►
From YouTube: Numenta Office Hours - April 14, 2020
Description
Jeff Hawkins and Subutai Ahmad take questions from the community. More details at https://discourse.numenta.org/t/numenta-office-hours-april-14-2020/7357
Broadcasted live on Twitch -- Watch live at https://www.twitch.tv/rhyolight_
A
Anyway,
so
let's
go
first
to
Rick
who's,
you've
been
a
rap
quite
a
while
I
recognize
you
going
way
back
to
the
mailing
lists.
Are
you
still
in
Australia?
You
know?
Is
that
the
yes?
Yes,
so
so
Rick
you
have
a
question
for
Jeff
and
it's
regarding
the
internet,
which
I
think
is
kind
of
a
deep
question.
Why
don't
you
go
ahead?
Yeah.
B
B
I
mean
understand
that
the
thousand
Pines
theory
is
in
its
infancy
and
all
the
details
as
to
all
the
naming
and
protocol
and
watering
and
so
on,
between
the
aquatic
columns
it
hasn't
been
not
at
all
yet
and
once
it
has,
can
can
other
distribute
systems
out
there
learn
from
that,
or
maybe
super
speculative
doesn't
work.
The
other
way
around
can
right
here
we
learn
from
existing
distributed
systems.
So
what.
C
You
directed
I
question
I
mean,
but
maybe
super
time
would
be
better.
The
answer,
but
I'll
take
a
stab
at
it.
First
I
mean
you
talked
about
a
lot
of
different
things.
They
write,
you
start
off
with
the
Internet,
but
it
a
couple
of
things.
First
of
all,
because
the
whole
distributed
nature
of
the
thousand
brains
theory.
C
We
didn't
know
that
until
recently
I
mean
it
wasn't
something
we've
presumed
and
asked
it
just
sort
of
it
sort
of
presented
itself
and
it's
a
natural
consequence
of
having
a
you
know
essentially
grid
cell,
like
representations
of
space
in
each
cortical
column,
and
then
the
voting
thing
came
out
of
that
too.
That
happened
fairly
recently.
This
is
not
something
we've
been
thinking
about
for
a
long
time,
but
I'm
a
little
I'm,
not
familiar
with
the
things
you're
talking
about
a
little
bit
more
familiar
with
a
complex
systems.
C
Theory
and
sin
issues
there
at
least
I've
read
some
say
that
expertise
wrong,
and
so
a
couple
things
came
up
from
your
questions.
One
is:
do
we
think
this
is
something
I
think
so
I
mean
it's.
It
seems
obvious
to
me.
Now
was
not
any
reason
so
recently
that
there's
a
lot
of
advantages
to
dividing
up
the
system
like
this
and
having
all
these
multi
agents
voting,
and
it
solves
a
whole
series
of
problems
that
people
have
been
trying
to
deal
with.
For
years.
C
It's
and
I
mentioned
the
complex
system
things
it's
almost
all
complex
systems
in
the
world,
work
in
the
distributed
fashion
like
this
and
somehow,
and
so
that
seems
natural
to
that
it
would
work
that
way.
So
first
answer
is:
do
I
think
this
is
the
doing
machine,
intelligent
machines
going
to
have
to
work
on
these
principles?
C
D
I
mean
I
can
I
can
add
to
that.
Maybe
ads
draw
some
parallels
between
distributed
systems
and
the
thousand
brains.
Theory
I
think
there
are
a
bunch
of
features
that
are
kind
of
in
common
between
the
two.
So
one
might
be,
you
know
most
distributed
systems,
there's
some
level
of
redundancy
in
them.
So
you
know
you
have
multiple
copies
of
something
stored
or
whatever
in
the
thousand
brains
theory.
D
You
know
you
have
cortical
columns
and
you
know
every
cortical
column
is
learning
complete
models
of
objects
and
you
almost
always
have
multiple
cortical
columns
that
are
looking
at
almost
the
same
information
or
the
same
object
for
sure
and
so
there's
a
fair,
quite
a
bit
of
redundancy
in
the
system.
It's
not
completely
redundant.
There's
there's
a
lot
of
you
know:
it's
not
like
every
cortical
column
is
looking
at
every
single
object,
but
there
is
quite
a
bit
of
redundancy
and
just
like
with
distributed
systems
that,
among
other
things,
can
lead
to
fault
tolerance.
D
E
D
Think
another
commonality
may
be
is
scalability
with
distributed
systems.
You
know
you
can
usually
keep
adding
components
and
things
scale
really
nicely.
It's
the
same
thing,
I
think
with
a
thousand
brains
theory.
Because
of
the
way
it's
done,
you
can
add
new
sensory
modalities.
You
can
add
new
cortical
columns,
you
know
different
parts
of
the
hierarchy
and
the
system
will
adjust
and
the
capacity
will
grow
and
we'll
be
able
to
do
more
with
that,
so
it
it
makes
it
much
easier
to
scale
as
well.
D
Another
thing
I
was
thinking
of,
as
you
are
talking
Rick
what
was
typically
with
something
like
the
internet
or
distributed
systems.
You
need
some
common
language,
you
know
like
tcp/ip,
or
something
to
communicate
and
same
thing
with
a
thousand
brains.
One
of
the
big
things
we
have
to
solve
is
you
know
if
you
have
cortical
columns
that
are
acting
independently?
Well,
they
need
some
common
language
to
communicate
and
that
language
is
the
one
that
they
you
use
for
boating
and
for
other
things.
So,
there's
I
can
definitely
draw
a
lot
of
parallels
between
the
III.
C
Could
jump
in
a
little
bit
too?
You
mentioned
the
scaling
stuff,
which
is
fascinating.
In
course,
mammals
have
neocortex
and
it's
fascinating,
the
variety
of
them
and
how
messy
they
are
and
how
it
literally
almost
appears
that
you
could
hook
these
things
up
together
and
lots
and
lots
of
different
ways
without
a
lot
of
concern,
and
it
still
works.
There's
a
lot
of
evidence
for
that.
There's
obviously
bed
with
designs
and
others.
C
But
the
point
of
that
nature
has
gone
to
the
scaling
exercise
in
mammals
and
primates
and
humans,
including
till
very
recently,
you
know,
and
and
there's
lots
of
varieties
of
how
cortical
columns
have
been
arranged
and
put
together
in
various
ways,
you've
different
sensors
and
they
all
seem
to
work.
