend
on
where
the,
where
the
company's,
if
I
tilt
the
cup
like
this,
it's
a
different
behavior.
If
it
still
delay
it's
a
different
position,
and
so
you
have
to
go
back
and
forth.
This
is
again.
E
This
has
been
known
in
robotics
for
a
long
time.
You
have
to
go
back
and
forth
between
a
movement.
You
know
if
I
want
to
make
a
movement
on
an
object.
I
have
to
transform
that
into
a
specific
movement
in
my
body,
which
is
depending
on
where
the
object
is
I
can't
always
say,
move
your
hand
and
certain
way
to
get
the
rim
too
handle
it
depends
on
where
the
object
is
you
have
to.
There
has
to
be
this
transformation
between,
like
hey
I,
want
to
make
a
movement
if
I
make
a
movement.
E
My
bodyspace,
what
see
a
pat
impact
on
the
on
the
object
of
our
maker
movement
on
the
object,
space
or
I
want
to
get
in
some
place
on
the
object.
How
do
I
have
to
make
that
happen
in
my
hand?
So
this
is
a
very
well
an
unknown
problem,
not
well
understood,
probably
very
known
problem,
and
you
can
solve
it
using
traditional
coordinate
geometry
type
of
stuff
at
the
bank.
Notes
able
to
do
it
that
way.
E
J
E
G
E
It
you
know,
I
can
see
a
little
bit
more
about
this.
We
there's
some
things
about
this.
These
mysterious
we're
very
excited
about
I'm,
very
excited,
but
assuming
we
just
don't
understand
in
this
transformation
issue,
the
nature
of
the
coordinate
space
that
objects
are
represented
is
confusing.
So
that's
an
active
problem
that
a
number
of
us
are
working
on
and
it's
not
easy
and
if
anyone
else
really
really
wants
to
get
into
it,
Rick
I
fairly
confident
will
solve
it.
E
But
it's
you
know
until
you
do
it,
you
never
one
hundred
percent
certain,
but
it's
a
challenge,
but
there's
so
much
compelling
about
the
rest
of
it
that
it
feels
like
okay.
This
is
a
problem
we
have
to
solve.
A
And
David
was
also
talking
about
some
process
questions
earlier,
a
little
bit
glazing
over
it
and
he's
asking
like
how
do
these
realizations
of
your
reserve?
These
aha
moments,
as
far
as
the
theory
goes,
get
translated
in
their
escalated
into
actual
theories
and
then
what's
the
process
for
turning
those
an
algorithms
like.
What's
the
plan
look
at
what
can
they
expect
right.
F
I
can
talk
a
little
bit
about
the
second
part
of
it.
I
think
we
talked
about
these
theories.
We
have
quite
extensively
and
we
walk
through.
You
know
the
assumptions
and
the
properties
we're
looking
for
how
it
might
be
manifest
in
the
biology
and
then
what
I
like
to
do
is
that
try
to
take
that
into
one
level
more
detail
of
at
least
pseudocode
level
detail.
F
So
we
can
understand
it's
precisely
how
it
might
be
implemented
in
a
neuron,
what
the
rules
might
be,
what
it,
what
the
different,
what
you
want
to
do
in
all
of
the
different
conditions,
so
that's
kind
of
the
level
of
detail
they
lied.
I
like
to
be
in
that
way
from
their
implementing
is
pretty
straightforward,
but
I
think
going
from
I
have
no
idea
how
Jeff
gets
is
breakthroughs,
but
but
from
the
theory
to
the
pseudocode
level
is
the
really
tough
process?
I
often
it's
for
me.
E
I
think
we
go
through
a
sort
of
a
cycle
here,
I
think
I,
think
in
I.
Don't
think
mathematically,
I,
don't
think
in
pseudocode
I
think
instead
of
visual
physical,
it's
hard
to
describe
sort
of
physical
construction
in
the
neuroscience
and
then
sumotori
tries
to
translate
that
you
a
is
also
really
good
at
the
neuroscience.
So
he
come
taxes.
That's
not
true-
or
this
is
true.
Scott
is-
is
sort
of
a
great
critic
and.
