►
From YouTube: Brains@Bay Meetup - A Thousand Brains: a fireside chat with Jeff Hawkins (Apr 14, 2021)
Description
To learn more about Brains@Bay, visit our Meetup page: https://www.meetup.com/BraIns-Bay/
Brains@Bay Meetups focus on how neuroscience can inspire us to create improved artificial intelligence and machine learning algorithms. In this meetup, neuroscientist and entrepreneur Jeff Hawkins, co-founder of Numenta, will be talking about his new book A Thousand Brains: A New Theory of Intelligence.
This is a fireside chat between Numenta's researcher Lucas Souza and Jeff, where they will cover the main aspects of the theory, what it represents for neuroscience and machine learning and how can we incorporate these breakthrough ideas into our existing learning algorithms. This is followed by a Q&A.
Link to meetup: https://www.meetup.com/Brains-Bay/events/277296472/
0:00 A Fireside Chat with Jeff Hawkins
1:27:26 Q&A
- - - - -
Numenta is leading the new era of machine intelligence. Our deep experience in theoretical neuroscience research has led to tremendous discoveries on how the brain works. We have developed a framework called the Thousand Brains Theory of Intelligence that will be fundamental to advancing the state of artificial intelligence and machine learning. By applying this theory to existing deep learning systems, we are addressing today’s bottlenecks while enabling tomorrow’s applications.
Subscribe to our News Digest for the latest news about neuroscience and artificial intelligence:
https://tinyurl.com/NumentaNewsDigest
Subscribe to our Newsletter for the latest Numenta updates:
https://tinyurl.com/NumentaNewsletter
Our Social Media:
https://twitter.com/Numenta
https://www.facebook.com/OfficialNumenta
https://www.linkedin.com/company/numenta
Our Open Source Resources:
https://github.com/numenta
https://discourse.numenta.org/
Our Website:
https://numenta.com/
A
A
Today
we
have
a
very
special
guest
which
is
jeff,
but
before
I
wanted
to
talk
a
little
bit
about
brands
of
faith,
so
branded
bay
is
an
online
machine.
Learning
meets
neuroscience
meetup
we
started
in
bayer,
hence
the
name
we
currently
have
about
a
thousand
members
of
it
was
just
last
week
with
milestones.
So
thanks
everyone
and
the
idea
is
to
discuss
the
brains,
part
machine
learning,
algorithms
and
from
the
perspective
of
both
neuroscientists
and
machine
learning
researchers.
A
So
we
had
previous
meetups
on
alternatives
to
back
propagation
there's
multiple
there
active,
dendrites,
lateral
connections,
predictive
processing
have
been
learning
and
a
lot
of
other
topics.
So
if
you
have
feedback
or
suggestions
for
next
topics,
just
leave
a
message
in
the
meetup
board
or
just
send
me.
A
B
A
And
so
the
way
we're
gonna
do
this.
We
have
we
booked
two
hours,
so
we're
gonna,
try
and
answer
all
the
questions
within
an
hour
and
a
half,
but
if
we
can
we
we
booked
for
two
hours,
so
we
might
extend
a
little
bit
and
I'm
the
way
we're
going
to
do
this.
I
have
some
questions
to
dab,
so
we're
going
to
have
like
a
small
conversation
about
thousand
brains.
A
Theory
how
it
applies
to
machine
learning,
and
my
goal
here
is
to
help
machine
learning,
practitioners
of
in
understanding
how
to
apply
this
theory
to
actual
machine
learning
algorithms.
How
can
we
use
these
ideas
to
to
improve
the
existing
algorithms,
so
we're
going
to
talk
a
little
bit
about
neuroscience
a
little
bit
about
machine
learning,
we're
not
going
to
cover,
in
the
main
conversation
we're
not
going
to
cover
other
topics
of
the
book?
A
C
Yeah
sure
sure
you
know
it's
as
lucas
and
I
were
talking
a
little
while
ago.
He
wants
to
focus
here
on
the
machine
learning
aspects
of
this.
Of
course
I'm
more
of
a
neuroscientist
point
of
view.
So
occasionally
I
may
be
pointing
putting
those
questions
back
on
lucas
and
others
to
help
me
out,
but
we'll
see
how
it
goes.
A
Okay
and
yeah
as
a
disclaimer
I
work
for
so
I've
introduced
jeff.
First,
I
think
jeff
doesn't
need
an
introduction,
but
anyway,
jeff
hawkins
is
a
serial
entrepreneur.
He
founded
several
companies,
including
palm
handspring,
graduate
neuroscience
institute
and
then
dementor
and
besides
being
an
entrepreneur
he's
first
and
foremost
a
neuroscience
and
an
ai
researcher.
A
Jeff
has
two
books,
so
on
intelligence
was
the
first
book,
one
that
was
very
inspiring
for
me
and
the
reason
I
came
to
work
at
the
menta
in
the
first
place.
So
I
just
want
to
leave
that
disclaimer
that
I
also
work
at
memento
with
jeff
and
he
just
released
the
second
book,
a
thousand
thousand
brains,
which
is
the
book
we're
gonna
focus
on
so
jeff.
I
wanted
to
ask
you
started
asking
a
question
about
your
motivation,
so
why
did
you
write
this
book.
C
C
But
you
know,
honestly,
those
scientific
papers
are
very
difficult
to
read.
They
don't
tie
together
too
well,
because
you
write
them
one
at
a
time
different
times,
and
so
I
felt
the
story
was
very
important
and
I
wanted
to
make
it
in
a
more
accessible
format,
because
I
think
they're
important
they're
important
ideas.
So
that's
why
I
wrote
the
book
I
felt
it's
like
hey.
This
is
part
of.
What
we
do
is
is
to
communicate
what
we
do.
A
C
Yeah,
so
we're
talking
neuroscience
now.
So
if
you
look
at
the
human
brain,
you've
got
this
neocortex,
which
is
you
know
the
big
wrinkly
thing
that
covers
the
rest
of
the
brain
and
it's
a
it's
a
sheet
of
neural
tissue.
It's
about
two
and
a
half
millimeters
thick
and
it's
you
know
1500
square
centimeters.
C
C
So
these
and
we
talked
about
these
columns
about
somewhere
between
a
half
millimeter
and
a
millimeter
or
maybe
a
little
bit
more
in
size
and
area,
but
they're
not
visible
under
the
microscope,
but
they
exist
because
they've
been
shown
some
sense
empirically
to
exist
in
that.
If
you
were
to
stick
a
probe
through
the
near
cortex
and
vertically
down
through
the
two
and
a
half
millimeters.
C
C
But
if
you
stick
the
probe
horizontally
through
the
neocortex,
what
you'd
see
is
for
about
about
a
millimeter?
All
the
cells
would
respond
to
the
same
patch
of
the
sensory
area,
the
retina
or
the
skin,
and
then
they
would
abruptly
change
to
an
adjacent
patch
of
the
retina
or
the
skin
and
abruptly
change
again
in
another
one.
C
So
the
cells,
even
though
you
can't
see
a
cortical
column,
they
the
tissue
in
the
in
about
a
column,
a
millimeter
square
column
of
cortex
all
responds
to
the
same
source
of
input
and
the
next
area
over
responds
to
a
different
source
of
input.
And
that's
where
the
idea
comes
from.
I'm
not
sure
the
first
person
to
propose
it.
A
A
It
was
like
I
was
there
with
you
and
watching
this
happening
and
talking
to
mont
casa
and
even
about
the
last
words
of
exchanges,
and
I
think
that
was
a
nice
touch.
So
some
of
the
ideas
discussed
in
this
book.
They
were
first
introduced
in
your
first
book
on
intelligence
in
2004
such
as
the
coracle
column.
We
just
talked
about
the
the
role
of
movement,
intelligence
and
the
brain
being
the
predictive
engine
that
builds
the
model
of
the
world.
A
C
I
think
in
general,
almost
most
of
what's
in
the
first
book
I
would
still
adhere
to.
I
haven't,
read
it
recently.
In
fact,
I
decided
not
to
read
on
intelligence.