So
it's
we
have
evidence.
That's
true,
there's
a
great
question.
B
B
This
an
energy
or
synergy
whatever
on
the
that
that
they
distributed
brain
and
distributed
systems
out
there
they're
engineered
are
somewhat
similar.
What
does
that
mean
for
organization
of
research
in
this
area?
I
mean
if
I
meet
a
distributed
systems
engineer
on
the
road
and
each
asked
me
all:
we've
been
trying
to
solve
those
problems
from
consensus
and
voting
and
language
can
I
send
your
way
in
saying
you
know
no
meant
as
working
on
this
exact
same
way,
but
you
may
get
some
inspiration
out
of
what
they're
doing
can
I
do
that
we
can.
C
Definitely
send
them
to
our
research,
I'm,
not
sure
they'd
wanna
have
a
conversation
with
us.
In
any
particular
day,
I
mean
it's
a
bit
outside
of
what
we're
doing
right
now,
databases
so
I
think
they'd
have
to
look
for
inspiration
and
what
we've
written
more
than
like,
oh
yeah,
we've
got
some
code
to
share.
You
know,
I,
don't
think
that's
gonna
happen
thanks.
B
D
Well,
we
mean
the
code
for
all
our
papers.
Is
there
and
that's
available,
as
you
know,
so
we
have
certainly
to
the
extent
that
we've
written
about
in
our
simulations,
you
know
so
we
have
stuff,
are
on
the
voting
and
an
interaction
between
the
the
sensory
layer
and
the
location
layers.
Those
you
know
there
is
code
for
that.
So.
F
D
C
Don't
get
snow
the
whole
time
on
this
topic,
but
the
the
real
key
thing
about
the
distributed
nature
of
the
cortex
is
it's
it's
the
sensory
motor
module
it's
and
that
makes
every
single
quarter
column
a
sort
of
a
complete
modeling
system
in
its
own
right
and
I.
You
know,
I,
don't
I,
don't
know
if
about
distributed
systems,
but
there'd
be
a
lot
of
other
different
systems
that
don't
have
that
sort
of
nature
to
them
that
completely
independent
bottles,
but
at
least
in
terms
of
AI
or
machine
intelligence.
That's
a
key
part
of
open.
A
A
G
Yes,
okay,
okay,
so
my
question
is
so,
let's
fast
forward
in
the
future
and
if
you
were
to
have
access
to
a
brain-computer
interface
that
has
high
temporal
and
spatial
resolutions,
resolutions
that
allow
you
to
zoom
in
and
see
separate
neurons
and
maybe
even
possibly
stimulate
them
and,
let's
assume
put
it
for
the
sake
of
the
discussion
that
is
completely
safe
at
that
point
and
it
said
they
approved,
etc.
What
would
be
the
first
thing
you
would
do
with
this
kind
of
technology?
G
What
what
you
would
test
with
it
like
there
are
a
lot
of
things,
obviously
that
you
can
do.
But
what
do
you
think
would
be
the
first
priority
to
test
with
it?
Do
we
do
you
have
any
specific
thoughts
on
how
it
could
help
in
further
development
of
the
theories
that
you
work
with
and
maybe
possibly
been
applying
it
to
the
real
world
like
real
life?
And
what
do
you
think
like
regarding
the
first
useful
applications
of
such
technology,
both
the
research
and
general
problem?
Well,.
F
D
Are
a
ton
of
predictions
that
come
out
of
a
theory
that
are
really
hard
to
test
with
today's
experimental
techniques?
So
you
know
if
we
could
look
at
the
activity
of
all
the
neurons
across
multiple
cortical
columns
and
across
layers,
while
animals
are
actually
doing
behavior
I
mean
there's
a
ton
of
stuff
that
could
come
out
of
it
and
some
of
which
we
listed
in
our
papers.
But
if.
C
I
mean
there's
literally
everyday,
as
we
think
about
this
stuff
and
there's
questions
like
Oh.
Could
it
work
this
way
that
way,
I
don't
know
it
could
either
way
this
way
that
way,
I
don't
know
we
don't
need.
You
know
we're
still
struggling
with
the
basic
grid
cells
work,
and
so
when
we
go
through
these
different
models,
as
thousands
just
there's
so
many
every
moment,
you
say
well
my
work.
This
way,
if
I
could
just
push
a
button
say,
did
it
work
that
way?
That
would
be
amazing.
That's
not
reality,
but
that.
C
G
Yeah
I
imagine
that
you
live
for
like
to
not
stop
creation
like
creativity,
process
and
also
a
if
I
may
another
question.
We
got
a
big
creativity,
specifically
Jeff
in
the
meeting
that
you
have
this
legs
statement
and
a
podcast
you
mentioned
in
an
approach
at
the
shoe
useful,
creative
thinking.
Like
you,
wake
up
at
night
at
you
stay
awake,
and
can
you
please
elaborate
on
specific
steps
like.
C
What
do
you
and
not
sure
I
don't
mind,
but
that
may
not
work
for
anyone
else.
I
have
no
idea,
if
that's
a
general
purpose
thing
and
it
by
the
way
it's
an
observation
of
mine.
Less
of
like
oh
I
thought
about
here's.
How
what
you
do
it?
No,
it's
just
something:
I
learned
many
years
ago
that
seems
to
work,
and
so
it
began.
I
have
no
idea.
That's
works
for
anyone
else.
So
for
what
I
do
is
I.
C
If
I
can
think
about
a
problem
before
I
go
to
bed
I,
just
not
hard
to
them,
usually
thinking
about
calm
before
a
little
bed,
especially
a
really
hard
one
and
and
and
then
you
just
go
to
sleep
and
then
I
always
wake
up
in
the
melon
I.
Don't
know
why
I
don't
do
and
I
don't
have
a
problem
lying
in
bed
perfectly
still,
with
my
eyes
closed
for
45
minutes
for
an
hour,
doesn't
I
don't
get
bored
doing
that
and
and
I
just
those
those
those
things.
This
often
I
just
bring
up
the
things.
C
G
C
C
That
usually
happens
to
get
up
and
go
to
the
bathroom
come
back.
Well,
you
know
you
kind
of,
like
I,
think
it's
there's
some
sort
of
relaxation
state
that
your
brains
in
five
minutes.
It's
just
it's
like
you!
