►
Description
Broadcasted live on Twitch -- Watch live at https://www.twitch.tv/rhyolight_
A
All
waiting,
so
so
that
there's
like
they,
this
is
a
common
term,
but
it's
also
so
microscopic
they're,
making
the
distinction
in
microscopic,
which
is
like
a
single
cellular,
recording
and
macroscopic,
which
is
fMRI
type
things
and
then
mesoscopic
is
in
between
well.
Is
it
and
so
they're,
specifically
in
this
case
the
making?
A
The
point
that
a
lot
of
this
fall
has
to
do
is
everything
we
talked
about
in
terms
of
bridge
salt
place,
those
on
the
effector
cells
and
so
on
and
they're
saying,
okay,
a
lot
of
this
work
has
been
done:
singer's,
cellular
recording
level
and
now
there's
a
bunch
of
literature
about
the
fMRI
level.
But
we,
these
researchers,
are
each
of
protocol.
Eeg
court
under
the
skull
into
protocol.
Eeg
is
imagine
what
local
field
potentials,
not
individual
self,
but
voltage
potential
group,
small
groups
of
cells
they're.
A
Actually,
the
resolution
that
they're
recording
from
and
these
in
their
world
is
getting
close.
It's
very
vocal.
We
started
a
lot
with
fMRI.
Some
seven
hundred
microns
over
a
few
part
of
a
millimeter
and
fMRI
is
getting
down
there.
But,
of
course,
with
the
EEG
which
you
get
you
get.
This
high
temporal
resolution
very.
B
C
A
B
A
They
talked
about
some
of
these
are
like
just
a
surface
implanted
like
like.
Instead
of
going
through
the
skull,
the
right
on
the
bias
to
surgery
or
monkey,
they
are
a
few
that
are
updating
planted,
but
they're
not
doing
any
singles
Pikesville,
it's
just
all
local
field
potential
like
they're
doing
frequencies
in
waves.
A
A
A
A
A
A
And
then
here's
how
people
get
it
at
the
fMRI?
These
are
experiments
you
could
refer
to
in
a
variety.
The
exact
same
equivalence
and
play
cells
or
cells
in
grid
cells
also
gives
us,
and
now
here's
how
we
getting
this
up
on
the
net.
With
these
oscillations
and
these
of
different
figures
for
our
these
papers,
so
the
tri-state
I,
can
we
into
playing.
This
is
all
part
two.
A
A
A
Section
dedicated
to
this
issue,
why
you
see
innocent
directional
activities
and
we
are
classes
a
lot.
Where
did
this
come
from
one
of
the
different
theories?
So
here
they're
saying
you
know,
here's
a
different
uses
like
in
a
microscopic
world
of
you
know
single
grid
cells,
and
you
know
in
one
direction
you
can
you
have
people
longer
before
we
did
to
repeat
when
you're
on
the
shorter
direction
you
go
this
way.
Why
is
it?
We
see
less
bold
activity
when
we
go
this
direction
versus
this
direction?
A
C
A
Particularly
great
party
Senate's,
both
let's
topic
good,
like
reputation
metropolis,
if
itself
would
be,
they
are
basically
saying
we
need
to
do
this
across
the
entire
brain.
There's
evidence
that
this
stuff
is
occurring
in
all
different
places
that
we
need
to
look
resolved
and
we
saw
five
things.
We
can
do
that
with
our
our
kind
of
technology,
with
our
representations
outside
of
the
veronik
projects
we
perform
complimentary,
computation
for
spatial
navigation
or
they
gave
or
other
cognitive
functions
that
exploit
similar
computational
principles,
so
our
papers
back
right,
that's
what
we're
proposing
it
is.
C
A
To
get
a
little
shout
out
there
well
they're,
basically
saying
you
know,
they're
touching
on
all
these
throughout
this
paper.
They're
talking
like
these
are
the
common
principles
off
the
brain
related
to
movement,
so
they're
your
scale,
the
related
to
either
what
you
move.
A
lot
of
stuff
I've
talked
about
with
how
the
data
frequency
could
impact
scale.
A
Behaviour
is
supported
in
this
paper.
So
there's
a
lot
of
evidence
for
that.
So
I
felt
like
this
paper
sort
of
ties
together
a
whole
bunch
of
the
things
we've
talked
about
using
new
techniques
we
haven't
discussed,
but
not
really
introducing
any
new
ideas,
just
more
evidence
and
more
sort
of
precise
definitions
between
a
prison.
A
A
A
C
A
A
A
A
B
A
A
I
didn't
talk
about,
you
know
mostly
I
think
the
way
they
knew
these
things
I
his
impression
is
they
localize
with
where
the
frequency
is.