Before
I
read,
I'm
sure
I
wrote
a
thousand
brains,
because
I
was
afraid
that
if
I
read
it
again
right
before
I
wrote
the
new
book,
I
would
subconsciously
pick
up
the
exact
same
words,
language
that
I
used
before,
and
so
I
didn't
want
to
do
that.
C
So
I
haven't
read
in
a
while
the
last
time
I
read
it
might
have
been
like
five
years
ago
or
something
like
that
and
I
was
I
was
pleasantly
surprised
at
how
much
of
the
book
I
would
stick
that
I
still
adhere
to
I'm
sure
now.
If
I
looked
at
I'd,
find
some
things
that
I
you
know.
You
know
that
was
wrong,
but
I
I
think
in
general,
I
I
don't
think,
there's
anything
fundamentally
that
I
would
disagree
with
in
the
first
book,
but
the
new
book
goes
well
beyond
that.
C
You
know
we
talked
about
the
sort
of
predictive
model
in
the
brain
in
the
near
cortex,
in
on
intelligence,
that's
been
sort
of
a
guiding
factor
for
ourselves
and
other
people
too.
It's
like
how
does
brain
build
this
model
of
the
world,
and
but
we
didn't
really
know
the
details,
any
real
details
about
it
in
when
I
wrote
a
thousand
brains,
I
mean
excuse
me,
I
wrote
on
intelligence,
we
didn't
understand
several
major
things
which
we
now
do
understand.
C
We
understand
better
which
form
the
heart
of
the
new
book,
and
one
of
those
is
exactly
how
information
is
processed
in
cortical
columns,
and
this
is
the
idea
of
reference
frames
that
the
cortex
actually
implements.
All
these
every
column
implements
all
these
reference
frames
of
which
the
sensory
input
is
associated
with
locations
in
reference
frames
relative
to
external
things
in
the
world,
and
we
didn't
know
about
that.
C
We
did
not
know
that
when
I
wrote
on
intelligence,
we
assumed
that
the
model
was
incorporated
in
the
hierarchy
of
cortical
regions
like
everyone
else
did,
and
there
is
still
this
higher
cortical
region.
But
we
now
know
that
every
cortical
column
is
a
complete
modeling
system
on
its
own,
which
is
a
totally
surprising
idea.
We
didn't
expect
that
and
then,
along
with
that,
we
also
now
understand
how
columns
are
voting,
which
we
didn't
understand
before
either,
and
that
gives
a
whole
other
way
a
filter.
C
Looking
about
how
the
connections
are
bringing
what
they
do.
So
it's
kind
of
like
we
had
the
big
idea
on
intelligence,
but
we
didn't
really
know
how
it
worked
at
all
in
terms
of
like
sensory
motor
integration
and
how
knowledge
is
stored
in
reference
frames
and
how
the
cortex
is
organized.
Overall,
we
got
that
kind
of
wrong.
I
guess
that
is
something
we
got
wrong
on
intelligence,
so
those
are
the
big
ideas
and
they're
really
big
ideas.
I
don't
know
you
know
the
downplay.
I
think
they're
really
important
ideas.
C
It
sort
of
explains
how
knowledge
is
stored
in
a
system
like
a
brain,
what
it
means
to
know
something
in
integrating
the
sensory
and
the
motor
interactions
with
these
reference
frames.
So
that
was
a
big
piece
and
then,
of
course,
in
the
book
I
go
beyond
the
biological
theory.
I
spend
much
more
time
on
the
new
book
talking
about
artificial
intelligence
and
the
prospects
and
issues
so
does
that
in
the
third
part
of
the
book.
A
C
Yeah,
our
thinking
of
hierarchy
in
the
past
was
was
much
more
mainstream
in
the
sense
that,
oh
yes,
you
know
it's
like
how
deep
the
deep
learning
networks
work
today
right.
You
know
you
have
those
inputs
come
in
and
you
extract
some
some
small
features
and
you
combine
them
into
other
features
and
you
find
them
other
features,
and
so
on
you
end
up
with
some
representation
higher
up
in
the
hierarchy.
C
That
just
throws
a
lot
of
the
ideas
about
hierarchy
out
the
window
not
completely,
but
some
of
them-
and
I
I
have
to
be
honest-
I
think
I
understand
hierarchy
much
less
than
I
thought
I
understood
it
before,
but
because
now
we
have
a
whole
series
of
questions
that
come
about
like
well.
If
every
column
in
the
cortex
is
learning
complete
models
of
the
world,
what
are
they
passing
between
each
other
and
so
we're
spending
a
lot
of
time
right
now
talking
about
hierarchical
composition?
C
How
do
you
do
a
nested
structure?
You
know
how
much
of
that's
currently
occurring
in
the
quarter
column.
How
much
is
it
occurring
between
regions
and
the
cortex
and
so
on?
So
in
some
sense
I
feel
more
confused
about
hierarchy
now
than
I
did
10
years
ago.
But
I
was
wrong
10
years
ago
and
I
think
most
I
think,
at
least
from
the
brain's
point
of
view.
C
I
think
we
all
had
it
wrong
that
no
one,
I
don't
think
anticipated
this
idea
that
columns
are
complete
modeling
systems
because
I'm
not
aware
of
anyone
saying
that
so
I
wish
I
had
better
answers
to
you
lucas
on
exactly
how,
but
you
know
we
I
talk
about
it
a
bit
in
the
book
and
I
make
some
calculations
about.
You
know
how
to
think
about
hierarchy.
Now.
C
A
C
Okay,
so
you
know
the
binding
problem,
people
you
hear
it
a
lot
and
I
we
have
to
be
careful.
Maybe
I
should
try
to
define
it
because
I
think
not
everyone
uses
it
the
same
way.
So
I'll
do
I'll
talk
about
some
aspects
of
what
the
problem
is
it
can
you
can
look
at
it
different
ways.
One
way,
the
simple
way
to
look
at
the
binding
problem
is
to
say:
well,
we
have
all
these
different
sensory
inputs
and
they're
being
processed
in
different
parts
of
the
neural
cortex.
C
How
do
they
get
combined
into
a
single
sensory
perception?
So
how
do
these
different
components
of
color
and
texture
and
sounds,
and
someone
get
mind,
a
single
sensory
perception
and
they're
thinking?
The
thinking
was
well
in
the
brain.
We
ought
to
see
this
convergence
of
all
these
inputs
into
some
spot.
Where
you
know
perception
occurs,
but
the
brain
doesn't
really
look
like
that,
and
so
that
was
one
of
the
problems.
We
don't
really
see
this.
You
know
it's
sort
of
the
inputs
of
the
brain
sort
of
go
all
over
at
once.
C
C
C
It's
very
distorted,
there's
holes
and
blood
vessels
and-
and
you
know,
nerve
holes
coming
through
the
blind
spot
and
so
what's
actually
projected
to
the
cortex.
If
you
could
imagine
what
the
what
is
being
projected
to
the
cortex,
it's
not
like
an
image
at
all:
it's
a
big
mishmash
of
of
stuff
and
it's
it
would
be
unrecognizable
and
somehow
we
still
perceive
the
world
in
some
sort
of
crystalline
beautiful
thing.
If
I
look
out
of
the
world,
I
don't
see
this
mess,
I
don't
see
the
holes,
you
know
so.
C
Lead
to
this
percept,
that's,
stable
and
and
uniform
and
nice,
and
the
same
thing
would
be
argued
about
touch.
You
know
if
you
touch
something
with
your
hand,
you
have
all
these
different
sensory
puts
from
different
parts
of
your
skin
and
they're
moving
all
over
the
place,
and
yet
you
perceive
that
I'm
holding
this
teacup
in
my
hand
and
it's
not
moving
around
and
it
doesn't
feel
weird.
So
these
are.
This
is
sort
of
the.
How
do
our
different
sensory
inputs
get
united
into
some
sort
of
singular
stable
perception?
C
C
So
the
answer
to
the
I
think
and
that
what
we
propose
the
answer-
the
binding
problem
comes
again
from
these
cortical
columns
that
each
are
a
complete
modeling
system,
so
you
have
150
000
or
so
of
these
cortical
columns
in
your
near
cortex,
and
each
one
is
trying
to
model
whatever
its
inputs
are.