If
you
try
too
hard
to
solve
a
problem,
the
problem
doesn't
get
solved,
but
if
you
don't
try
too
hard
and
it
just
pops
in
your
head.
C
G
C
C
A
H
Not
a
centralized
one,
but
if
you
look
at
it
a
huge
chunk
of
the
brain
and
the
rest
of
this
neural
network
is,
is
all
about
running
a
body
yeah.
So
that's
that
was
around
since,
before
there
was
a
cortex
as
the
cortex
is
developed
throughout
a
lot
of
your
earlier
work,
particularly,
you
had
a
location
signal
that
just
sort
of
appeared.
It
was
just
there
and
no.
C
No
I
I
think
the
location
think
has
been
around
for
a
long
time
and
many
I
think
it's
throughout
the
old
brain
from
bringing
parks
lots
of
Pizza
they'll.
You
know
you
can
consider
that
neuronal
cortex
hippocampal
complex
is
older
than
in
your
cortex
I.
Don't
think
we
know
that
for
certain,
but
and
but
you
know
the
concept
any
animal
that
moves
in
any
kind
of
coordinated
fashion
of
body,
it
has
to
have
some
reference
frames
to
do
that.
I
think
and.
H
C
There's
a
problem
with
that,
and
so
you
know
the
problem
with
that
is
that
and
I
and
I
described
this.
We
describe
this
in
the
frameworks
paper
and
is
that
your
human
body,
your
brain,
your
brain,
is
making
predictions
many
many
many
predictions
simultaneously.
Every
part
of
your
skin
is
you
know
coffee
cup
example.
Every
part
of
the
skin,
touching
the
copper
Cup
is
predicting
its
own
thing.
C
H
H
H
A
prediction
in
vision:
the
vision
was
coming
from
the
other
direction
by
both
of
them
go
through
a
great
deal
of
mapping
by
the
time
they
meet
each
other,
and
you
also
didn't
saying
that
yeah
I'm,
saying
most
of
the
parietal
cortex
is
mapping
between
the
aloe
and
egocentric.
You
know
the
entire
Pryda
lobe
is
basically
doing
that
are.
C
H
Not
challenging
or-or-or
and
in
that
area
I'm
saying
that
the
location
signal
is,
is
a
distributed
form
now
I
didn't
care
what
the
form
is
that
originates
across
the
entire
Sylvian
fissure
and
each
one
of
the
senses
learns
to
map
from
what
it's
sensing
through
the
entire
parietal
lobe
I.
Guess,
I,
guess
I'd
understand.
C
H
C
H
H
C
Exists:
I
guess,
I'm
saying:
if
you're
not
rejecting
the
idea
of
every
cortical
column
has
its
own
location
signal,
then
you
can
apply
the
location
signal
idea
to
every
single
thing.
The
brain
does
and
you
don't
have
to
make
an
exception
for
somatosensory
cortex,
there's,
obviously
a
very
complex
that
of
coordinate
transformations
that
are
going
on.
It's
my
centric
cortex
and.
A
Want
to
as
the
moderator,
let
me
jump
in
here
for
a
minute,
because
I
I'm,
pretty
familiar
with
mark
Browns
writings
on
the
forum,
he's
been
talking
about
the
ideas,
these
ideas,
a
lot
and
they've
stuck
in
the
heads
of
a
lot
of
the
people
who
are
community
is
it?
Let
me
say
this
mark,
and
you
can
tell
me
if
this
is
right
or
not.
A
C
C
H
Me
interrupt
you
if
I
make
sure
I
get
to
the
point
where
your
finger
is.
You
know
where
your
hand
is,
you
know
where
your
arm
is
you're
the
entire
inverse
kinematic
chain
back
to
your
support
structure,
which
is
gravity,
is
that
is
the
your
body
in
space?
That's
all
part
of
the
same
map
that
eventually
turns
into
your
fingertip
on
the
so
I
guess.
C
Maybe
maybe
we
just
think
about
this
I,
don't
think
that
the
brain
tries
to
figure
out
where
the
fingertip
is
relative
to
some
central
body
space.
You
can
figure
out
where
your
fingertip
is
to
relative
to
anything,
and
you
can
figure
out
where
it
is
relative
to
the
hand
independent
where
the
hand
is
well
to
the
body.
So
it's
it's
very
fluid
system
here
and
I
think
it's
sort
of
a
scaffolding
system
as
opposed
to
oh
there's,
one
body,
centric
reference
frame,
I,
don't
think
that's
likely
to
not
be
true.
Although.
C
H
C
C
H
H
C
C
And
but
the
point
is,
we
think
that
there
was
he
sort
of
specialized
systems
that
were
developed
or
evolved
in
the
past
and
what
happened
in
the
New,
York
or
Texas
became
very
generalized,
and
so
the
exact
same
system
works
for
language
and
for
high-level
thought,
and
it's
not
like
it's.
You
know
it's
really
hard
to
say
but
you're
thinking
about
something
like
democracy
or
something
like
that.
You
know
what
is
that
body
centric
space
it's
meaningless,
but
we
think
the
same
mechanisms
that
you
use
to
navigate
your
body
through
space
is
the
same
actions.
C
You
use
navigate
your
thoughts
through
high-level
concepts
space,
so
we're
looking
for
sort
of
the
the
highly
the
sort
of
the
genericized
version
of
the
system,
not
not
the
specific
ones
that
evolved
a
long
time
ago,
who
think
these
specific
systems
evolved
a
long
time
ago,
but
now
sort
of
this
general
purpose.
One
created
like
this
to
make
hundred
two
hundred
thousand
copies
of
it,
and
so.
H
H
Alright
doodles,
the
second
one
is
you
had
one
you
were
discussing
a
grid
paper,
some
other
some
earlier
topics.
I
have
a
pointer
to
it.
In
my
question
on
the
forum,
I'd
asked
you
particularly
a
question
about
the
separation,
doing
spatial
memory
and
spatial
processing
and
I'd,
ask
Matt
and
when
he
asked
you
I
think
he
missed
the
the
point
of
what
I
was
trying
to
ask,
and
that
is
that
I
think
that
spatial
memory,
formation
and
spatial
processing
are
two
different
locations
in
the
brain
spatial
memory.
H
C
That
question
and
I
don't
think
I
think
it's
very
compatible
with
everything
we've
done
here
and
proposing
again
remember
we
start
off
early
I
mean
here's
distributed
right.