You
know
some
part
of
the
cortex
or
whatever,
and
that's
how
the
okay
I
mean
happy
Terrace
is
going
to
be
in
this
category
is
opposed
to
it's
more
like
they're
looking
forward.
Where
are
these
frequencies
are
very
vocally
localized
and
that's
the
finger
focused
on
as
opposed
to
that
looking
at
a
different
way.
So
yeah
they
didn't
say.
Let's
talk
about
now.
A
A
B
A
We
haven't
really
had
a
solid
theory
about
it
and
I
work
and
we've
had,
but
we've
recently
started
thinking
about
how
to
correlate
no
I
can
take
it
off.
Like
you
go
yes
question
to
that,
then
it's
been
two
basic
areas.
One
is
awesome.
The
difference
like
the
different
phases
of
certain
oscillations,
appear
to
being
related
to
learning
and
inferences,
so
that
solves
some
of
our
learning.
Inference.
A
Problems
that
were
all
simultaneously
looks
like
you
might
actually
have
two
different
cell
populations
are
awfully
me
back
and
forth
between
learning
and
inference
day
and
the
other
area
that
just
briefly
mentioned.
Is
there
the
idea
of
these
oscillations
using
the
thalamus
and
the
cortex,
which
are
prevalent
and
well
study,
but
have
no
threat,
achill
sort
of
theory
behind
them
and
I?
A
Did
that
I'm
not
sure
completely
generated
myself
from
some
of
us?
I,
don't
remember,
but
I've
been
running
with,
is
the
idea
that
that
frequency
is
that
there's
some
models
of
grid
cells
which
are
based
on
that
the
progression
of
the
the
bump
in
the
grid
cells
is
based
on
an
oscillatory
frequency
and
and
that
also
at
write
frequency
represents
the
rate
of
movement.
That's
one
of
two
major
classes.
A
Defense
helps
and
said:
Papa
C's
that
that
I've
been
working
on
recently,
which
I'm
pretty
excited
about
is
that
this
frequency
of
oscillation,
between
cortex
and
thalamus
is,
is
essentially
it's
controlling
scale,
but
it's
going
to
control
scale
of
distance
and
size
which
are
related
so
like
going
to
move
my
item
you're
right.
Alright,
how
much
I
move
depends
on
for
where
you
are
so
by
actual
frequency,
we
decode
your
distance
and
we're
not
to
encode
how
much
a
movement
right
Lou.
A
A
A
E
Do
what
one
just
tiny
addition
on
this
paper
I
read
part
of
it,
because
I
was
curious
to
know
on
their
theories
of
when
they
lay
out
the
reason
why
the
hex
directional
signal,
Trivium
cells
might
be
there
one.
One
thing:
I
talked
to
one
of
the
authors
of
this
paper
and
and
one
thing
that
they
don't
say
in
this
paper,
because
you
wouldn't
want
to
say
in
a
paper
because
it
might
be
overruled
it's
yeah,
it
might
might
not
age.
E
E
So
the
a
bit
of
information
that's
hard
to
get
it's
hard
to
learn.
Unless
you
talk
to
people
is
multiple
labs
have
had
trouble
replicating
that
result.
Multiple
labs
have
had
trouble,
replicated
replicating
the
the
idea
that
these
conjunctive
head
Direction
driven
cells,
are
not
evenly
distributed.
The
idea
that
they're
kind
of
at
the
peaks
of
a
hexagon
or
nebula
multiple
people
try
to
find
that
available
and,
and
basically
that
that's
anyway.
A
A
C
D
E
Different
things-
and
that's
that
was
part
of
the
motivation
for
the
dole
lab,
which
is
like
the
fMRI
people,
and
basically
they
moved
to
Norway
to
where
the
most
or
lapis
to
collaborate
in
part.
For
that
reason,
they
wanted
to
figure
out
where
the
heck
this
exa
directional
signals
coming
from,
like
that
was
one
of
their
main
motivations
and
driving
forces.
You.
A
A
E
E
E
A
E
A
Can
summarize
what
I
tell
the
story
got
this
shirt
god
so
I
sent
our
frameworks
paper
to
jump
in
prior
to
submitting
them
and
you're
thinking
hey.
You
know
this
is
so
much.
What
are
you
doing
with
that
capsules?
And
you
know
maybe
I
just
whatever?
Maybe
you
can
suggest
something,
and
so
he
responded
right
away,
but
with
almost
no
text,
I
think
his
email
said
read
this
this
paper.
That
was
it
being.
E
C
A
E
B
E
A
E
C
E
Because
because
he
drew
in
the
crown
so
that
you
can
see
the
crown,
but
if
you
only
pay
attention
that
these
solid
black
lines
ignore
the
crown
underneath
the
point
who
is
making
is,
you
can
also
see
kind
of
a
zig
zag,
where
this
is
a
surface.