Some
are
getting
inputs
from
the
skin
and
the
ears
and
the
eyes
other
getting
inputs
from
other
parts
of
the
neocortex,
but
they're
trying
to
model
something
and
each
one
can
take
it
sort
of
has
hypothesis
of
what
it's
perc.
C
None
of
this
makes
sense
in
sort
of
classic
view
of
hierarchy,
but
what
those
connections
are
doing,
we
believe
are
voting,
and
so
individual
columns
was
saying.
Okay,
I
think
I'm
seeing
something
I
think
I'm
seeing
you
know
a
coffee
cup
is
the
example
I
use
in
the
book
and
another
one
saying
I
think,
I'm
seeing
this
part
of
a
coffee
cup.
C
Sometimes
I
think,
I'm
feeling
something
that
could
be
this
or
could
be
that
but
they're,
not
certain
that
there'd
be
a
lot
of
ambiguity
about
which
each
column
is
is
perceiving,
but
by
voting
then
they
all
reach
a
consensus
and
they
all
say:
okay,
we'll
all
agree
that
this
has
got
to
be
a
coffee
topic.
You
know,
individual
columns
could
have
be
questioning
that,
but
together,
that's
the
only
thing
that
makes
sense.
C
So
it's
these
long-range
connections
and
these
voting.
That
brings
us
our
singular
perception
of
the
world
and
what
we
perceive
you
know
we're
going
about
our
day.
We
don't
perceive
most
of
the
stuff
that's
happening
in
these
quote:
unquote,
columns,
we
don't.
We
cannot
perceive
all
the
processing
that's
going
on
the
messy
stuff.
That's
happening
as
your
eyes
are
moving.
Your
fingers
are
moving,
but
we
perceive
the
devoted,
so
the
voting
neurons
are
actually
stable,
they're
all
saying
yep
we're
all
looking
at
a
coffee
cup
at
this
position.
C
This
angle,
that's
what
we're
seeing,
even
though
your
eyes
are
moving
around
the
inputs,
the
columns
are
changing,
the
their
voting
output
is
the
same,
and
so
that's
so.
That's
why
we
have
the
single
perception
and
we
we
actually
aren't
unable
to
perceive
the
vast
majority
of
what's
going
on
in
these
cortical
columns.
A
C
Sure
I
mean
you
know
it's
interesting,
so
many
people
have
trouble
understanding
reference
frames.
You
know
highly
educated
people
who
are
not
engineers
and
scientists,
some
some
of
them.
Just
I
find
I
just
can't
even
get
this
at
all.
You
know,
and
I
said,
don't
you
remember,
the
xyz
coordinates
you
learned
in
high
school
and
they
go
not
really.
C
So
I
should
point
out
that
you
know
reference
frames
have
been
a
useful
tool
for
thinking
about
cognition
for
a
long
time.
There's
a
lot
of
cognitive
theory.
That's
based
on
use
of
reference
frames,
robotics
people
use
reference
frames,
some
of
the
early
ai
attempts
use
things
called
schemas,
which
are
a
type
of
reference
frame,
but
in
neuroscience
in
in
in
most
of
classic
machine
learning
techniques.
C
Today,
reference
frames
don't
really
play
a
part,
and-
and
so
we
weren't
as
we
were,
studying
the
brain,
we
weren't
thinking
about
reference
frames
and
and
in
the
cortex
as
far
as
I
knew
no
need
to
really
talked
about
there,
but
you
asked
me
to
describe
what
the
what
their
properties
are
I'll
try
I'll
do.
The
best
I
can
first
of
all
in
the
brain-
and
I
think
it's
this
important
part
of
the
theory
is
that
the
reference
names
are
tied
to
movement.
C
So
these
are
not
some
abstract
mathematical
concepts
describing
some
then
dimensional
space.
These
are
the
brain
model
spaces
using
reference
frames
that
are
based
on
how
we
move
in
the
world,
and
so,
if
a
rat
could
only
move
in
two
dimensions
in
a
in
a
world,
then
it
would
develop
two-dimensional
reference
frames
and
if
a
bat
can
move
in
three-dimensional
spaces,
it
might
develop,
three-dimensional
reference
frames-
and
you
know
we
don't
really
know
the
all
the
we.
We
can't
really
model
most
of
the
world.
C
C
The
properties
you
asked
about
are
cell,
but
one
is
we
like
to
say
they're
metric,
so
you
think
of
a
reference
frame
as
a
way
of
specifying
locations
in
relative
to
each
other
and
a
metric
means
that
the
same
even
very
simple.
You
might
say.
Oh
this,
you
know
it's
not
distorted.
You
know
I
can
have
like
a
grid
if
all
the
squares
and
the
grid
are
the
same
size.
So
then
it's
a
metric
space,
I
can
say
I
can
measure
from
any
two
points
I
can
say.
C
Well
they
go
two
over
and
three
up.
Then
then
it's
going
to
be
the
same
distance.
If
I
did
that
same
movement
from
someplace
else,
so
it's
not
a
distorted.
You
really
like
it
to
be
a
metric
space,
so
you
could
say
how
what
direction
I
have
to
go
to
get
from
one
point
to
another
point
in
this
in
this
reference
frame,
and
I
can
make
that's
the
correct
answer,
no
matter
what,
where
I
start
from
that's
one
property,
another
property
that
that
is,
we
know
that
the
brain
does.
C
This
is
is
path,
integration
which
comes
from
the
being
the
metric
space
path.
Integration
is
saying:
well,
if
I'm
at
some
location,
where
will
I
be?
If
I
move
in
a
certain
direction
and
for
a
certain
distance
or
if
I
move
so
much
to
the
left
and
then,
if
someone's
up
and
then
somebody
left
again,
where
will
I
be?
And
so
you
can,
without
even
observing
anything
you
could
just
integrate
your
behaviors
over
time
and
and
be
pretty
have
a
pretty
good
guess
of
where
you
are.
C
After
doing
that,
it
would
be
like
you
know,
getting
up
out
of
your
bed
at
night
and
you're
standing
next
to
your
bed,
and
you
say:
okay.
Well,
I'm
now
going
to
take.
You
know
six
steps
towards
the
bathroom
and
if
it's
completely
dark-
and
you
don't
see
anything
you're
internal,
you
know
you're
no
longer
out
the
bed.
You
can
now
visualize
where
you
are
in
that
room.
You
said
I'm
in
a
different
part
of
the
room,
even
though
you
didn't
see
anything
you're
just
doing
path,
integration.
C
So
that's
another
property
of
reference
frames
that
is
really
key
to
the
whole
theory
is
how
we
make
predictions
about
what
will
happen
when
we,
if
we're,
if
we
were
about
to
make
a
movement
or
do
something,
so
I
think
I
mean
probably
if
anyone
wants
to
speak
up.
I
probably
forgot
something
else,
but
those
are
some
of
the
key
properties
of
reference
frames.
A
C
C
The
reference
frame
of
a
map
would
be,
it
could
be
the
latitude
and
longitude
lines
or
it
could
be
the
you
know
if
the
map
was
like
a
road
map,
and
you
just
had
you
know,
columns
a
from
f
and
rows
one
through
ten.
Something
like
that,
the
it's.
The
the
reference
frame
itself
is
the
grid
underlying
the
map.
It's
just
a
metric
space,
and
then
you
populate
the
reference
frame
with
things
that
are
observed
at
different
locations.
C
So
you
know
a
map.
You
say,
oh
well,
at
this
point
in
this
map.
This
is
what
I
observed
there
and
then,
until
you
end
up
by
doing
this,
you
by
you
can
actually
build
a
model
of
something.
Let's
say:
if
it's
a
map
of
a
town,
you
would
build
a
model
of
the
town
by
moving
around
observing
what
you
see
at
different
locations
and
mapping
right
and
down
on
this
map.
C
And
now
now
you
have
this
model
of
the
town,
it's
it's
more
complicated
than
that,
but
that's
the
basic
idea
you
have
to
you
have
to
in
reality.
You
have
to
you,
have
to
also
note
the
the
orientation
of
the
pose
of
these
objects
and
we
don't
think
that
you
don't
distort
them
at
some
location,
but
that's
the
basic
idea,
and
so
it's
a
that's
a
useful
analogy.
Although
maps
are
two-dimensional,
the
same
basic
idea
works
for
any
dimensional
spaces.