You've
got
multiple
models
of
everything,
so
it's
not
like
one
model,
only
one
type
of
model
in
just
one
place
and
something
else
someplace
else
you
know
I
would
argue
my
knowledge
of
a
coffee,
coffee
cup
with
spatial
memory
right.
It's,
it's
not
the
kind
of
memories,
I
think
about
hippocampus,
but
it's
still
spatial
memory
and
so
I.
H
C
If
we're
not
tuned
to
these
things,
so
you
know
exactly
what
parts
or
things
are
occurring
in
the
hippocampus
and
what
parts
appearing
in
prefrontal
cortex
and
what
parts
appearing
in
parietal
cortex
is
a
little
bit
less
interesting
to
us.
It's
it's
a
it
we're
looking
for
clues
as
the
core
core
methods
and
underlying
all
these
things
and
not
about
where
the
difference
is
off
and
why
you
know.
Why?
Is
it
one
type
of
memory
over
here
over
there?
It's
like?
C
What's
common
between
these
things
and
what
are
the
common
components
that
that
all
these
things
have
I'm
just
I'm,
not
saying
your
point
of
view?
Is
it's
just
different
you
coming
out
at
the
point
of
view?
We
don't
we
don't
really
focus
as
much.
We
look
at
those
differences,
maybe
in
sort
of
some
sort
of
clues
to
us,
but
they're,
not
we're
not
trying
to
explain
that
stuff.
Yeah.
H
C
Because
the
zillion
other
things
right,
there's
there's
no
shortage
of
other
things
that
people
don't
understand
so
and
off
to
make
our
choices
I'm
doing
this
about
as
long
as
you
have
yeah,
oh
no
one
can
do
all
of
this.
You
know
it's
good
that
we're
all
good
new
different
things.
I
mean
lately
I've
been
focusing
a
lot
about
old
literature
on
v1
and
what
we
know
about
that,
because
there's
a
lot
of
good
data
there
yeah-
and
you
know
this
resilient
questions
about
that
to
know.
Why
is
it.
C
So
I
point
out
is
that
I
think
it's
good
to
take
different
approaches?
It's
good!
You
look
at
it
one
way
or
the
other,
but
there's
no
shortage
of
ways.
We
can.
You
know
problems
things
to
look
at
the
time
differentiate.
What's
going
on
here.
So
when
you,
if
you
sent
a
paper,
you
know
I'll
look
at
it,
but
it
doesn't
mean
I'm,
gonna
I.
Maybe
I
might
still
decide
to
focus
on
something
else.
That's
like
whatever
we
find
fruitful,
that's
what
we
have
to
pursue.
So
you
know
this
way.
It's
good!
H
C
A
Mark
brown-
that's
a
bit
King
on
the
forum,
actually
made
of
a
moderator,
is
helped
out
with
hoping
with
spambots,
and
it's
not
for
diapers.
So
the
next
hand,
up
and
by
the
way,
if
you
want
to
speak,
try
and
put
your
hand
up
and
zoom,
which
is
a
participants
venue
and
then
raise
hand.
Falco
is
up
next
and
I
know
it's
super
late.
There
is
coming
from
Belgium.
I
I
The
question
is
a
bit
specific
to
the
last
research
meeting,
which
I
found
really
interesting
and
at
some
point
Jeff,
you
explained
you
show
the
ice
cube
representation
of
mini
columns
in
v1,
the
one
that
splice
up
and
then,
if
you
zoomed
in
you,
could
see
the
little
orientation
markers
and
and
then
you
start
explaining
how
how
this
is
supposed
to
be
working
or
and
you.
You
specifically
said
that
you
move
in
three
dimensions:
you
go
forward
and
then
you
go
left
or
something
like
that
or
you.
I
C
First
of
all
just
prepare
mind.
What
we
were
talking
about.
Research
recently
is,
of
course,
a
new
idea.
Is
that
something
else
other
people
talked
about
as
far
as
I
know?
So
just
keep
that
in
mind.
It's
now
that
is
knows
been
doing.
This
is
just
an
idea,
and
the
idea
is
that
part
of
what
a
mini
congas,
not
remember
those
orientations
as
reference
receptive
fields
at
the
cells
and
mini-com,
only
applied
about
half
themselves
and
many
columns
there
and
all
those
reporting
out
is
that
those
cells
seem
to
represent
one
way.
C
If
we're
discussing
that
this
week,
not
only
shows
meanings,
but
when
the
only
thing
it
requires
is
that
there
are
some
must,
there
are
bits
coming
into
the
column
that
can
be
interpreted
as
flow
or
movement
and
they
can
and
the
mini
column
will
learn
or
the
column
of
sedimenting
columns.
Will
learn,
say
having
all
these
bits
are
coming
into
representing
movements
in
different
directions,
or
some
sort
and
it'll
it'll
divide
them
up
using
our
spatial
pool
or
mechanism
so
that
each
column,
inch
mini
column
will
essentially
representing
unique
combination
of
flow
bits.
C
If
you
will-
and
it's
easy
to
think
about
that
as
like.
Oh,
if
I
had
a
two-dimensional
space
that
we
each
one
would
be
sort
of
a
radio
dividing
up,
the
dimension
did
some
number
of
radial
vectors
or
you
can
imagine
three
dimension.
What
we're
talking
about
with
vision
that
the
eyes
have
enough
information
to
detect
movement
in
3d.
Now
you
can
detect
movement
you're
moving
forward
in
and
out
of
the
plane,
you're
rotating
your
head,
we're
shifting,
left
and
right.
C
So
there's
enough
information
in
the
optic
flow
there
that
the
v1
and
the
retina
could
detect
that
in
two
eyes
and
and
b1
could
represent,
that
could
have
represent
dimensions
in
3ds.
There's
not
treaty
input
coming
in,
all
you
need
is
going
in
is
to
look
at
all
the
different
literally
you
all
you
have
to
do
is
look
a
lot
of
two
different
I.
Imagine
retina
and
you've
got
it's
just
a
flat
tree
and
things
are
moving
around
on
it
because
you're
moving
right
there
moving
around
cause
you're
moving.
C
Well,
the
things
on
the
flat
sheet
are
moving
together.
That
means
you're
you're
moving
back
and
things
moving
apart.
That
means
you're
moving
forward.