This
there's
this
little
rectangle
right,
here's
the
surface!
Lately
it's
like
a
piece
of
paper
with
two
folds
in
it,
but
it
doesn't
illustrate
this
very
well
anyway.
C
A
E
So
I
mean
he
goes
into
I,
don't
have
much
to
say
from
this
paper.
It's
it
sort
of
just
describes
a
lot
of
what
we
have
to
say.
Yes,
his
own
notation
of
how,
if
you
see
the,
if
you
can
see
within
your
viewer,
centric
reference
frame
where
component
a
is
like
what
is
the
relationship
between
like
object,
a
and
your
and
your
viewer
centric
frame,
and
if
you
have
that
for
object,
B
or
or
with
a
parent
object,
etc.
C
E
So
this
one
shape
representation
and
parallel
systems
is
really
the
most
hardcore
of
all
these
papers.
He
just
goes
into
topic
after
topic.
The
first
couple
pages
I
would
suspect
we
would
find
you
would
find
like
I,
don't
know
if
this
is
relevant
and
then
then
he
gets
the
meat
of
it
and
suddenly
every
part
of
its
relevant,
so
I'll
scroll
down
past
this
first
ones,
where
he's
talking
about
early
visual
processing
and
instead
move
on
Xhosa.
E
Finally,
the
first
thing
that
I
highlighted
before
shape
constancy-
oh
so,
a
few
people
in
the
room
know
that
many
many
times
I
have
driven
I've
drawn
something
on
the
board,
which
Jeff
calls
the
flux
capacitor.
That
involves
a
population
of
cells
representing
some
kind
transformation
between
two
other
populations
and
causing
kind
of
a
routing
between
them,
and
then
there
needs
to
be
some
mechanism
for
that.
In.
C
E
E
E
C
E
So
another
thing
that
we've
talked
about
a
little
bit
that
gets
to
something
I've
been
talking
about
lately,
others
feed
this
sense
he's
connecting
us
to
Gestalt
psychology
and
he
interprets
it
a
little
bit
differently
to
what
he
thinks
is
going
to
be
true.
With
these
neural
network
models
of
object,
recognition
so.
F
E
We
attend
to
a
whole,
we
do
not
see
its
parts
as
holes,
because
the
representation
of
the
whole
does
not
in
any
way
involve
or
require
the
representation
of
the
parts
of
schools
that
might
be
a
little
producing.
The
second
second.
Second
sense:
it's
better
when
a
part
is
seen
as
a
constituent
of
a
larger
whole.
E
It
is
given
a
quite
different
internal
representation
from
the
one
that
has
when
it's
seen
as
a
whole
and
its
own
right
is
a
high
spot
like
RIT,
oh
yeah,
thanks
a
right
mess,
so
I've
been
thinking
lately
about
the
topic
of
compositional
objects
and
recognizing,
like
higher-level
objects,
without
having
to
drill
down
to
like
its
lowest
level
components,
and
he
shows
a
picture
here.
That
I
think
is
a
useful
example
in
this
next
page,
so
I'll
making
an
expedition
no
coming
the
wrong
paper.
E
This
in
the
next
paragraph,
but
I'll
go
ahead
and
I'll
kind
of
go
through
these
papers,
like
a
treat.
So
welcome
back
to
that
one.
So
the
example
he
uses
to
illustrate
this
is
when
you
first,
when
you
first
see
this
in
passing,
you
probably
just
a
messy
face,
then
you
look
closer
and
realize
that
wait.
E
Was
a
banana
area
of
watermelon
apples
and
there
Casey
makes
is
that
when
you
see
this
you're
not
building
up
from
the
fruit
to
the
face
you're,
you
kind
of
when
you
see
it
from
the
top.
When
you
see
the
whole
thing
you
see
the
face
and
then
you
can
attend
to
the
individual
pieces
and
suddenly
you
see
the
fruit,
but
there's
something
nice.
A
D
It's
it's
kind
of
somebody's
making
the
point
that
the
representation
of
this
object
we're
not
representing
as
a
composite
object
of
specific
things.
Like
proofs.
We
have
a
relationship
between
generic
things
that
makes
up
a
face
and,
oh
by
the
way,
those
generic
things
could
be
bananas
or
pears,
yeah.
A
Yeah
I
want
you
to
learn
this
object.
I
will
see
them
instantly
and
I
were
part
of
one
novels
about
the
object.
Is
that
there's
two
hours
of
their
opinion
to
that
representation?
Is
there
as
a
composite
object,
although
when
I
first
saw
it
I
didn't
have
that
repetition
with
more
some
finishing
shapes
of
some
sort?
Is
that
my
interpreting
that
the
way
there
is.