D
C
D
And
so
the
map
is
another
useful
thing
there.
If
you
look
at
the
map,
think
of
the
map
of
the
world
there's
no
natural
zero
point,
but
instead
you
can
pick
a
point
any
point
and
use
that
as
an
anchor
or
use
that
as
a
and
you
can
talk
about
things
relative
to
relative
to
that
yeah.
Not
maybe
not
an
anchor,
but
just
something
you
know
you
can
talk
about
things
relative
to
that.
So.
C
Specifically,
that's
a
good
point,
and
maybe
I
should
we
should
just
clarify
the
way
the
brain
does.
This
is
unusual
right
that
we
know
how
the
neurons
do
this
and
supersite
you're.
Referring
to
that
in
the
brain.
We
know
there's
these
things
called
grid
cells
and
they
create
a
type
of
reference
frame
and
one
of
the
very
and
I'll
just
rephrase
that
you
said
super
tight,
one
of
the
surprising
things
about
them
is
there
is
no
origin
to
these
things.
C
There
is
no,
you
know,
point
of
reference
the
way
that
the
way
they
the
way
the
neurons
implement
these
reference
frames.
They
do
it.
We
could
go
into
it,
but
it's
clever.
It's
complicated.
It
takes
a
while
to
really
understand
it,
but
you
end
up
with
no
there's
no
points.
It's
like
it's
like
it's
like
it's
over
origin.
It's
like
the
neurons
can
represent
this
huge,
huge
space
of
possible
representations
and
any
particular
object.
You
want
to
assign
to
a
reference,
give
a
reference
frame
to
it.
C
C
A
C
C
C
Well,
brain
does
it
differently?
It
appears
at
the
moment
that
the
brain
has,
as
you
said,
an
over
complete
set
of
vectors,
a
few
dimensions.
If
you
will
so
it
might
take
a
three-dimensional
space
and
represent
it
by
50
different
dimensions.
So
it's
an
overcomplete
set
and
it
has
these
interesting
properties
of
of
these
grid
cells,
where
there's
no
origin
and
and
how
they
work
and
so
on.
So
now
the
question
is:
do
you
have
to
do
it
the
way,
the
brain?
C
If
we're
trying
to
build
an
ai
system,
it's
very
risky
to
move
away
from
the
neuroscience
until
you
completely
understand
the
neuroscience,
because
the
neuroscience
has
got
these
tricks
just
over
and
over
again
surprises
us
with
things
like.
Oh,
I
wouldn't
have
done
that.
Oh
I'm
surprised,
and
then
you
find
that
these
properties
that
come
from
the
way,
the
neurons
do
this,
that
you
don't
get
other
ways
I
mean
the
biggest
simplest
example
is
sparsity
right,
most
people.
We
we
design,
computers
and
electronics.
C
We
don't
build
sparse
systems,
we
build
dense
representations,
ones
of
ones
and
zeros
are
equal,
but
the
brain
uses
sparsity,
and
so
it's
mostly
neurons
are
off
or
silent,
and
only
a
few
percent
are
on.
Well,
that's
not
an
optional
thing
right.
It
turns
out
there's
some
really
interesting,
computational
properties
that
come
from
sparsity,
and
so
I
know
I
don't
think
you
can
build.
C
You
know
intelligent
machines
using
dense
representations,
and
so
the
same
question
could
come
around
for
the
reference
frames
and
at
the
moment
I
would
I
would
say
I
would
be
very
very.
I
would
be
confident,
but
not
100,
confident
that
we're
going
to
have
to
incorporate
some
of
the
properties,
the
the
neurons,
how
they
implement
reference
frames
and
that
we
shouldn't
just
say:
okay,
we
can
start
doing
this.
C
You
know
like
a
roboticist
might
do
and
they'll
just
say:
okay,
this
is
a
simple
problem:
we're
just
going
to
define
each
joint
on
the
robotic
arm
and
give
it
its
own,
its
own
reference
frame
in
its
own
orientation,
and
we
can
do
the
mathematics
to
calculate
where
the
hand
is
oh,
that
works,
but
it
may
prevent
you
from
actually
doing
getting
all
the
desirable
properties
that
we
have
as
we
move
about
in
the
world.
So
at
the
moment
I
would
be
very
cautious
to
to
to
say.
C
A
A
100
agree
on
that,
so
you
just
mentioned
sparsely
as
being
one
of
the
key
elements
which
is
maybe
missing
in
modern
machine
learning
implementations.
So
can
you
elaborate
on
that?
Do
you
see
sparsely
as
a
functional
requirement
of
the
welfare
like
you
need
it
to
keep
it
running
at
30
watts
without
exceeding
a
heating
threshold,
or
is
it
the
key
property
of
how
the
brain
learns.
C
Well,
it's
a
key
property
how
the
brain
learns
and
works
and
represents
information.
There's
no
doubt
about
that.
So
I
I
don't
think
it's
optional
in
any
way,
just
to
be
clear.
There's
there's
all
kinds
of
sparsity
that
we
have
we're
talking
about.
One
type
of
sparsity
is
that
you
have
a
whole
bunch
of
neurons
and-
and
you
want
to
represent
something
well,
the
brain
does
it
by
activating
a
very
small
percentage
of
those
neurons.
C
So
if
I
have
10
000
neurons-
and
I
activate
two
percent
of
them-
I'll
have
200
active
neurons
and
it's
only
going
to
use
200
active.
It
never
is
going
to
use.
You
know,
5000
neurons,
to
represent
something,
and
so
it's
a
few
ones
and
mostly
zeros,
and
that
that's
activation
sparsity
there's
also
sparsity
of
connections
between,
like
you
know.
If,
if
I
wanna,
if
I
want
to
recognize
a
pattern,
that's
200
active
neurons.
C
C
C
I
can
represent
2
to
the
64
different
patterns
with
those
64
bits
and
that's
great,
but
the
problem
is:
if
you
have
a
single
error
in
one
of
those
bits,
the
whole
thing
gets
screwed
up
right
and,
and
so
you
can't
have
you
have
to
have
this
perfect
accuracy
and
there's
no
there's
no
relationship
between
the
bits.
So
I
look
two
bit
patterns.
I
can't
say
they're
related
or
not.
You
know
I
go
from
you
know.
C
You
know
if
I'm
just
counting
up
integers
using
binary
on
you,
I
can
flip
all
the
bits
in
one
one
fell
swoop
between
two
adjacent
numbers.
So,
but
with
sparse
representation,
you
get
all
these
wonderful
properties.
You
get.
You
get
a
very
large
representational
space,
it's
it's
much
smaller
than
a
dense
space,
but
it's
still
more
than
the
number
of
atoms
in
the
universe.
You
know
you
could
say
how
many
ways
you
can
choose:
200
out
of
10
000,
it's
a
very
astronomical
number,
you're,
very
robust
meaning.
C
C
C
000
neurons
200
are
active.
That
represents
something
now
I
say
what
if
I
activate
two
patterns
at
the
same
time,
so
now
I
have
roughly
400
bits
active
a
little
bit
less
and
I
activate
a
third
pattern
or
fourth
pattern.
When
I
mix
in
all
these
patterns
together-
and
so
it's
a
you
might
have
thought,
I've
messed
the
whole
system
up,
but
the
system
will
still
work,
it'll
still
process
all
those
patterns
simultaneously
without
getting
confused.
C
So
it's
it's
a
different
way
of
representing
uncertainty.
That
neurons
can
represent
multiple
patterns
at
the
same
time,
using
the
same
neurons
and
up
to
a
certain
point
and
and
nobody
gets
lost
or
you
know
you
can
process
in
parallel,
multiple
things
at
once.
So
that's
another
property
of
sparse
representations,
which
is
we
rely
on
a
lot.
C
You
know
there
was
a
in
our
2016
paper,
titled
white
on
the
runs
of
thousands
of
synapses
subatai
did
a
nice
little
analysis,
I
think
was
in
that
paper
did
nicely
analysis
of
sparsity
and
some
of
the
mathematical
properties
of
them
and
we've
also
done
some
other
ones
since
why
so
dense
is
another
paper
bin?