The
things
shift
this.
That
means
you're
doing
one
thing:
if
they,
if
the
things
move
like
this
you're
moving
one
way
the
things
move
like
this
you're
moving
other
way.
So
the
the
information
you
just
look
at
a
collective
set
of
movement
bits
in
the
retina
and
that
none
of
those
bits
are
3d.
C
C
Yeah
I
mean
optic
flow,
has
been
around
and
talked
about
for
a
long
long
time.
I
think
what
we're
doing
here
is
we're
bringing
into
them
into
the
idea
bringing
into
them
into
v1
and
in
specifically
trying
to
think
more
generically
about
many
columns
and
how
many
columns
represent
space
and
how
a
column
can
represent
space
in
any
kind
of
space.
It's
in
this
idea
of
the
movement
of
a
bit
is
there's
no
Priya
student,
knowledge
of
and
v1
that
this
is.
This
is
a
vision
and
it's
moving
in
certain
direction.
C
D
H
I
But
each
each
individual
mini
column
receives
a
sample,
a
limited
sample
and
for
itself
the
depths
or
tries
to
figure
out.
This
is
what
I'm
seeing
I'm
now
moving
in
in
this
one
flow
that
I
am
particularly
attuned
to
and
now
I.
Is
that
what
you're,
saying
or
sort
of
but
I
can
you
think
of
a
better
way
of
saying
it.
D
Yeah
I
mean
each
each
mini
coal
and
milk
have
to
figure
out
the
structure
of
its
input
and,
if
you
think
about
a
spatial
pool
or
working
on
these
bits
coming
in,
it's
been
up
for
out.
You
know
what
the
common
coincidence
of
these
points
are
and
we'll
figure
out
the
natural
kind
of
underlying
dimensionality
yeah,
okay,.
C
Alright,
this
is
a
complete
reversal.
How
people
thought
about
many
columns
in
the
past,
the
the
classic,
google
and
bezel
interpretation
that
the
receptive
field
or
accepted
fields
are
a
a
piece
of
you
know
it's
a
feature
in
this
in
the
visual
space
and
I
think
that
actually
occurring
there
too,
but
most
of
those
cells
are
movement,
sensitive,
meaning
most
of
the
the
cells
in
in
a
mini
column
in
p1
or
highly
they're,
completely
tuned
to
movement
in
one
direction
or
another,
not
just
a
spatial
time
layer.
C
Four
cells
have
a
spatial
pattern,
but
everything
almost
everything
else
is
directionally
sensitive
and
there's
no
explanation
for
that.
This
provides
an
explanation
that
these
are.
These
are
movement
vectors,
not
just
spatial
patterns,
okay,
I
think
I
start
to
understand,
though.
Oh
oh
yeah,
look
it's
this!
You
know
this
is
the
risk
of
broadcasting
our
research
meetings,
because
sometimes
we
don't
even
understand
anything
until
you
know
three
months
later,
so
it's
it's
really
hard
to
punch,
put
words
on
these
things
for
us.
E
C
A
E
E
Two
questions,
one
for
Jeff
and
another
for
time
for
Jeff.
In
the
last
research
meeting,
you've
said
the
location
signal
in
v1
is
egocentric
and
the
output
is
more
allocentric,
and
this
sounds
like
to
me
implies
the
location
signal
gets
more
and
more
allocentric
as
echoes
of
the
hierarchy.
The
question
is:
how
how
is
this
electrification
in
order
and
actually
I
will
follow
up
question
which
is:
is
this
also
imply
abstract
concepts
really
in
higher
levels
of
the
hierarchy?
E
C
Well,
you
can
answer
the
question
for
a
super-tight,
oh
okay,
so
so
in
a
frameworks
paper,
we
basically
deduce
that
columns
all
making
predictions
and
they
make
two
independent
predictions
or
whatever
reference
frame,
they're
working
in
and
there's
all
eccentric
reference
range
and
the
egocentric
reference
frames,
and
we
speculated
the
reference
range
for
high
level
concepts
as
well,
and
we
did
not
talk.
We
talked
about
how
it
is
that
one
column,
a
simple
mechanism,
could
do
an
egocentric
reference
frame.
C
How
1.com
would
do
it
else
at
you
perpetrate,
but
we
didn't
actually
say
how
analysis
on
to
reference
frame
would
be
created.
We
do
we
didn't
know,
we
don't
know,
and
we
just
said
that-
hey,
if
you
know,
if
you
had
the
right
type
of
you
know
inputs
to
it,
it
would
do
that
now.
What
happened
recently
was
is,
as
we
were
going
through.
This
mini
comrade
processes
and
I
started
working
on
the
mini
hypothesis,
as
as
a
real
basic
mechanism.
C
By
how
sensory
bets
are
input,
bits
to
a
column
could
automatically
determine
the
dimensionality
and
of
the
space
of
the
thing
that's
being
represented,
they
couldn't
determine,
could
figure
out
the
dimensionality
of
whatever
thing
is
being
observant
column
and
when
I
did
that
in
our
follow
that
logic,
if
not
many
common
hypothesis
is
correct.
What
you
would
see
in
v1,
you
would
see
that
the
dimensions
were
retina
centric
space
and
so
not
not
allocentric
space.
So
that's
fine,
there's
nothing
wrong
with
that.
C
It
still
could
recognize
things
like
written
letters
and
things
like
that,
but
it
means
it
would
be
difficult
for
B
wanted
to
somehow
learn
a
three-dimensional
object,
so
I
just
speculated
very
much
off
the
top
of
my
head
in
that
meeting.
A
lot
of
great
thought
that
hey,
wouldn't
it
be
nice
if
I
had
an
all-electric
me
scuse
me
an
egocentric,
v1
representation
of
space
and
I
mean
I,
know
I
need
to
get
to
an
outside.
You
learn
really
quickly
that
somehow
it
occurs
either
in
that
column
in
between
or
in
between
hierarchical
regions.
C
We
don't
really
know
it,
but
even
if
it's
between
hierarchal
regions
it
needs,
it
has
to
occur
somewhat
in
that
in
v1
itself.
So
every
cargo
column
in
v1
would
happen
to
be
saying:
okay,
I'm,
starting
out
in
some
sort
of
egocentric
or
an
eccentric
space,
but
I'm,
probably
gonna.
You
know,
I
have
to
be
able
to
start
converting
it
to
an
egocentric
space
right
away.