B
A
E
E
E
Encoding
multiple
multi-dimensional
features-
I
will
talk
about
this
now
so
like
he
talks
about
in
through
the
earlier
part
of
the
paper.
He
was
using
single
units
for
everything,
but
then
he
says
he
talks
about
abandoning
the
naive
idea
that
each
specific
feature
is
represented
by
activity
and
exactly
one
unit.
Instead,
each
unit
can
be
more
coarsely
tuned
tuned
so
that
it
is
activated
by
a
range
of
possible
features
and
the
ranges
of
different
units
can
be
made
to
overlap
so
that
each
feature
activates
many
different
units.
E
Through
the
it's
okay
so
far,
I've
been
focusing
on
areas
where
he
came
to
the
same
conclusion
as
us,
but
in
this
paper
I
thought
it
was
noteworthy
some
something
where
he
says.
Something
then
I
would
say:
I,
don't
know
if
he's
right
that
we
might
agree
with
him
some
days
other
days.
We
wouldn't
we'll
see
so
I'll
talk
about
it
here
when
he
talks
about
perception,
grow
people
moving
the
eyes
around
building
up
an
image
or
whatever
inference
process
is
going
on.
E
As
you
move
your
eyes
around,
he
jumps
to
the
idea
that
I'll
read.
The
second
part
is
what
he
thinks
is
the
answer
there.
What
happens
that
might
be
read
the
Woolsey,
so
what
is
needed
is
a
way
of
representing
where
the,
where
the
pieces
are,
that
is
not
affected
by
eye
movements
or
even
by
movements
of
the
whole
person
through
space,
so
well
sort
of
you,
here's
where
I'm
confused.
This
is
a
paper
on
spatial
working
memory
of
remember,
remembering
like
what's
out
there
any.
Where
is
it?
Where
is
it
and
he,
but.
E
Right,
we
would
definitely
agree
with
this
when
it
comes
to
the
learning
object,
yeah,
but
he
says
so
when
I
was
reading
this,
how,
when
I
was
reading
his
other
work,
I
was
like
wait.
Why
is
he
jumping
with
this
conclusion
that
careless
read
the
signs?
Oh,
it's
just
here's!
Where
he's
he's
saying
something
that
I
think
is
plausible,
but
he
says
it's
very
unlikely.
E
He
says
it's
just
as
conceivable
that
as
we
move
our
eyes,
the
internal
records
of
all
the
previously
perceived
place
pieces
are
correspondingly
altered
so
that
the
records
always
encode,
where
the
piece
is
relative
to
the
current
retinal
position,
but
this
seems
very
unlikely.
This
paper
was
written
before
psychotic
remapping
is
written
before
object,
effector
cells.
It
was
written
before
all
these
things
that
actually
resembled
what
he
what
he
is.
A
Saying
but
the
phrases
of
eternal
records
are
our
correspondent,
Walter
yeah.
That
seems
odd,
maybe
I'm
just
going
the
word
in
there,
I
mean
when
I
think
about
psychotic
remapping,
for
example.
It's
not
like
a
record
of
being
Walter,
it
is
it's
more
I.
Have
one
I
have
this
three-dimensional
map
and
or
some
N
dimensional
map
and
I'm
moving
to
a
new
location?
Therefore
I
can
my
inputs
are
changing,
but
it's
not
like
I'm
I,
don't
believe
the
idea
that
I'm
moving
something-
and
you
know
I'm
saying
it's
not
like
they're
altering
something.
C
A
E
E
Throughout
the
rest
of
this
relic,
oh
okay,
this
is
just
linked
to
the
rest
of
his
writing
and
he
never.
This
is
the
only
time
he
considers
the
idea
that,
instead
of
maybe
what
you
do,
is
you
keep
track
of
like
okay,
there's
a
chair
over
there
there's
a
chair
over
there
and
as
I
move
around
I
update
those
egocentric
locations
of
those
chairs,
which
I
would
argue
optic.
Effector
cells
essentially
do
all.
A
E
A
B
A
The
co,
the
other
part,
this
requires
you
to
sort
of
requires
a
thousand
brains
there
right.
It
basically
says
you
know,
there's
one
thing
to
say:
I
have
a
you
know:
I
have
a
reputation
of
all
these
locations.
How
can
it
all
be
updated?
But
if
you've
got
thousands
of
the
same
module
all
peak
on
one
location,
then-
and
it
seems
more
plausible,
so
that
it
would
also
so
maybe
getting
the
cue
point
during
about
the
computational
efficiency.
A
It
remind
me
that
this
is
only
possible
because
we
have
thousands
of
columns
were
all
doing
in
some
sense,
and
so
it's
not
like
you
have
to
go
through
Sara,
Lee
or
think
about
it
in
massive
scale.
I'll
take
all
these
things
at
once,
I
imagine
doctor
knees
day.