So
we
think
it's
fundamental
and
in
our
current
work
right
now
is,
as
you
know,
we
as
long
a
lot
of
other
machine
learning.
A
C
That's
a
good
question.
I
don't
know
if
I
know
the
answer
to
that
question.
It's
certainly
varsity
occurs.
First
of
all,
not
everywhere
in
the
brain
is
sparse,
not
every
part
of
the
mineral
tissue
at
all
parts
of
your
brain
sparse.
But
when
you
get
to
the
cortex,
it's
a
prominent
feature
and
different
types
of
cell
population
of
different
levels
of
sparsity.
C
C
In
the
example
in
the
book
I
used
to
talk
about
the
you
know:
the
mental
coffee
cups,
here's
my
method,
coffee
cup,
so
I
said
how
many
columns
are
going
to
represent
that
coffee
cup
are
going
to
have
models
of
that
coffee
cup
and
it's
not
going
to
be
all
150
000.
It's
going
to
be
a
small
subset
of
them.
It's
not
going
to
be
two.
It's
going
to
be
probably
a
few
thousand
and
that's
where
the
the
term,
a
thousand
brains
theory
comes
from.
C
So
you're
going
to
have
a
if
I,
if
I
ask
myself,
which
columns
how
many
columns
are
going
to
represent
any
particular
thing,
it
will
be
a
sparse
set
of
columns
out
of
a
very
large
set
of
columns.
But
you
question
a
different
question,
which
is:
will
they
be
active
and
I'm
trying
to
think
on
my
feet
here
about
that?
What
are
all
the
other
columns
doing
at
that
moment
in
time?
And
I
don't
really
have
a
good
answer
to
that.
C
I
would
say
probably
that
many
of
the
columns
would
be
would,
with
the
inactive
completely
nothing's
ever
inactive
completely
in
the
brain.
So
I
I
would
say
probably,
but
I
really
don't
know
the
answer.
That
question
is
it's
I've
never
asked
that
question.
I
haven't
thought
about
it
before,
so
it
tripped
me
up
a
little
bit.
C
C
80
of
the
neurons
are
excited
to
ourselves,
and
these
are
the
ones
we
mostly
think
about
when
we
think
about
representing
something
and
20
of
the
neurons
are
inhibitory
cells
which
keep
the
excitatory
cells
from
getting
too
excited,
and
so
these
are
everywhere
and
there's
different
types,
but
it's
it's
generally
built
into
the
tissue
that
the
neural
tissue
does.
It
wants
to
be
sparse
it
just
it's
sort
of
it's
just
trying
to
be
sparse,
all
the
time
and
you're
never
wanting
there
is
there's
basically
gain
control
everywhere.
That
says,
we
cannot
form.
C
You
know,
dense
representations,
that
that's
like
a
seizure,
and
we
don't
want
that.
So
this
is
built
into
the
neural
tissue,
that
sparsity
is
itself
almost
guaranteed
and
forced
by
the
mechanisms
of
the
inhibitory
cells
and
now
you're.
But
you
asked
in
the
question
of
well:
how
do
we
know
we
get
the
right
response?
C
I
don't
know
what
that
means.
The
right
response,
I
mean
you
know,
some
responses
are
driven
by
sensory
input.
Some
are
driven
by
our
motor
behaviors.
They
should
reflect
what
the
sensory
input
should
be.
They
should
reflect
what
the
motor
behavior
should
be.
Some
are
representing
location,
so
they
should
reflect
the
locations.
Each
of
these
cell
populations
has
its
own
mechanisms
for
determining
what
the
right
representation
is
and-
and
so
I
think
you
know
if
you
want
to
get
down
to
like
well,
what's
the
right
representation
for
any
particular
cell
population.
A
A
So
it's
not
just
about
which
neurons
are
we
using,
but
the
fact
that
we
are
using
those
neurons
are
the
representation
itself
right
and
that's.
I
think
how
sparsely
super
representations
are
built,
it's
about
which
neurons
are
active
and
not
about
like
this
notion
of
weight
that
we
have
in
deep
learning
or
how.
C
C
C
C
However,
there
are
other
parts
of
the
cortical
tissue
which
are
less
sparse
if
you
think
about
activation
of
mini
columns
or
something
like
that
and
where
you
have
maybe
a
few
dozen
different
items
and
a
handful
are
active.
Well,
those
are
that's
the
the
the
rate
coding
there
is
going
to
be
more
important.
So
it's
not
it's
not
black
and
white,
but
I
would
say
it
both
exists
rate
coding
and
population
coding,
but
I
think
most
people
have
missed
and
understood
the
importance
of
population
coding
in
in
a
lot
of
research.
C
A
I
I
have
harder
questions.
Okay,
so
so
does
the
theory
predict
fully
redundant
cortical
columns
as
we
have
multiple
cortical
columns
with
same
inputs
predicting
the
same
object
or
we
have
multiple
coraco
columns
predicting
the
same
objects,
but
each
one
has
a
different
input.
So
it's
similar
to
the
concept
of
bagging
that
we
have
in
statistical
learning
theory.
It's.
C
More,
it's
more
of
the
latter
that
I
it's
it's.
I
wouldn't
say:
there's
redundancies,
it's
not
like.
Oh,
this
is
a
redundant
machine.
We're
going
to
make
three
copies
of
something
and
compare.
You
know,
compare
their
outputs
to
make
sure
that
you
know
which
ones
are
correct,
or
something
like
that.
Every
quartical
column,
in
essence,
has
its
own
inputs.
Now
it's
not
the
same.
C
The
two
columns
couldn't
have
very
similar
inputs,
but
they're
different,
even
if
I'm
two
columns
looking
at
two
different
patches
from
the
retina
or
two
different
patches
on
your
skin
that
are
next
to
each
other.
Those
inputs
are
different,
they're,
not
always
going
to
be
the
same,
and
each
column
is
going
to
do
its
own
modeling.
C
You
can
even
think
of
like
a
visual
column,
some
columns
might
get
might
rely
on
on
color
and
other
columns
may
not
have
color
or
less
color
input,
and
they
can
both
model
the
visual
world,
and
so
you
might-
and
so
you
might
have
a
model-
that's
more
reliant
on
the
properties
of
the
of
the
grayscale
and
the
low
light
you
can
get
with
that
versus
some.
That
might
be
more
focused
on
color
differences,
and
so
I
think
the
general
way
to
look
at
it.
There's
no
redundancy.
C
You
just
got
a
lot
of
columns.
Each
one
is
modeling
its
own
thing,
even
if
multiple
columns
are
getting
input
from
the
same
sensory
organs
such
as
the
skin
or
the
ear
or
the
eyes,
they're
not
going
to
build
identical
models,
they're
not
being
copied,
they
are
in,
they
some
sense
learn
independently
to
some
extent
and
and
so
they're
you.
But
on
the
other
hand
you
could
knock
out
a
bunch
of
these
columns
and
the
system
works
just
fine
right.
So
it's
not
it's.
C
C
You
know,
I
might
you
know,
there's
you
know.
For
some
reason,
I've
lost
some
touch
sensation
in
my
in
my
face.
I
don't
know
why
that
is,
but
I
have,
and
so
I
can't
always
feel
things
on
my
face,
but
I
feel
pretty
normal.
It's
like
you
know
it's
not
like.
I'm
lost
in
the
world,
so
they're.
So
these
are.
These
are
not
redundant
they're
very
complementary.
C
A
Okay,
so
so
to
combine
these
columns,
we
need
to
vote
right,
which
is
a
key
component
of
the
thousand
print
series.
So
let's
talk
about
that,
so
there
are
many
ways
we
can
vote
right.
So
voting
between
different
models
is
the
whole
superior
machine
learning
and
makes
them.
For
example,
we
explored
this
idea
of
voting
as
unions
of
sparse,
distributed,
binary
representations
and,
like
I
did
with
reference
frames,
I
wanted
to
try
to
deconstruct
the
idea
of
this
voting
mechanism
into
a
set
of
properties
that
we
can
use
as
requirements
to
build
the
system.
C
Oh,
these
columns
are
building
these
models.
Yet
we
have
this
this
this
and
each
model
is
changing,
as
your
inputs
are
changing,
your
eyes
are
moving
and
all
these
things
are
changing
yeah.