I,
don't
know,
that's
gonna
happen
yet
I,
don't
know
zero
idea.
C
How
that's
gonna
happen,
it's
just
a
speculation
that
it
has
to
occur
someplace
and
so
it's
kind
of
Kerr
there.
It's
gonna
crab!
Beware!
If
it's
something
one
of
my
rules
about
the
new,
your
cortex
is
if
the
new
Kirk,
just
as
a
whole
does
something
every
cortical
column
that's
to,
and
therefore,
if
it's
some
basic
function
like
that,
so
so
we
don't
know
yet
now
it's
gone
I'm,
hoping
to
figure
that
out
all
right
to
be
taught
your
question.
E
Okay,
the
the
question
for
citize,
the
human
one
of
the
goals
of
momenta
is
to
apply
the
principles
you
found
in
the
neural
cortex
to
machine
learning
platforms,
and
the
question
is:
how
could
the
sequence
memory
be
implemented
into
a
deep
learning
system?
Is
it
going
to
be
like
almost
the
same
or
fully
differentiable
operations,
or
are
you
going
to
adopt
it
to
fit
better
to
deep
learning
system
like
you
from
the
paper?
How
can
they?
How
can
we
be
so
dance
yeah.
D
You
know
it's
a
it's
a
great
question.
We
did
a
little
bit
of
investigation
on
this
last
summer
and
I'm,
one
of
our
interns,
Jeremy
Gordon,
actually
published
a
paper
on
archive
with
an
initial
take
on
this
and
the
basic
idea.
There
is,
let's
say
you
know
we're
trying
to
really
map
almost
exactly
what's
going
on
in
the
temporal
memory
over
to
you
know
more
a
deep
learning
framework
in
there.
D
We
assume
that
there's
some
sort
of
a
what
we
create
a
sparse
representation,
that's
coming
in,
and
then
we
have
a
set
of
cells
that
are
recurrently
connected
just
like
in
in
temporal
memory,
and
we
have
you
know
you
can
even
organize
them
into
many
column
structures
so
that
the
cells
along
a
mini
column,
share
kind
of
the
same
feed-forward
weights,
and
you
try
to
get
the
system
to
learn
things
and
predict
things.
So
there's
a
paper
I
think
so
the
interns
name
is
Jeremy
Gordon
and
you
can
look
it
up
on
an
archive.
D
There's
a
sort
of
an
initial
take
on
this
and
what
we
tried
to
do
there
is
try
to
make
sure
that
the
learning
rules
and
the
error
rates
were
very
local.
So
there's
no
back
prop
through
time,
and
so
you
know
lot
of
the
recurring
networks
to
get
them
working.
Well,
you
have
to
propagate
the
gradient
through
sort
of
in
Fulton
to
the
past,
sometimes
an
infinite
number
of
times
steps
into
the
past
so
that
causes
a
you
know:
huge
memory,
usage
and
resource
issues
and
stuff
and
the
brain
that
doesn't
happen.
D
D
Everything
has
to
be
completely
local
to
the
neuron,
and
everything
has
to
work
with
very,
very
kind
of
simple
localized,
heavy
and
style
kind
of
learning
rules,
but
that
is
one
of
our
goals
is
to
set
of
take
this
next
step
and
go
towards
putting
in
something
like
a
you
know,
much
closer
to
a
temporal
memory
structure.
One
of
the
things
we're
looking
at
right
now
and
thinking
about
is
well.
How
can
we
put
in
active
dendrites?
D
You
know
when
you
have
really
sparse
representations
and
once
you
have
a
dendritic
structure
on
the
neuron,
you
can
start
thinking
about
doing
continuous
learning,
because
each
dendritic
segments
can
learn
a
very
muse.
You
know
a
new
sparse
pattern
and
when
things
are
really
really
sparse,
they
don't
interfere
with
the
other
patterns.
So
you
don't
get.
You
know
something
like
catastrophic,
forgetting
we
showed
this
in
our
kind
of
neuron
paper.
In
the
context
of
you
know
biological
theory
but
I,
think
the
same
principles
will
apply
in
practical
systems
as
well.
D
F
Hi
I
posted
this
on
the
forum.
I,
don't
know
if
Jeff
saw
it
basically
I
actually
yesterday
had
the
idea
that
perhaps
the
solution
to
where
to
folding,
which
cells
into
a
cortical
column
is
not
to
fold
them
into
a
cortical
column,
but
rather
to
have
them
external
and
being
shared
by
all
the
cortical
columns
in
in
that
particular
sensory
area,
and
so
did
you
did
you
see
that
question?
Yes,
I
did
yeah.
C
What's
your,
what
do
you
thoughts
on
that?
Well,
I
talked
about
I.
Think
I
talked
about
this
a
little
bit
earlier
and
first
of
all,
just
this
is
talk
a
bit
about
grid
cells
and
what
we
know
about
a
little
bit.
You
know
the
only
real
knowledge
we
have
about
pixel
modulus
is
mostly
from
rat
and
surrounding
projection.
There's
a
paper
that
I
referred
to
a
lot
right
there
tank
and
he
he's
just
first
I
sort
of
mapped
out
with
these
things,
a
little
bit
more
like
in
in
in
detail
and
like
an
entire
grid.
C
C
Nobody
knows
how
grid
so
modules
actually
allow
us
to
encode
space
because
individual
module,
a
lot
of
it's
not
equity
notice,
has
its
periodicity
to
it,
which
allows
doesn't
mean
you
can't
you
can't
specify
the
location
very
specifically
and
we
wrote
in
our
papers
one
hypothesis
behind
that
happened.
I,
don't
do
that
or
pick
the
method
we
talked
about
it's
true
anymore.
So
at
the
moment
we
don't
really
know.
We
do
know
that
grids
or
modules
are
smaller
about
the
size
of
a
quarter.
We
don't
know
actually
how
they
work.
C
C
So
it's
not
it's
not
a
question
to
say
each
cortical
column
has
a
good
soul,
module
or
whatever
it
takes,
because
you
know
that's
what
it
looks
like
mine
equipment.
The
the
only
issue
I
have
with
with
the
routing
idea,
I
think,
is
what
you're
suggesting
like
you
could
have
set
someplace
and
use
them
judiciously.
C
Is
it
just
as
I
mentioned
earlier?