E
F
But
when
one
technique,
I
just
seen
you
use
over
the
years,
that's
been
very
effective,
is
don't
get
too
narrowly.
Thinking
about.
One
thing
is:
there's
always
like
a
hundred
different
constraints
that
have
to
be
satisfied
simultaneously
or
many
many
many
constraints,
it's
easy
to
think
of
a
solution
that
solves
one
particular
problem
or
one
aspect
of
it,
but
it's
really
hard
in
some
sense
to
think
of
a
solution.
That
means
all
of
the
constraints
yeah.
E
E
And
what
happens?
Is
it
the
more
constraints
you
throw
onto
the
problem
that
harder
to
see
you
initially
it
just
like?
Oh
my
god,
somebody
things
like
it,
but
if
you
can
get
when
the
answer
comes,
it's
more
clear.
That's
the
right
answer
in
this
particular
case.
Not
only
trying
to
figure
out
what
I
like
to
figure
out,
how
we
learn
to
learn
and
infer
what
objects
in
the
world
are
to
touch
and
vision
and
so
on.
E
E
That's
all
happening
everywhere
in
the
cortex,
and
so
it's
you
know
we
really
like
to
do
is
come
I'm
trying
to
do
nice
if
we
come
up
with
theory
explains
not
only
to
how
do
we
infer
what
these
objects
are,
but
we
also
at
the
same
time,
movement
not
only
tells
me
what
to
predict,
but
it
also
sometimes
change
the
subject.
So
those
are
just
there's
a
series
of
large
number
of
problems
we
might
be
able
to
solve
all
at
once
here.
E
How
do
you
behave
and
how
objects
transform
under
behavior
I'm,
not
sure
we
get
all
that
done,
but
we
managed
to
try
that
you
know
what
we
did
it.
The
sequence
of
memory
many
years
ago
with
a
simple
memory.
Is
we
just
picked
a
thing
of
hiding
form
representations
of
our
sequences,
because
then
that
was
that
was
a
good
chunk
to
pull
off
and
it
took
us
a
real
long
time
to
really
get
a
deep
understanding
of
it.
E
J
G
A
You
have
12
viewers
in
addition
to
these
guys.
I
had
a
question,
don't
mind
sure
you're
talking
about
an
object
being
represented
an
object
space,
and
maybe
talking
about
and
like
in
your
example
I'm
assuming
that
that
object
being
represented
in
your
brain,
is
a
generic
representation
of
a
pen
right,
a
base,
all
the
different
as
objects
like
that.
No,
maybe
it's
both
okay.
E
That's
all
right
computer,
so
here's
here
is
this
is
what
the
unova,
your
ball
pen
I,
know
this
pen
I
have
this.
You
know
I
play
with
these
pens
cried
a
lot,
and
you
know
this
is
super
ties
and
I.
So
you
have
certain
expectations
about
this
very
pen.
You
know
so
and
then
I
have
other
pens,
I
know,
but
then
I
can
get
a
new
pen
which
is
some
more
generic
right.
So
the
problem
is,
you
have
to
be
able
to.
E
Sometimes
you
have
a
very,
very
specific
knowledge
about
something,
and
you
have
you
know
the
details
of
that
and
it
can
vary
much
from
that.
All
your
predictions
are
very
precise.
Other
times
you
have
more
generic
version.
That's
one
of
the
problems
we're
trying
to
solve
it.
It's
easier
actually
to
define
the
precise
pit,
but
exactly
I
know
exactly
what
this
pen
suppose
babe
like,
because
it's
the
one
on
familiar
with
my
train
down,
it's
a
little
harder
to
say
like
oh,
how
do
I
understand
what
a
generic
panels
and
have
expectations
about
it.
E
A
Talk
about
that
post,
earrings,
Oh,
Fergus
know
the
ones
the
picture
of
the
neurons,
then
one
else
beat
them.
I
know
yeah.
I
read
it.
I
think
I
think
we're
going
on
it's
a
bit
of
that.
That
photographs
is
not
the
detail
that
we're
talking
about
them,
no
temp
but
yeah.