We
have
the
singular
perception,
and
it's
like
how
do
you
get
the
singular
perception
that
was
like?
We
talked
about
the
binding
problem,
and
so
the
the
whole
idea
of
voting
came
out
of
that,
as
opposed
to
some
sort
of
machine
learning
concept.
It
was
more
like
hey.
C
How
do
we
have
this
singular
perception
of
the
world,
and,
and
so
when
you
think
about
your
perception,
if
I'm
looking
out
at
something
looking
at
my
window
right
now-
and
I
see
other
things
and
I
see
them
at
distances
well,
my
neurons
are
voting
on
what
I'm
seeing
and
what
the
pose
is
relative
to
me
right
now.
That's
what
they're
voting
on
and
that's
that's
what
I
perceive.
C
So
what
are
the
properties?
I
can
talk
about
how
voting
occurs
in
the
neurons
and
we
can
maybe
try
to
tease
out
some
of
the
properties
about
that
in
the
brain.
A
key
part
of
the
voting
is
this
union
property.
We
talked
about
earlier
that
if,
if
a
column
has
you
know,
ten
thousand
neurons
that
are
representing
something
and
an
object,
identity,
for
example,
and
it's
not
certain
what
the
object
is.
It
can
invoke
multiple,
a
union
of
representations
at
the
same
time
saying.
C
C
Than
this,
but
it's
a
little
bit
about
you
can
think
of
it.
It's
sort
of
like
you,
know
the
well
anyway.
That's
the
basic
idea
and,
and
so
the
mechanism
of
underlying
voting
in
in
the
brain
is
it's
not
like.
I
have
a
probability
distribution,
because
the
unions
aren't
really
a
probability
distribution
they're.
C
Not
they
don't
sum
the
one,
and
you
know
it's
it's
like
it's
like
you
know
it
could
be
these
five
things
and
they're
all
equally
likely
right
now,
but
but
you
know,
they're
all
good
and
and
we
want
to
eliminate
the
ones
that
don't
that
don't
aren't
consistent
with
other
columns
and
so
a
couple
interesting
properties
about
the
neuroscience.
Like
I'm,
not
sure
we'll
get
to
your
question
exactly.
I
don't
know
how
to
answer
it.
A
couple
of
interesting
questions
about
neuroscience.
Let's
say
I
have
thousands
of
these
columns.
C
Each
one
has
10
000
neurons,
representing
some
something
that
we're
going
to
vote
on
in
any
any
one
of
these
comps
200
of
those
things
would
be,
you
know,
one
object.
The
200
cells
would
be
another
object
of
itself,
so
you
might
have
you
know
a
thousand
cells
active
in
each
of
these.
You
know
thousands
of
columns
and
they
have
to
make
learn
these
connections
between
them.
C
The
first
thing
that's
interesting
about
it
is
not
all
columns
have
to
connect
to
all
other
columns.
In
fact,
you
can
have
a
very,
very
sparse
connectivity
between
these
columns,
meaning
it's
important
that
the
whole
network
settles,
but
not
everybody
has
to
connect
the
other
one.
So
if
I
was,
if
I
would
say
a
column
and
I'm
I
have
to
get
input
from,
I
mean
maybe
I
have
to
to
do
my
voting.
I
need
to
get
you
know
five.
Ten
thousand
other
senate
axons
coming
into
me.
C
Well,
those
ten
thousand
accents
could
come
from
any
set
of
columns
all
over
the
place.
It
doesn't
really
matter.
I
could
randomly
sample
where
they're
coming
from
the
whole
system
will
settle.
That's
an
unusual
property
of
them
and
the
mechanism
pretty
straightforward
and
easy
to
make
this
work.
So,
but
what
is
the
fundamental?
C
C
D
I
think
one
of
the
kind
of
fundamental
properties
we
talked
about
in
the
kind
of
the
columns
paper
is
that
you
can
in
a
you
know,
cortical
columns
are
getting
evidence
from
their
input.
They
have
these
multiple
hypotheses
that
they
represent
and
they
can
integrate
evidence
in
two
different
ways.
One
is
spatially
you
can
get
information
from
lots
of
different
other
cortical
columns
and
you
come
to
a
consistent
interpretation
of
which
hypotheses
are
most
like
most
consistent
with
all
of
the
information
available
at
a
particular
point
in
time.
D
But
you
can
also
integrate
information
over
time.
As
you
move
your
senses
around
and
gather
more
information
like
you
know,
touching
your
coffee
cup
or
whatever
you
can
sort
of
narrow
down
that
the
evidence.
So
you
can
narrow
down
the
set
of
hypothesis,
so
there's
two
set
of
orthogonal
ways
of
narrowing
things
down,
and
this
allows
you
to
very
very
rapidly
come
to
consistent
interpretation,
yeah.
C
C
So
what
super
tight
saying
is
you
know
you
can
a
single
column
from
from
your
finger,
let's
say:
can
learn
and
infer
a
cop
or
object
just
by
moving
over
time
and
and
of
course,
you'd
have
to
move
over
time
to
do
that,
because
one
touch
isn't
sufficient
and
the
visual
equivalent
will
be
looking
at
the
world
through
a
very
narrow
straw.
You
know
you
can't
see
everything,
but
you
could
move
it
around.
So
I'm
just
paraphrasing
what
super
type
brought
up
is
that
columns
can
resolve
ambiguity
by
movement.
C
That's
the
the
default
way,
but
they
can
they
can.
They
can
also
resolve
ambiguity
by
voting,
which
is
sort
of
the
supplemental
way
in
some
in
some
ways
the
fundamental
mechanisms
through
movement,
and
so
as
supertai
just
said,
if
you
vote
you
can
you
can
reduce
the
amount
of
movement
you
have
to
do
to
infer
or
to
recognize
something.
A
C
I
mean
it
obviously
takes
some
amount
of
time
to
resolve
the
voting
right.
Neurons,
don't
have
a
clock.
They're,
not
working
on
you
know,
passes
through
some
coding
loop.
So
I
think
from
a
neuron
point
of
view,
it's
just
a
natural
settling
process
of
you
know
it
will
take
a
little
bit
of
time,
but
it's
very
very
fast.
I
don't
you
wouldn't
you
know.
D
Yeah,
at
least
the
mechanism
we
proposed
was
extremely
fast.
The
neurons
that
are
getting
kind
of
most
evidence
laterally
through
their
distal
dendrites,
will
be
stronger
and
they
will
kind
of
compete
and
beat
out
the
other
neurons
that
are
not
getting
as
much
so
it
as
much
evidence.
So
it
does
take
a
little
bit
of
time,
but
it's
super
fast.
C
You
could
think
about.
You
know.
I
use
the
example
in
the
book
of
these
bistable
images,
like
everyone's
seen,
the
vas,
and
that
looks
like
two
faces
right
and,
and
you
look
at
that
and
it's
the
same
visual
input
and
then
it
it
instantly
switches
between
vibes
and
you
sit
there
for
vibes
for
a
while
and
then
oh
now
I
see
the
faces
and
you
see
the
face
as
well
and
then
now
you
see
the
vowels.
C
You
can't
see
both
at
the
same
time,
and
and
what's
literally
going
on
in
that
situation,
is
these
voting
neurons,
just
they're,
locking
in
quickly
onto
one
of
the
a
pattern
that
says:
okay.
This
is
consistent
with
being
of
oz
and
for
some
reason
we
don't
know
why.
But
some
there's
some
mechanism
that
says
okay
enough
of
that,
is
there
another
another
interpretation
and
it
locks
into
the
next
one?
It's
just
the
exact
same
evidence.
There's
no
change
in
the
evidence.
There's
two
interpretations,
but
that
switch
is
very
rapid.
C
A
C
And
that's
what
I
think
is
going
on:
it's
very
consistent
with
the
mechanism,
all
what
it
requires
beyond
what
super
tight
just
said
is.
It
requires
there's
some
sort
of
mechanism
that
that
either
fatigues
or
forces
a
new,
a
new
interpretation.
You
know
so
the
neurons
say:
okay,
we
can
only
do
this
for
so
long
and
now,
let's
try
something
else.
I
I
don't
know
what
that
mechanism
is.