This
seems
to
be
too
much
parallel
processing
going
on
and
it
seems
like.
There's
thousands
of
these
things
these
predictions
occurring
every
moment
in
your
life
and
each
one
has
to
have
a
location
and
and
idea
that
you'd
be
routing
this
information
around
and
updating
things.
It's
really
it's
just
it's
difficult
for
me
to
believe
it
doesn't
mean
it's
not
true,
but
it's
just
it's
hard
to
believe
it's
much
more
copacetic
answer
to
say:
hey,
you
know
it's
just
like
the
anti
wanna
cortex
you
just
have.
C
You
know
you
have
a
whole
bunch
of
these
things
and
and
that's
just
it
but
simpler
explanation
and
it's
faster
and
it
leads
to
the
whole
parallel
processing.
Every
call
you're
working,
simultaneous
with
every
other
one,
as
opposed
to
like
oh
I,
have
all
these
resources
and
I'm
rounding
up
some
here,
a
routing
from
there
and
I'm
rounding
some
here
it
just
doesn't
seem.
The
distance
like
me
is
what
the
brain
looks
like.
F
You
know
I
can
understand
what
you're
trying
to
say
here,
but
I
think
of
some
of
the
things
that
it
does
solve.
For
instance,
assuming
that
you
do
have
in
each
cortical
column
and
then
do
you
have
to
set
the
grid
cells,
then
you
have
the
problem
of
scaling
or
having
just
a
centralized
grid
cell
that
serves
a
whole
lot
of
cortical
columns,
so
things
yeah.
C
But
they've
been
disassociated
from
from
the
cortex
and
put
into
one
central
location
and
so
and
if
you
look
at
the
anatomy
and
the
way
the
cells
projected
from
the
cortex
and
back
again,
it's
very
suggestive
that
this
would
be
a
perfect
place
for
scale.
It
would
be
like
okay,
if
I'm
gonna,
scale,
vision,
I'm
gonna
scale,
the
same
amount
for
a
good
portion
of
you
want
I've
I'm,
going
to
scale
anything
I'm
going
to
do
for
a
good
portion
of
the
cortex,
not
all
the
projects,
because
I
would
scaled
with
important.
C
Quite
different,
but
it's
not
going
to
be
a
column
by
column
basis.
I
think
maybe
maybe
I
read
too
much
in
your
question,
but
the
idea
is
that
yeah
I'm
going
to
scale
a
whole
bunch
of
columns
at
once,
if
they're
a
whole
bunch
of
visual
columns
looking
at
something
each
one
can
ever
told
location
I
need
to
scale
them
simultaneously.
The
thalamus
is
beautifully
situated
to
do
that
exactly
it
has
all
of
my
connections
at
the
right
spot
so
that
working
on
so
that
wouldn't
be
a
centralized.
A
scaling
would
be
well.
F
That
yeah
sorry
times
trouble
and
that's
exactly
why
I
thought
that
would
also
be
a
good
place
to
have
this.
The
grid
cell
modules
sitting
right
every
car
they
would,
they
would
deal
with
the
scaling
themselves
the
same
way
they
scale
when
you
go
into
different
sized
rooms.
The
if
object
is
further
away,
you'd
scale
it
down
and.
C
Yeah
so
I'm
agreeing
with
it,
the
scaling
requires
a
centralized,
but
for
individual
predictions
made
by
by
the
cortex
are
thousands
and
thousands
of
one
simultaneously.
So
that's
not
a
good
place
to
centralized
things.
It's
just
you
know,
I,
don't
I,
don't
know
if
any
dog
Fisher
could
even
do
that.
You
know
in
the
court
in
the
brain,
you
can't
move
things
around.
C
Usually
everything
has
to
be
burned
every
one
week,
chances
to
learn
anyway,
so
to
me,
I'm,
sort
of
proposing
that
it's
sobbing
we're,
probably
looking
for,
which
is
a
centralized
scaling
system,
but
that
doesn't
mean
I
should
centralize
all
my
grid
cells
or
my
location
signals,
because
I
have
thousands
and
thousands
and
thousands
of
them
constantly
making
predictions
in
multiple
modalities,
and
so
that
doesn't
work.
Well,
you
don't
want
to
centralize
that
that
would
be
kind
of
a
super
bottom
line.
F
C
F
C
To
this
tremendous
amount
of
evidence,
suggesting
in
my
mind
and
suggesting
that
that
the
location
signal
is
local
for
comment
and
that
there
aren't
there
isn't
a
method
or
the
connections
or
even
the
neural
possibility
of
calculating
location,
someplace
and
then
putting
it
over
here
and
then
the
moment
they
are
putting
it
over
there
and
so
on
its.
And
we
don't
have
any
evidence
that
Antron
of
cortex
either,
which
we
think
cortex,
is
just
a
grownup
version
of
mental.
You
know,
so
you
might
be
right.
F
D
E
D
A
A
J
C
A
J
C
C
J
Yes,
essentially
like
a
way,
basically
like
a
way
of
demonstrating
the
new
like
how
to
demonstrate
the
new
aspects
of
it
like
like,
like
in
terms
of
incorporating
behavior,
for
instance,
like
understanding
the
stapler.
It's
not
just
about
how
it
feels,
but
about
how
it
moves
and
how
it
behaves.
And
then
part
of
the
model
would
and
then
there's
like
the
idea
of
displacement.
Just
basically
displacing
the
cells
of
representing
how
like
an
object,
can
behave
or
how
it
can
move
or
how
it
can
change.
C
D
Well,
let's
say
you
know
if
you
know
where
my
fingertip
is
and
as
I'm
moving
along
an
object
like
like
my
cell
phone,
how
do
I
convert
in
my
movements
the
movements
of
my
fingertip
into
movements
on
the
reference
frame
of
the
object,
and
you
know
you
can
have
the
exact
same
movement
but
depending
on
the
orientation
of
the
object.
You
end.
D
In
very
different
locations,
so
that's
you
know.
We've
often
thought
about
that
as
that
or
the
orientation
problem,
and
some
of
I
think
what
Jeff's
been
thinking
about
now
could
could
lead
to
that
as
well.
So
that's
a
problem
that
I'm
I'm,
particularly
interested
in
and
the
other
side
of
it
is
you
know
how
do
we?
We
talk
a
lot
about
modeling
the
object,
but
then
how
do
we
turn
that
around
and
generate
behavior?