All
the
research
all
were
meeting
I
didn't
these
in
this
field
in
neural
networks.
Back
then
was
there
was
no
concept
of
anything
like
that.
It
was
always.
A
A
A
E
E
His
view
of
what
does
Jeff
have
you
ever
seen
such
that?
What
does
he
think
of
it
happens?
And
it's
interesting
to
consider
the
idea
that,
like
as
you,
walk
around
the
world,
you're,
always
choosing
a
a
rough
frame
of
reference
for
the
scene
and
like
if
you
see
a
stop
sign
over
there,
you
put
that
somewhere
in
your
current
scene
and
then
then
it
stays
still
as
you
move
and
you're
walking
around.
And
then
you
periodically
change
scenes.
You
periodically
choose
a
new
fragrance,
but.
A
Okay,
so
well
I,
think
about
that
when
you
move
through
the
world,
there's
juice
I
get
one
of
them:
I
didn't
even
the
world
of
static,
walk
to
each
other,
but
then
your
perception,
where
they
are
relative
view,
is
everything's
changing
constantly.
You
know
I'm
right
in
my
pocket
or
every
little
piece
out
there
at
the
position
and
it's
flowing
past
me
citizens.
You
know
constantly
in
it
I'm
aware
of
all
that.
That
just
tells
you
that
the
brain
is
doing
is
that.
A
A
E
E
The
idea
that
that
our
brain
learns
a
set
of
primitive
features,
really
well
and
then
builds
things
out
of
them
and
such
that,
once
you
once
it's
learned
every
once,
it's
learned:
how
do
what
something
looks
like
from
all
different
viewing
angles
they
can,
then
they
can
then
recognize
objects
can
posted
those
from
multiple
viewing
angles
and
he
and
I
specifically
use
the
word
primitive
features
when
I
was
talking
about
this,
so
he
this
kind
of
recursive
definition
of
a
shape
in
terms
of
shapes
of
its
parts
leads
to
Rodriguez
the
only
term
names,
a
hypothetical
primitive
features,
then
the
fruit
face.
E
E
A
E
C
C
E
E
You
what
you
basically
have
done
is
you
created
this
for
current
network
that
will
settle
if
one
of
these
is
right,
it's
going
to
cause
one
of
these
objects
to
be
recognized,
which
is
going
to
cause
everything
to
settle
down
into.
Basically,
it's
a
unions,
narrowing
approach
of
whichever
one
of
these
transforms
leads
to
successfully
recognized
out
that
there's
going
to
win
out
in
this
competitive
battle.
I,
don't
know
if
I'm
being
very
clear.
I
was
trying
to
be
quick
with
this,
because
it
really
is
it's
an
idea
of
I.
D
E
D
E
C
E
E
Rather
than
at
the
same
time,
we
can
talk
about
that.
This
wasn't
the
main
point
of
this
main
point.
This
is
going
to
be
when
he
talks
about
what
he
calls
the
N
squared,
probably
when
he
talks
about
the
idea
that
these
transformations
these
displacements,
whatever
would
I
call
them
they.
They
require
kind
of
this
everything
to
everything.
Mapping
you
need
to
be
able
to
map
any
anyone.
E
He
brought
up
something
that
was
kind
of
doctor.
Looking
for
me,
I
don't
know
if
this
is
gonna
work,
but
it
might
be
a
really
good
idea.
So
first
I'm
now
reading,
first
n
itself
and
being
like
the
number
of
okay,
the
N
squared
problem.
How
do
you
reduce
entered
first
n
SL
can
be
dramatically
reduced
before
the
mapping
by
distributed
encoding
of
the
individual
features
and
mappings
as
patterns
of
activity
and
many
different
hardware
units.
So
now
I'm
going
to
bring
back
the
station
for
thing
and
now
I'm
gonna
started
playing
the
recording.
E
Yeah
so
Ohio
idea,
one
afterwards,
since
we're
are
young
this
so
once
again,
I
don't
know
that
this
I
don't
have
it
and
working
version
of
this
in
my
head.
But
the
fact
that
into
developing
writing
about
it
and
said
it
worked,
and
I
did
it
is-
that
is
intriguing
so
I'm
talking
about
now,
but
it
might
under
program
so
yeah
transforming
between
spatial
cooler
style
representations.
C
C
E
C
C
E
E
E
It
can
encode
an
immense
number
of
features
because
it
had
course
post
them,
so
can
it
can
it
can
take
a
novel
feature
and
transforming
it
into
photography
while
I'm
jumping?
Oh,
you
can
take
it
a
novel
feature
and
representative
somehow
and
the
thought-provoking
idea
here
is
one
if
you
could
use
a
similar
type
of
idea
to
transform
those
two.