C
Maybe
somebody
maybe
someone
has
discovered
in
some
neuroscience
research,
but
the
point
of
that
story
was
that
there
are
two
interpretations
of
the
evidence
that
both
can
be
consistent
and
they
both
can
be
voted
on
and
then
that
switch
happens
very
rapidly
and
you're.
Not
you
don't
spend
any
time
in
intermediate
states.
Thinking
about
it.
Like
you
know
you,
don't
you
never
see
the
votes
in
the
face
at
the
same
time,
you
know
they're
there,
but
you
can't
really
perceive
them
at
the
same
time.
It's
not.
A
Yeah,
that's
that's
great,
so
those
are
questions
I
always
had
about
voting
in
the
thousand
brains.
So
let's
talk
about
movement
a
little
bit.
That's,
I
think
a
very
important
aspect.
You
mentioned
right
in
the
beginning,
and
one
of
the
key
ideas
we
have
in
predictive
processing
is
that
movement
can
be
used
to
disambiguate,
meaning
during
learning.
We
are
driven
towards
actions
that
will
help
us
capturing
the
missing
pieces
of
the
puzzle
and
then
improve
our
understanding
of
something
so
in
the
thousand
brains.
A
C
H
C
Their
it's
related
to
movement,
so
these
are
the
movement
output
cells
in
the
near
cortex.
So
like
I'm
speaking
right
now,
there's
parts
of
the
cells
in
my
language
areas
in
my
cortex
that
are
firing
right
now
and
those
are
layer,
five
cells
and
they
project
down
to
areas
would
that
would
control
my
voice
box
and
breathing,
and
things
like
that
sweet
now.
The
surprising
thing
about
these
cells,
they're
kind
of
unique
you
can
see
them
they're
well
known,
is
that
apparently
they
exist
in
every
cortical
column.
C
They
they
don't
exist
in
motor
areas
of
the
brain.
That's
what
people
used
to
think
many
years
ago,
like
oh
there's,
a
motor
section
of
the
brain?
Well,
not
really!
That's
just
the
motor
somatosensory
section,
but
even
like
primary
visual
cortex,
the
first
columns
they
get
input
from
the
eyes.
Have
these
cells,
which
project
and
create
movement,
and
they
project
back
to
a
part
of
the
old
brain
called
the
superior
colliculus
which
is
involved
in
controlling
the
movement
of
the
eyes?
C
And
so
it's
you
can't.
The
evidence
is
super
strong
that
you
can't
really
separate
out
movement
from
sensations.
It's
part
of
the
part
and
partial
the
same
tissue,
and
this,
of
course,
has
been
pretty
much
ignored
in
many
many
models
of
vision
and
other
things,
and
it's
hard
to
avoid
it
if
you're
trying
to
think
of
modeling
something
with
touch
or
but
you
can
avoid
it
with
the
hearing
and
vision.
C
So
there
it's
it's
integrated
there
and,
of
course
the
whole
theory,
as
we
just
talked
about,
is
that
you
know
we.
We
learn
and
we
infer
through
movement
right.
It's
the
simple
gambling
you
just
think
about
looking
at
the
world
through
a
straw
or
touching
an
object
within
one
finger.
That's
like
close
to
like
one
column,
that's
moving
and
building
a
model
of
something
by
by
integrating
its
location,
where
it's
viewing
something
in
space,
in
a
reference
frame
and
building
a
model
of
something
based
on
sensing
at
different
locations.
C
And
you
can
integrate
this
together
in
a
single
column
to
build
a
model
of
an
object.
So
one
of
the
things,
a
thousand
brain
theory
says,
is
that
even
low-level
visual
regions
of
the
cortex
are
building
complete
models
of
the
world
because
they
can
integrate
over
time
now.
The
second
part
of
the
question
is
harder
to
answer,
which
is
how
do
these
columns
integrate
their
different
movements?
I
think
that's
what
you're
asking
right.
So
it's
interesting
if
you
think
about
vision.
Well,
all
the
corner
columns
getting
input
from
the
retina.
C
They
have
to
somehow
agree
on
the
movement,
because
your
eyes
can
either
move
one
way
or
the
other.
You
can't
move
them
both
ways
at
the
same
time
right.
So
that's
a
situation
where,
okay,
somehow
in
a
way
that
I
don't
understand
yet
they're,
going
to
agree
that
okay,
there's
one
there's
one
movement,
we're
going
to
make
right
now
and
we've
agreed
that
it's
going
to
be
go
this
way.
C
When
you
come
to
something
like
your
fingers
and
your
hands
and
your
limbs,
it's
different
right.
My
fingers
can
move
independently
of
each
other
somewhat,
and,
and
so
they
don't
have
to
agree,
but
there
are
patches
of
my
finger
that
are
going
to
have
to
move
together,
but
the
patches
in
the
next
finger
over
can
move
independently
of
that
and
they
may
be
represented
right
next
to
each
other
in
the
cortex.
C
So
in
the
cortex,
the
columns
representing
this
finger
and
this
finger
might
be
right
next
to
each
other,
but
they're
going
to
generate
different
behaviors,
one
could
say,
extend
and
one
can
say
flex
something
like
that.
We
don't
have
a.
We
haven't
developed
a
concrete
theory
about
this
yet
about
how
they
communicate
with
each
other
to
form
coordinated
movements.
I
don't
know
the
answer
that
question.
C
I
just
gave
you
some
of
the
parameters
around
it.
Sometimes
it's
they
all
move
together.
Sometimes
they
move
partially
together
and
partially.
Not
it
doesn't,
they
can
move
independently.
Even
even
columns
are
right
next
to
each
other,
the
cortex
could
be
generating
movements
or
different
movements.
Sometimes
they
generate
the
same
movement.
C
A
C
Well,
you
have
to
be
an
optimist
right,
you
got
to
say
like
there's
an
answer
to
these
questions,
and-
and
the
best
thing
you
can
do
is
just
sort
of
tease
apart
the
problem,
as
you
said
earlier,
defining
the
problem
and
putting
parameters
around
it
and
say:
okay.
Well,
it
has
to
do
this.
It
has
to
do
this
and
have
to
do
that
and
have
to
do
that.
We
know
this.
So
what
can
be
the
answer
and
somehow,
if
you
spend
enough
time
thinking
about
it,
usually
the
answer
comes
in.
A
So
I
want
to
switch
gears
a
little
bit
to
talk
about
reinforcement,
learning.
I
think
that's
a
question
that
comes
up
a
lot,
so
in
reinforcement
learning
we
generally
divide
algorithms
between
model
free
and
model
based,
and
the
latter
requires
the
agent
to
learn
a
model
of
the
word
that
can
be
used,
predict
fitter,
state
dependent
on
current
state,
and
we
have
this
middle
ground
with
a
successful
representation
where
the
agent
learns
the
model
of
the
word,
but
that
model
also
imbues
a
prediction
of
its
policies
of
its
behavior.
A
So
the
models,
the
model
that
agent
learned
knows
which
regions
of
the
state
space
are
visited
more
frequently,
and
then
it
can
use
that
information
to
better
predict
future
states
or
prioritize
how
information
is
starting
retrieved.
So,
if
you
think
about
a
map,
if
you
know
that
you're
going
more
to
like
the
center
of
the
town,
you
have
you
know
the
center
of
the
town
drawn
in
more
details.
A
C
C
You
know
there
are
models
that
say:
okay,
an
agent
should
spend
so
much
time
exploring
novel
places,
because
you
don't
know
what
you're
gonna
learn
an
agent
may
say
no,
it's
safer
to
explore
the
previous
export
spaces
because
that's
less
dangerous,
maybe
something
like
that.
Is
your
question
related
to
that.
A
Maybe
a
different
way
to
put
it
is,
is
the
model
we
build
dependent
on
behavior.
So
we
take
that
into
account
to
decide.
You
know
how
do
we
model
and
with
which
parts
of
the
model
are
we
gonna
dedicate
more
information
pieces
to
it
or
or
is
it
independent?
So
we
just
learn
a
model
of
the
word
that
we
could
just
attach
to
any
behavior,
and
we
don't
really
take
that
into
account.
C
D
Maybe
it's
more
like
you
know,
you
have
some
sort
of
a
complex
goal.