So
if
I
know
I.
F
D
C
Out
even
worse
and
further,
you
know
you
know
we
this
this
idea
of
each
column
having
its
own
reference
frame
and
doing
complete
modeling
and
voting.
That's
pretty
new
I.
Don't
think
you
know
that
existed
before
just
a
really
big
idea,
but
it
left
out
all
the
details
about
exactly
how
is
it
column?
Does
this
right
really
do?
C
How
do
the
layers
of
the
cells
and
all
that
stuff
actually
implement
this,
and
we
speculated
a
bit
about
it
here
and
there
too,
but
I
just
meant
saying
orientation
problem
so
that
the
next
big
goal,
if
you
will
would
be
to
fill
in
all
those
pieces
to
say?
Oh
here's,
how,
in
the
actual
structure
of
the
column,
does
these
things
and
in
a
sort
of
a
complete,
less
hand-wavy
sort
of
form?
C
What
we
have
right
now
is
it
deduced
properties
of
columns
which
there
is
empirical
support
for
that's
great,
but
we
didn't
explain
how
it
actually
worked,
because
we
didn't
know
and
I'm
hopeful
that
we
can
figure
that
out
and
it
won't
take
it
15
years.
If
it's
a
I
feel
we,
we
have
a
lot
of
the
pieces
lying
around
and
the
sisters
up
we've
been
talking
about
with
many
columns
and
grid
cells
and
orientation
cells
that
there's
there's
just
so
many
pieces
here.
C
We
it's
like
a
puzzle
and
eventually
figure
out
how
to
put
those
pieces
together.
I
think
the
whole
things
coming
pop
out
at
the
moment.
We
have
like
15
pieces
like
this
right,
so
I
mean
that's
the
dream.
It's
a
you
know
and
I'm
really
excited,
but
I
think
this
many
common
hypothesis
might
be
the
key
to
unlocking
how
those
pieces
go
together.
But
it's
too
early
to
know
right.
A
H
H
Look
basically
the
same
everywhere,
but
there
can
also
be
differences
in
lengths.
The
collateral
between
distribution
of
inhibitory
networks,
that
sort
of
thing
I,
don't
think
he
necessarily
is
that
far
off
I
don't
think
it
goes
into
the
thalamus,
but
at
the
cosine
reviews
he
covered
some
hinting
at
some
papers.
I've
seen
some
other
ones
that
follow
the
same
Avenue,
that
is,
that
grid,
so
hex
is
hex.
Coding
has
shown
up
in
the
hub
regions
all
around
the
brain
and
just
there
they
haven't
been
able
to
find
it
anywhere
else.
H
So
the
grid
cell
behavior,
in
other
words
coding
for
located,
starting
that
the
location
signal
to
go
backward
through
that's
distributed
through
the
rest
of
the
the
cortex,
could
easily
be
in
the
hubs
and
go
out
from
there
by
that
same
reasoning,
there's
a
reason
why
they
call
it
the
striate
cortex.
It
looks
different
and
again
changing
the
links
for
the
various
parts
of
the
cortical
column.
You
end
up
with
a
good
bore,
filter,
so
I,
don't
think
that
the
fred's
necessarily
wrong
I,
just
think
he's
putting
the
pieces
in
the
wrong
place.
Oh
my.
C
H
It's
it's
explaining
his
idea.
It
makes
sense
to
me
I'm
just
fleshing
out
how
it
could
actually
be
implemented
and
still
me
you
have
some
really
valid
concerns
about
it's
a
huge
parallel
system.
It
can't
you
you
don't
pipe
things
around
like
a
computer
program.
That's
just
silly!
It
doesn't
work
that
way.
Everything's
hooked,
everything
with
a
fixed,
permanent
connection.
It
has
to
stay
in
the
same
place
all
the
time.
Well,
more.
C
Than
that
it
has
to
learn
it
almost
all
of
the
rules.
I
have
is
almost
all
long-range
connections.
The
cortex
have
to
be
learned
through
associative
learning.
There's
nothing
they're,
not
determined
genetically,
but
but
that's
that
that's
is
a
constraint
on
the
solution.
So
I
don't
think
there
was
a
question
there
on
that
right,
not.
C
F
C
Was
visiting
check
our
YouTube
yeah,
it
was
a
postdoc
with
them
for
six
months
or
so
and
before
he
came
back
and
one
of
the
things
they
one
of
the
things
that
people
have
have
argued
about.
The
nature
of
many
columns
in
particular
and
and
an
argue
that
they're
not
functionally
relevant
is
because
they
were
not
really
visible
in
rodents,
they're,
visible
in
primates,
but
not
in
rodents
and
everyone.
Its
agreed
upon
that
many
columns
are
a
part
of
the
arms
of
genetic
development
of
the
cortex.
C
That
is
that's
how
the
cortex
grows
from
all
mammals,
but
whether
they're
functional
it's
a
different
question,
and
so
in
mammals
you
see
in
the
primates.
You
see
these
of
orientation
columns,
but
in
the
rodents
sake
they
didn't
see
this
stuff,
they
couldn't
see
physically,
they
couldn't
see
them
in
ecology,
but
you
can
see
him
in
primates
and
they
and
they
and
they
couldn't
detect
the
gradual
changes
of
orientation
so
on
on
the
surface
of
the
rat
brain,
but
then
what
what
Florian
reported
back
from
from
cosine
was
that
did
this.
C
They
weren't
looking
small
enough
that
that
the
it
looks
like
those
all
that
stuff
is
there
and
it
didn't
specifically
say
penguins,
but
the
changes
in
orientation
and
those
other
things
which
people
didn't
extinct
existed
in
the
rat
where
the
rodent
cortex
are
there.
But
you
have
to
look
at
a
much
smaller
scale,
so,
instead
of
the
mini-com
mini-com
being
50
or
there's.
C
Like
it's
more
like
there,
it's
I've
learned
there.
You
know
20,
microns
or
less
in
a
party
they
just
couldn't.
They
couldn't
move
they're,
they're
so
close
together
that
the
probes
were
just
too
accurate
in
to
figure
it
out,
anyways
and
now
they're,
starting
to
think
that
those
things
do
exist
in
the
road.
So
it
was
just
an
interesting
sort
of
experiment.
The
five
are
we
are.
We
can
end
it
there
matter.
We
can
do
arm.