So
I
don't
see
any
other
questions
on
the
forum
or
imagined.
A
A
D
I
E
It's
a
very,
very
generic
struck,
I
dia,
but
then
we
wanted
something
practical
with.
It
was
a
lot
of
practical
things.
The
first
thing
you
think
about
you
can
build
being
a
robot,
robotic
arm,
something
that
pick
things
up
and
mainly
know
what
it
is
not
manipulate.
It
I
think
we
do
ready
very
rapidly
than
simply
that's.
That's
an
example
that
you
might
just
you
know,
put
a
human
level
behavior,
but
you
know
it
could
be
anything
it's
well
consider.
Actually,
if.
A
Could
have
an
AI
into
the
end,
video
game
that
can
move
for
the
whole
world.
But
if
you
think
about
the
Internet
as
a
world-
and
you
can
wrap
this
type
of
aia
entity
that
has
some
motor
integration
with
an
interface
that
understands
cute
asian
protocols,
it
could
explore
the
entire
internet,
like
yeah,
given
some
goal
or
goals
to
attain
yeah.
Just
by
following
those.
A
E
Think
it's
you
know,
I
think
it
even
brought
in
that
you
think
about
all
the
things
humans
are
able
to
do.
I
think
I've
said
this
in
previous
post
I:
don't
there
are
really
only
two
ways
of
discovering
the
structure
of
the
world
one
is
the
world
is
moving
and
the
others
you
move
through
the
world
in
interactive
world
and
the
vast
majority
of
learning
is
the
latter
type.
Everything
work
we've
done
up
to
now
is
a
forum
on
time.
So
this
is
a
bigger
I.
Think
mathematics
is
is
a
derivative
of
this.
E
That
is
mathematicians.
You
have
concepts
that
are
like
have
structure
and
you're,
manipulating
them,
and
you
flip
them
around
and
you
do
transformations
on
them
which
are
like
motor
behaviors.
This
is
a
very
abstract
thing,
but
the
generic
algorithms
that
the
cortex
implements.
Oh,
they
were
evolved
for
vision,
touch
and
hearing.
They
obviously
work
for
other
things
that
are
not
like
that
at
all,
and
so
I
believe
that
this
generic
algorithms
can
say
like
I'm,
not
it's
not
restricted
to
three
dimensional
spaces.
E
It's
not
restricted
to
movements
that
are
in
three
dimensional
spaces
that
we're
used
to,
but
they're
learning
algorithms
really
could
be
applied
to
any
kind
of
structure
and
any
kind
of
high
dimensional
space.
Any
kind
of
transformations
and
behaviors
I
think
it's
what
we
do
when
we
do
mathematics
and
when
we
do
physics
when
we
do
computer
design
or
software
designer
and
ever
so
very
very
high
level.
I
think
this
gets
at
the
core
of
how
we
model
the
world
and
how
we
discover
structure
and
how
we
manipulate
that
to
structure
two
jeeps
or
net.
E
So
it's
a
you
know.
This
is
not
going
to
be
like
I,
don't
think,
there's
a
third
or
a
fourth
way
of
a
building
models
of
the
world.
These
are
the
two
basic
core
things,
and
so
this
should
go
a
long
way.
We
can
figure
this
out
if
we
can
do
it,
it's
kind
of
its
kind
of
have
a
lot
of
legs
to
it
over
time.
It's
a
very
powerful
idea
that
you
have
you
know,
but
but
again
you
can
build
video
games.
It
can
build
robots.
You
could
you
know,
you
know
things
search.
E
A
Glenn
as
a
question,
we
chat
about
the
quite
quite
an
ounce
of
thousands
of
Sasuke
/,
so
this
good
form
to
join
him
on
that
tree
there.
It's
pretty
specific,
so
understand.
If
no
one
remembers
up
top
of
your
head.
The
paper
ends
within
HTM
curve
versus
a
first
order
model
and
the
system
of
fifty
percent
patterns
and
fifty
percent
noise.
Do
you
remember
how
many
different
patterns
there
were
mixed
in
with
the
noise?