C
A
So
the
first
thing,
first
again
sometime,
oh
what
a
great
idea
and
the
thing
I
found
really
attractive
about
it,
which
I
think,
oh
maybe
this
whole
transformation
could
be
cool,
but
it
seems,
like
you,
just
draw
a
good
idea.
Any
couple,
any
common
translation
means
I,
I
you're,
showing
them
chunky
over
to
about
the
direction
you're
like
lesbian
comments
about
that.
Many
comments
being
comments
back
so
like
this
idea
of
compartmentalizing
the
computation
just
single
many
comments.
Really
writing
it.
Sort
of
like
saying
yeah,
I
think
I
can
remap.
C
A
A
E
E
E
D
A
A
He
says
some,
he
says
here:
specific
features
at
villages
below
the
moon
line,
a
specific
features
code
right
opinion,
all
the
units
whose
regions
contain
strike
sue
me
saying:
okay,
all
those
those
are
like
the
green
circles,
right,
yeah,
different
features,
which
is
something
that
was
a
set
of
these
unit.
Can
they
greatly
exceed
the
number
of
units?
Yes,
so
I
could
have
like
the
ten
mini
100
mini
columns
button
in
code.
A
Billions
of
features,
this
type
of
encoding
is
especially
effective
when
this
is
the
censorship
now
especially
effective
on
the
number
of
features
present
at
any
one
time.
Well,
how
do
you
have
a
number
of
feature?
That's
like
the
holy.
Do
you
have
one
feature
present.
Any
one
time
is
much
smaller
than
the
number
of
possible
discriminable.
C
E
E
A
A
D
A
A
A
A
A
E
E
A
A
E
E
A
C
E
A
Of
this
game
just
makes
it
he
says
here.
First
talk
about
distributed,
voting
first
and
itself
can
be
dramatically
reduced.
We
do
my
oh.
We
have
lots
of
fewer
mini
columns
before
the
mapping
by
distributing
code,
yeah
of
the
individual
features
and
mappings
there's
patterns
of
activity
and
many
different
hard
minutes.
Okay,
so
that's
all
consistent
with
our
work
so
far.
A
A
E
B
E
A
A
Have
to
do
it
any
any
any
type
of
a
mapping
between
this
code
in
that
code,
I
can
use
it
using
fewer
than
octave
requires
every
human
here
to
connect
every
other.
There
I
can
use
fewer
units
because
I'm
using
sparse,
distributed,
representations
I.
So
I
get
my
capacity
by
that's
the
art,
as
opposed
to
dedicated
grandmother
cells
and.
E
A
A
A
A
A
E
E
A
It's
implied
I've
not
stated
explicitly
he's
going
to
go,
show
the
fact.
Well,
I
could
one
that
one
critical
interpretation
he
could
say
like
well,
if
I
want
to
do
a
mapping,
because
I
think
I
know
what
the
mapping
function
is
to
something
I
did
mapping
function,
it
could
require,
and
it's
extremely
limited
to
require
n
squared
mappings
between
every
divisions.
Every
digital.
A
A
E
Think
I'm
jumping
I,
think
I'm
correctly,
showing
what
he's
saying,
but
he
I
don't
I'm,
not
confident,
well,
whether
it
won't
work,
he
doesn't
make
it
obvious
that
won't
work
but,
but
it
might
is.
It
seems
possible
to
me
that,
okay,
like
suppose,
the
object,
placed
features
and
reclamation
reduce
this
is
just
kind
of
proof
of
concept.
Suppose
that
they
are,
you
know
exactly
the
same.
We're
just
they're,
essentially
the
same
thing,
just
view
from
a
different
reference
right.
So.
C
C
E
There's
a
equivalent
one
over
here
and
you're
always
going
to
be
able
to
I.
Don't
know
the
record
of
Java
said
you're
always.
The
point
is,
though,
imagine
just
as
the
proof-of-concept.
These
are
the
exact
same
code
just
from
different
projects
right
then,
it's
not
hard
to
believe
that
some
mapping
will
exist
from
from
unit
to
unit.
B
C
B
E
A
D
D
Yeah.
If
you
take
a
diagonal
line
and
you
rotate
it
by
that
orientation,
it
becomes
a
horizontal
line
or
if
you
take
a
horizontal
line
and
you
tilt
it
this
way,
then
it
becomes
a
vertical
line.
Those
are
the
pairs
of
feature
plus
orientation
is
how
eccentric
representation
of
the
feature
and
then,
if
all
three
of
them
are
there,
then
you
get
an
H.
Then
you
get
the
most
votes
for
H.
A
E
A
Advice
if
I
thought
those
triangles
as
opposed
to
reading
columns
if
I
thought
in
terms
of
modules
like
a
good
cell
module
to
ribs
on
module,
then
that's
is
what
we
said,
my
god
this
year,
the
displacement
idea
that
a
maza,
what
was
kind
of
modular
module,
but
then
these
are
no
longer
just
you,
because
this
would
be
a
a
module
or
mythological.
This
is
the
grid
cell
module,
and
then
that
case,
these
connections
really
are
just
like
that.
A
They
don't
go
across
like
this
right,
because
the
protein
in
one
module
encoding
in
another
module
and
I
do
this
encoding
I.
Do
this
transformation,
this
transformation
under
this
transformation
and
the
whole
thing
works,
so
that
idea
of
our
one-to-one
was
definitely
in
the
displacement
cell
idea,
but
it's
not
anywhere
in
our
army.
A
C
C
A
Great
so,
but
I've
never
been
able
to
unite
that
many
columns.
That's
why
I
hope
we
struggling
to
come
with
the
whole
of
many
problems
in
the
whole
court
economy,
so
at
first
Oh
baby.
This
is
a
clue
to
that,
but
this
is
really
pretty
much
describing
what
you've
done.
The
displacement
models
is
it's
a
basic
idea
of
the
model.
A
C
C
E
E
E
D
A
A
E
Should
it
should
I
go
ahead
and
this
one's
quicker
yeah,
the
other,
the
other
idea?
So
this
is
going
all
the
way
back
to
the
first
paper,
the
one
that
hinton
sent
you.
So
you
may
have
read
this,
but
I'm
gonna
I'm
gonna
put
an
extra
interpretation
on
top
of
it.
So
first
of
all,
I've
drawn
a
letter
here
and
that
has
been
rotated
and
what
has
been
observed.
E
Experimentally
is
that
if
people
are
instructed
to
look
at
a
letter
that
and
look
at
a
letter
like
this,
that
has
been
rotated
and
answered
quickly,
whether
it's
backwards
are
not
of
whether
it
like
if
this
were
top
right,
wouldn't
be
forwards
or
backwards.
People
aren't
able
to
do
that
immediately.
They
they
have.
A
E
E
Artists
backwards,
the
in
the
paper
then
that
first
one
about
reference
frames,
they
actually
did
some
experiments
where
they
actually
came
to
the
conclusion
that
people
probably
do
have
these
viewer
and
variant
bottles
and
objects.
However,
it
put
it
first
in
their
terminology,
however,
those
models
don't
prescribe
a
handedness,
which
is
to
say
the
axes
of
this
reference
frame.
Might
you
right-handed
and.
E
Which
is
to
say,
like
you
would
think
that
when
you
worry
yourself,
what
you
need
to
do
is
figure
out.
Okay,
which
way
is
up
and
which
way
is
north
and
north
is
always
some
quotation
marks
here.
It's
not
a
cold
R
and
once
you
know
which
way
is
up
and
which
way
is
north.
You
know
hundred
that
you
don't
orientation,
but
under
this
view,
this
idea
that
there's
actually
a
little
bit
more
flexibility
there.
You
also
need
to
figure
out
which
way
is
left
or
which
way,
which.
A
E
An
alternate
interpretation,
and
they
did
some
experiments
to
suggest
that
to
conclude
that,
actually
people
can
recognize
letters
immediately
if
you
run
the
experiment
correctly,
if
you
run
the
experiment
in
a
in
a
clever
way,
I
could
talk
about
that
more,
but
the
idea
is
the
inference
from
their
results.
Is
the
idea
that
your
your
mental
models
for
objects.
E
A
E
E
A
C
E
A
E
I,
don't
have
a
good
answer.
Is
that
no?
And
so
just
this,
is
this
idea,
though
of
it
like
now,
I'm
thinking
an
orientation
a
little
bit
differently
and
that
we
like
a
show
right
here,
surrounded
by
these
few
familiar
landmarks,
and
it
remember,
is
like
oh
I,
know
this
environment
but
in
fact
this
environment
that
kind
of
their
image.
So
this
half
integration
system
is
gonna
have
to
be
such
that,
like
it's
gonna,
be
backward
just
like
turning
to
turning
clockwise,
it
has
to
cause
his
head
direction
themselves
to
turn
counterclockwise,
but.
A
A
A
A
One
day
the
gravity
vectors
you
know,
but
that
doesn't
apply
to
the
cortex,
because
I'm
driving
vectors
from
my
hand
and
object
and
holding
in
my
hand,
and
so
that
you
know
maybe
the
cortex
really
is
truly
three-dimensional
and
anymore,
inherently
and
and
maybe
the
entrada
cortex
is
born.
Apparently
two
dimensions
so
I
see
that
reflection
problem
in
Rose
babies,
which
is
that's
a
lot
of
cortex
I,
see
that
as
the
out
different
backwards,
I
don't
have
to
reorient
to
room
the
nuances
backwards.