Was
it
one
pattern
and
then
one
pattern
changed
during
the
transition.
E
F
So
that's
the
cook,
so
the
code
for
that
is
actually
a
new
pic
research.
Anyone
can
take
a
look
at
that.
Basically,
it's
a
it's
a
sequence
where
you
have
six
elements
in
a
row
and
then
four
elements
of
noise
and
then
six
elements
in
a
row.
So
you
know
to
see
Quentin,
oi,
sequence,
noise
and
I.
Think
there
were
trying
to
remember
I
think
there
were
like
six
different
sequences
or
10
different
sequences:
I,
don't
I,
don't
I
thought
it
was
eight,
but
but
I,
don't
I
didn't
do
the
work.
I
thought!
F
E
D
F
F
E
A
D
A
E
Were
I
thought
the
lessons
confuse
me,
I
mean
cicadas
just
do
involve
physical
movement
right.
That's
a
physical
moment.
That's
that's
equivalent
to
moving
your
fingers
over
something
right.
You
you're
moving
in
your
sensor,
you're
moving
your
sensor
over
an
object
and
the
way
we're
thinking
about
it
now.
Is
you
don't
think
it
is
one
thing
you
think
about
the
sensor.
As
the
surgeons
columns,
you
know
series
of
little
pieces.
You
know
the
retina
would
work,
even
if,
even
if
it
was
like
lots
of
little
holes,
you
know
it's
like
you're.
E
Yeah
and
you
see
anyway,
so
the
point
is
yes,
so
the
rent
is
just
like
moving
your
fingers
and
and
you're
sensing
different
parts
of
the
object,
and
that's
a
motion
to
in
this
theory
says
you
have
to
do
that
to
learn.
You
cannot
infer
something
until
you've
touched
it.
Many
many
different
ways.
We
looked
over
at
many
many
different
ways
after
you
learned
that
you
can
sometimes
infer
in
a
single
touch
or
a
single
vision
of
single
glance.
E
E
A
M
E
Go
back
to
her
soon,
as
we
have
some
progress
on
that
will,
let
you
know
I
mean
I'm
an
impatient
to
make
progress
on
this.
You
know
we're
talking
about
it
every
week,
a
couple
times
a
week.
It's
the
only
thing.
I
really
want
to
work
on
right
now,
so
I
think
Marcus
is
working
on
it
too,
but
I
was
working
on
it.
A
A
F
F
You
know
you
could,
in
temporal
pooling
over
a
sequence
of
notes
and
a
melody,
for
example,
allows
you
to
generalize
and
recognize
a
particular
song,
and
maybe
variations
of
that
of
that
song
as
well.
So
this
the
encoding
is
part
of
it,
but
it's
not
the
whole
story.
There's
many
other
aspects
of
translation
yeah,
we
don't
understand,
that's
right
and
the
spatial
cooler
does
generalization
to
it,
figures
out
which
bits
in
the
SDR
to
ignore
and
that
becomes
more
noise.
More
robust
annoys
as
a
result,
in
that
you
understand.
A
E
Oscillation
frequencies
in
the
vent
and
gamma
frequencies
and
theta
frequencies
are
two
of
the
examples
and
they
clearly
play
an
important
part
in
how
the
biological
tissue
works.
The
moment
we
don't
incorporate
them
in
HTM
three
and
we're
at
the
moment.
We
don't
have
any
evidence
that
they
are
actually
information,
theoretic
components,
the
our
current
believe
and
it
could
change-
is
that
these
rhythms
are
ways
for
neurons
to
behave
together
properly.
E
So
in
HTM
theory
we
require
that
a
whole
bunch
of
synapse
has
become
active
at
the
same
time
in
the
biology
they
have
to
become
iconic,
dendritic
ranch,
that's
the
that's
the
key
part
of
this
temporal
memory,
algorithm
and
the
active
dendrites
and
in
biology
those
synapses
have
to
become
active
within
a
few
milliseconds
of
each
other,
which
is
very,
very
tight
constraint.
We
don't
have
that
constraint
in
software,
so
maybe
you
know
some
here's
an
example.