C
A
Feels
like
so
many
things
that
were
done
too
early
on
and
pay
I
mean
today
it's
vision,
based,
and
so
so
much
of
that
is
based
on
two-dimensional
images.
You
just
have
to
be
cautious
of
you
know.
Maybe
that
really
impacts
other
thing
about
it,
since
the
idea
that
this
art
of
backwards
and
looking
his
own
in
concept
for
a
two-dimensional
projection.
C
A
A
B
B
B
A
C
B
B
Is
the
art
I
think
it
requires
this
mental
partition
because
it's
unstable
prediction
you
would
make
you're
looking
at
the
head
of
the
are
a
I
know,
I
can
go
down
in
the
reference.
Remedy
are
and
goes
have
legs,
but
what
they
were
ditching
the
leg
deliver
exactly
with
its
you
know,
straight
or
bang
a
little
after
maybe.
B
A
B
I
think
it
so
to
me,
it
comes
down
to
the
ability
to
make
predictions
oriented
and
then
validate
those
predictions
with
you
know
attention
or
anything
else,
and
when
the
backward
that
you
show
that
the
informations
are
very
easy
to
generate
because
of
the
clear
perceptible
orientation
of
the
beat.
You
know
that
the
pointing
in
the
direction
of
the
rest
of
this
word
that
you've
seen
before,
and
so
you
can
make
a
very
prospect
prediction
and
I'm
guessing
that
yeah.
C
B
D
A
A
E
B
D
F
F
A
B
B
B
A
A
E
A
D
A
D
D
C
D
D
D
A
D
D
D
D
B
D
D
C
F
D
A
A
A
A
A
E
A
C
A
A
A
A
Well,
okay,
I
thought
the
common
referencing
would
be
here
because
I
could
say
something
like
well.
This
feature
this
feature
these
around.
Those
could
appear
different
places
on
here,
but
I
see
your
point
all
right,
all
right,
so
keep
going
so
I
have
to
I,
have
to
I.
Have
you're
saying
this
only
is
back
three
years
about
what
I
do
so
I
need
to
say,
but
he
don't
know
the
reference
and
I
know.
E
A
A
E
E
D
C
A
C
C
A
B
A
E
A
A
A
Yeah,
that
is
the
fact
that
you're
moving
you
the
protected,
removing
you
how
you
can
you
can
I
pop
this
is
by
your
location
and
has
removed
all
the
hypotheses
get
updated,
and
then
you
eliminate
them
yeah.
So
the
movement
itself
eliminates
the
possibilities
we
don't.
That
is
the
solution.
The
basic
solution
to
one
finger
you
have
to
generally
have
eliminates
that
narrows
down
the
Union
possibilities.
Here
we
have
no
moment
you
have
Union
and
possibilities,
and
we
have
some
other
mechanism
which
would
somehow
relate
these
guys
to
each
other.
A
A
E
F
A
The
thing
I
talked
about
with
her
I
no
longer
view
the
problem.
It's
like
remember,
I,
said.
The
problem
we
have
to
solve
now
was
a
single
sense
of
this.
That's
moving,
you
know
changing
its
angular
position
and
that
the
idea
of
moving
your
finger
over
the
object
is
it
sort
of
that's.
That's
like
the
your
sighted
I
asked
about
different
positions:
it's
not
like
you're,
just
sensing,
a
point,
you're
sensing
part
of
a
substructure
here,
yeah.
A
Obviously,
with
that
argument,
I
made
over
the
last
week
my
gos
maker
on
there
about
the
difference
between
just
think
about
the
problem.
Now
is
right
there
right
right.
We
give
this
thing
here
that
this
is
good
the
problem,
and
this
is
a
sort
of
it
and
that
the
real
system
here
you're
trying
to
if
we
solve
this
problem,
then
this
one
will
come
out
this
side
I
think
that's
I'm,
going
to
just
handling
this
I
think
that's
gonna,
be
the
solution
to
the
single
finger
problem,
but
we
still
have
this
problem.
Okay,
I.
C
A
You
know
there
are
some
pretty
much
smell
like
saying
these
guys
together,
say:
hey,
there's
only
a
few
things
from
about
your
futures
that
are
going
to
make
sense
here,
yeah
that
would
be
a
pretty
big
shower
and
bath
dip
and
now,
but
now
shut
down.
Then
I
go
back
to
mr.
clausten
here,
yeah
mr.
body,
whatever
this
yeah,
oh
yeah,
that's
a
good
idea.
I
mean
the
general
idea
that
you
have
multiple
these
multiple
different
ways
of
doing
this
now
write
down
before
it's
happening
simultaneously,
so
you've
got
the
voting.
You've
got
the
movement.