E
Maybe
some
of
those
oscillations
are
just
to
make
sure
that
the
neurons
are
firing
at
the
same
time
so
that
they
actually
work
on
active
dendrites,
but
if
we
can
relax
that
constrain
and
software
and
we're
not
actually
modeling
the
when
I,
not
Mac
modeling
the
ion
channels
in
a
dendrite
and
therefore
we
don't
have
this
constraint
of
you
know
within
three
milliseconds
or
something
like
that.
So
that's
kind
of
example
where
it
may
be
absolutely
required
in
biology.
E
A
F
We've
been
working
with
a
dataset
from
another
organization
that
unfortunately
can't
share,
but
we've
been
exploring
various
ways
of
using
the
cortical
I/o
encoding
to
you
know,
figure
out
which
documents
are
more
similar
to
the
other
and
so
on
and
and
overall
I
think
the
cortical,
I/o
and
coatings
work
pretty
well,
and
you
know,
there's
not
a
whole
lot
to
talk
about
their
I.
Think
one
of
the
one
interesting
aspect
of
that
is
that
for
that
Marian
had
actually
built.
F
Part
of
this
is
sort
of
you
can
think
of
a
paragraph
as
a
sequence
of
words
and
then,
when
you're
trying
to
recognize
similar
paragraphs
you're
in
some
sense
doing
classification
on
a
sequence
of
words.
So
it's
a
type
of
sequence
classification,
so
Marion
is
built
up
a
nice
little
software
framework.
F
That's
in
new
pic
research
for
doing
sequence,
classification
and
ua
has
also
been
doing
a
bunch
of
experiments
on
not
specific
to
NLP,
but
just
how
to
use
HTM
for
doing
sequence,
classifications
and
we'd
be
talking
about
putting
some
of
that
code
potentially
into
new
pic
even
soon.
So
that's
that
that
code
is
in
new
pic
research
as
well.
So
here's.
A
A
D
B
A
B
E
Absolute
proof
that
the
transforms
happening
and
as
I
said
there
is
no,
we
don't
not
aware
of
any
theories
about
how
does
being
implemented
in
the
neurons
yeah
so,
which
is
something
by
the
way.
You
know
the
GPS
encoder
we
did,
which
we
did
before
you're
thinking
about
play
cells
and
grid
cells.
This
is
very
similar.
It
was
with
a
brace
it's
almost
the
dynamical
type
of
representation.
So,
and
you
know
we
made
that
upsets
and
that's
another
good
evidence
like
we
thought
about
that.
E
So
we
understand
that
encode
it
pretty
well
and
how
those
representations
work.
So
we
think
something
like
that's
being
used
here
and
you
can.
You
can
talk
about
the
properties
but,
as
you
a
said,
we
don't
really
know
how
it's
how
would
be
implemented
in
neurons.
So
we
need
to
figure
that
out.
Yeah.
Oh.
E
A
A
N
So
yeah
I
guess
the
question
is:
what
are
the
improvements
that
can
have
first
Python
of
being
list?
I,
don't
know
if
that
itself
has
lunch,
something
we
found
beneficial,
maybe
possibly
in
cortex
I'm
more
experienced
with
cortex,
which
is
Felix's
and
implementation.
It
has
this
useful
thing
called
immutability
and
which
tends
to
make
things
inherently
visualize
abode
so
that
you
bring
up
sanity,
which
is
the
visualization
project
for
this.
It
kind
of
goes
hand
in
hand
with
vortex,
and
so
that's
one
of
the
fundamental
things
that
has
other
things
is
really
cool
for.
F
N
That's
another
yea
and
I
brought
up
immutability
as
good
for
visualizations,
but
it
can
also
be
good
for
you
can
just
try
new
things
that
you
can't
really
try
elsewhere
you
they
take
a
lot
more
work.
For
example,
if
you
want
to
compute
the
anomaly
score
for
a
ton
of
inputs,
there's
not
really
a
clear
way
to
do
that
with
HTM,
because
every
time
you
every
time
you
give
it
a
new
in
but
you're
going
to
change
it
in
place.
N
A
C
L
F
I
mean
part
of
this
part
of
this
new
idea
that
Jeffrey
talk
about
is,
if
you
have
these
sort
of
cortical
columns,
each
cortical
column
may
be
sensing
part
of
an
object,
let's
say
think
about
fingers
and
a
hand
could
be
touching
a
coffee
mug
and
each
finger.
It's
touching
a
different
part
of
an
object
from
any
one
finger.
You
may
not
be
able
to
identify
what
the
object
is,
but
if
you
touch
several
places
at
once,
you
may
instantly
be
able
to
identify
it.
F
So
that's
sort
of
like
having
multiple
things
which
have
portions
of
patterns
or
maybe
unions
of
patterns,
and
then,
together
they
kind
of
collaborate
and
figure
out
what
is
the
most
consistent
and
unique
representation
for
the
particular
sensations
that
you're
getting,
and
so
that
has
a
lot
of
similarities
with
some
of
the
associative
memory
networks
that
have
been
out
there
and
so
we're
kind
of
under
that
context,
we're
just
looking
at
some
of
the
literature
and
playing
experimenting.
One.
That's.
E
E
You've
got
multiple
ones
at
the
same
time,
almost
in
region
right
well,
not
there
in
the
same
region,
but
the
point
of
using
fingertips
are
not
necessarily
immediately
adjacent
to
each
other
I.
You
know
I
kept
done
that
to
beef.
It
could
be
just
different
parts
of
my
hand.
So
the
point
is
it's
not
like
at
the
red
I
could
fool
you,
because
you
think
it
already
to
always
get
this
picture
right.
Well,
when
you
think
about
the
touch
fingers
come
to
touch
coming
from
your
fingers,
it's
not
a
picture
anymore.
E
E
It's
like
basically
saying
I.
Have
this
there's
aught
associative
memory,
I
have
partial
inputs
to
each
different
part
and
each
part
is,
is
each
column
a
sort
of
ambiguous,
but
then
the
thing
settles
all
at
once
on
the
right
answer,
which
is
what
auto
associated
members
do,
but
they're
not
not
typically
distributed
like
that.
They're
not
typically,
should
really
like
these
different
columns
that
are
communicating.
That's
usually
like
you
just
get
partial
input
and
that
it
resolves
it,
but
the
same
principles
are
applying
there.
E
A
E
E
K
E
E
Magical
from
now
it's
not
so
we
got
a
lot
of
neural
machinery
in
that
region
and
far
more
than
you
get
between
Rachel,
and
so
we
want
to
understand
the
hierarchy
we
have.
We
have
theories
about
what
it's
doing,
but
you
know
you
can
make
a
useful
system
with
a
single
region.
We
really
want
to
try
to
do
that.
So.
A
K
L
F
F
Mean
from
the
software
infrastructure
standpoint,
it's
actually
the
network
API
supports
all
this
already
pretty.
Well,
so
implementation
wise,
it's
pretty
straightforward,
algorithmically
know
what
is
the
information
that
has
to
flow
across
these?
You
know
cortical
columns
and
how
do
they
resolve
down
to
the
unique
representation?
Well,
that's
some
of
the
stuff
we're
trying
to
figure
out
and
the
auto
associative
memories
might
give
us
some
clues
into
it,
but.
A
F
Yes,
some
of
the
so
one
of
the
really
interesting
things
is
the
temporal
memory
structure
and
the
spatial
pooling
structure
is
is
repeated
across
the
layer.
So
all
of
the
intuitions
that
we
have
from
temporal
memory
really
helped
us
in
coming
up
with
understanding
these
new
ideas
and
involving
them.
So
we
we
are
definitely
building
on
top
of
the
existing
structure,
some
of
the
new
things
they
were
adding
here
as
one
is
this
concept
of
apical
dendrites,
which
is
another
type
of
active
dendrites
and
those
take
typically
take
input,
feedback,
input
from
layers
or
regions.
F
Above
so
that's
something
that
we're
adding
to
the
system
and
then
the
other
big
new
thing
that
we're
playing
around
with
is
the
idea
of
these
lateral
connections
that
convert
on
a
unique
representation.
So
these
are
all
kind
of
generalizations
of
the
infrastructure
that
we
already
have
algorithmically.
L
F
E
We're
sticking
a
lot
of
the
basic
stuff
I
mean
well
one
way
to
think
about
what
do
we
wear?
The
couple
things
were
adding
well
we're
adding
this
idea
of
multiple
vertical
columns
that
are
communicating
across
long
distances,
which
we
know
it's
just
the
frame
and
then
we're
also
adding
more
layers
to
the
cortical
column.
So
to
do
the
transformation
I'm
also
working
on
layer
5
is
what
it
does.
So
those
are
the
sort
of
two
dimensions
but
they're
all
built
on
STRs
all
built
on
parameter
ons.
They
all
have
many
columns
inhibition
similar.
E
E
Memory
hugely
so
we
have
to
give
her
this
to
touch
of
scaling.
Pumps
was
the
one
user
furture,
which
is
a
practical
one
like
hey.
We
need
to
build
a
bigger
system
and
it's
running
through
slowly
there's
a
theoretical
scaling
issues
we
have
to
avoid,
which
are
you
know,
is
this
just
we
gonna
work
when
just
is
it
possible
in
a
physical
universe,
to
build
something?
That's
big
and
so
like?
E
We
did
some
work
two
summers
ago
with
sensorimotor
inference
a
first
attempt
that
this,
where
we
didn't
really
think
about
this
coordinate,
transform
and
that
didn't
scale
well
I
mean.
Has
the
reticle
scaling
problems,
not
practical,
scaling
problems?
So
we
have
to
make
sure
that
the
theories
don't
have
the
reticle
scaling
problems
and
then
then,
then,
after
that
is
engineering
and
the
viewers.
A
F
D
A
K
E
A
E
E
They
were
interested
in
understanding
how
you
know,
keeping
a
little
bit
deeper
into
our
algorithm.
Talking
about
how
to
accelerated.
I
think
the
conclusion
from
that
was
it's
more
promising
than
we
regionally
thought,
but
you
really
have
to
spend
some
time
working
on
it
to
find
out
what
GPUs
could
do
for
us.
So
was
that
a
good
summary
yeah.
A
Money:
okay:
let's
see
one
last
question,
we
only
have
about
a
minute
I,
really
several
questions
of
this
kid
named
matlind.
He
says
in
the
new
model
are
the
sensor
hit
was
from
different
figures
fed
into
the
same
region
or
each
been
in
two
different
fingers.
Fingers
are
different
region,
so
we
did
ask
that,
but
from
most
likely
the
same
region,
but.
E
This
is
all
old
papers,
I
forget
who
did
them,
but
they
talked
about
the
mapping
of
the
the
skin
to
the
to
the
primary
central
cortex
and
they
need
those
that
figure
the
really
distorted
figure
of
a
monkey
map
onto
a
swan.
If
you've
seen
this
fake
famous
up,
you
know
who
made
that
paper,
you
I
do
I
I.
F
E
All
laughing
anyway,
you
just
look
at
homunculus
searching
month
with
s1.
You
almost
certainly
will
find
it
and
was
interesting
about
is
he's
like
okay,
so
the
whole
body
is
represented
in
s1,
but
it's
really
distorted
just
like
the
regular
distorts
v1,
it's
really
distorted
by
the
area.
The
amount
of
sensory
bits
coming
in
skin,
so
your
fingers
and
your
lips
have
a
lot
of
sensory
bids,
and
so
the
lips
get
really
blown
up
on
this
one.
The
finger
the
hands
are
really
weirdly
distorted.
So
that's
really
odd.
Looking
figure
the
interesting
thing
about
it.
E
So
it's
a
good
it's.
Those
papers
are
probably
34
years
old,
but
they're
still
ballot
and
that's
a
big
clue
and
ones.
Why
I'm
thinking
about
cymatics
the
touch
it's
easy
to
get
around
some
of
the
things
you
naturally
think
about
vision
which
misleading?
Oh,
that
man
must
answer.
Hey
so
homunculus.
It
says
how
much
villains,
okay.