You
want
to
achieve
yeah.
How
do
you
decide
what
movements
to
make
to
achieve
those
in
the
most
efficient
way?
And
you
you
know
if
you
have
to
explore
a
lot
of
different
places
or
different
sensory
states,
you
know,
maybe
you
don't
want
to
go
back
somewhere.
You
visited
before
you
know
how.
C
C
It
well
the
general.
The
general
answer
to
the
question
is:
how
do
you
achieve
a
goal?
If
you
have
a
good
model
of
the
world,
then
that
then
the
model
tells
you
how
to
achieve
that
goal.
Even
if
it's,
even
if
it's
a
goal,
you've
never
done
before
we
talked
earlier
about
metric
spaces
and
one
of
the
nice
things
about
metric
space
is
that
you
can,
you
can
describe
you
can
decide
what
are
the
behaviors?
You
need
to
do
to
achieve
a
goal,
even
if
you've
never
actually
achieved
that
goal
from
the
particular
position.
C
You're
in
you
know,
the
simple
example
is
you
know
you
you're
you're
you're
walking
around
town,
and
you
know
where
the
library
is,
and
you
know
where
the
the
post
office
is
and
you've
never
you've
gone
from
your
home
to
the
post
office
and
you've
gone
from
your
home
to
the
library.
You've
never
gone
from
the
library
to
the
post
office.
So
you
know
how
to
get
from
your
home
to
the
post
office,
but
now
you're
at
the
library,
but
by
having
a
model
of
the
town.
C
You
know
you
can
you
can
say
exactly
how
you
get
from
the
post
from
the
the
post
office
to
the
library
or
whatever,
which
way.
I
said
it,
even
though
you've
never
done
that
before
that's
what
you
get
from
having
a
metric
based
space
and
a
model
of
the
town,
I
think
now
you're,
maybe
you're.
Introducing
the
question
super
type.
Perhaps,
is
that
what,
if
I'm
trying
to
achieve
a
goal
that
I
really
don't
have
an
obvious
way
how
to
do
it?
What
should
my
behaviors
be?
Is
that
the
question
then.
A
On
the
town
example,
so
let's
say
you
have
the
post
office
in
the
the
town
center,
but
you
never
had
to
use
the
post
office
before
so
you
go
by
every
day,
post
office
there,
but
you
never
had
to
use
it.
So
when
you
do
have
to
use
the
post
office,
is
that
going
to
be
already
part
of
your
model?
So
you
can
just
go.
C
And
enjoy
it
yeah
you're
going
to
learn
everything
as
if
I
I
mean
let's
say
I
I
just
were
you
dropping
her
from
the
new
town
it's
a
sunny
day
and
I
just
started
walking
around
right.
I
will
learn
that
town
I
will
learn.
I
will
remember
what
I
saw
at
different
locations
now
and
I
will
build
up
a
model,
as
I
walk
around
now,
take
a
real
human
doing
this.
That
model
may
not
be
very
long
left
lived,
you
might
forget
things
very
quickly.
You
might
you
know
an
hour
later.
C
You
might
have
forgotten
what
was
on
that
corner
back
there
or
something
like
that,
and
if
something
very
important
happened,
you
might
remember
it
longer,
and
so,
if
something
emotionally
salient
happened,
you
might
say.
Oh
my
god.
I
remember
that
intersection,
that's
where
the
accident
occurred.
So
we
can.
We
can
separate
out
that
sort
of
longevity
of
the
memory
from
the
basic
process.
The
basic
process.
Is
you
observe
you
assign
what
you
observe
to
a
reference
frame?
C
You
observe
you
move
you
go
to
a
new
location,
you
observe
you
sign
what
you
learn
for
the
reference
and
you
move.
You
observe
you
assign
what
you
learned
to
the
new
location,
the
reference
frame.
That
basic
process
is
how
so
you
wouldn't
go
past
the
post
office
and
not
notice
where
the
post
office
was.
You
could
forget
it
very
quickly
again.
C
I
can
say
that
you
might
not
remember
very
long,
but
that
basic
process
is
you
remember
it
the
example
that
I
think
marcus
first
came
up
with,
but
I
use
now
a
lot
is
you
know
you
sit
down
at
your
dining
room
table
and
you
see
where
all
the
dishes
are
for
tonight's
dinner
and
you
see
all
the
salads
over
here
and
the
potatoes
over
there
and
the
vegetable
casseroles
over
here.
Something
like
that
and
in
an
instant
you
remember
that
you
built
the
map
of
the
table.
You
know
and
said.
C
C
So,
yes,
you
will
learn
this
model
and
if
you
have
observed
it,
you
will
learn
it
and
you
might
forget
it,
but
you
will
learn
it,
and
so,
in
this
case
we're
talking
about,
if
I
you
know,
walk
around
town
and
I've
never
needed
to
use
the
post
office
board,
but
I
recognized
I
saw
that
I
know
what
a
post
office
was
and
someone
says:
how
do
you
get
to
the
post
office?
I'll,
remember
I'll
know
how
to
get
there.
I
don't.
C
I
won't
remember
how
to
get
there
I'll
remember
where
it
is
and
then
I'll
be
able
to
calculate
from
where
I
am.
Where
would
I
go
to
get
to
the
post
office?
Animals
can
do
this
very
well.
So
this
is
the
kind
of
thing
that
rats
do
all
the
time.
Things
like
that
insects
do
this.
So
in
that
sense
we
have
this
model
of
the
world.
A
C
It's
it's
a
great
question
and
an
interesting
thought.
We
were
first
and
foremost
just
trying
to
understand
how
the
model
comes
about
at
all.
You
know
how
is
knowledge
represented
in
the
brain?
How
can
you
do
these
things?
How
can
you
perceive
things
now
you
can
you
can
discuss
that
in
the
principles
like
we're
just
doing
without
bringing
up
any
kind
of
motivations
or
goals?
C
You
know
I
could
wander
around
town
aimlessly
just
enjoying
the
day
and
build
a
model,
the
town
I
don't
have
to
be
achieving
anything,
but
clearly
as
biological
organisms,
we
are
heavily
influenced
by
goals
and
motivations,
and
so
we
don't
typically
just
wander
around
randomly.
We
might
be
looking
for
food
or
we
might
be
looking
for
shelter
or
we
might
be
looking
for.
You
know,
friends
or
something
like
that
and
so
and
so
and
those
but
those
emotional
states
or
those
goal-driven
states.
C
C
Oh
my
god,
this
is
really
dangerous.
You
you
don't
want
to
forget
this
and
and
it
it
floods
in
your
cortex
with
with
neuromodulators,
and
that
will
essentially
say
remember
this.
Don't
forget
this.
You
know
so
you're
going
to
form
the
same
synapses
in
some
sense
as
you
did
before,
but
these
are
going
to
be
more
permanent
and,
of
course,
there's
these
interactions
between
the
lower
the
basal
ganglia,
the
emotional
centers
of
the
brain
and
the
cortex.
C
C
And
similarly,
you
know
an
emotional
center
can
influence
what
the
cortex
is
going
to
the
goals
it's
going
to
achieve
and
how
it's
going
to
perceive
things
but
we'd
like
to
think
of
the
cortex
first.
You
know,
let's
understand
how
it
works,
independent
of
those
effective
issues,
and
just
you
know
what
is
the
basic
mechanism
for
how
information
is
stored,
and
then
we
can
later
layer
on
these
goals
to
achieve
certain
things.
A
C
Would
say
it's?
Yes,
that's
what
I
would
say
and
I
would
say
that
with
a
bit
of
humility
and
like
well,
it
could
be
much
more
complicated
than
I'm
thinking
about
it,
but
that's
how
I
think
about
it
right
now.
Yes-
and
you
know
anytime
you're
trying
to
solve
a
difficult
problem,
you
have
to
parcel
out
the
part
that
you're
gonna-
you
can't
solve
it
all
at
once.
A
Nice,
so
I
had
a
few
implementations,
a
few
questions
about
implementation,
so
we
know
static
image.
Recognition
is
not
a
good
way
to
train
or
evaluate
gender
intelligent
agents.
I
think
I
heard
you
talk
about
this
before,
so,
how
do
you
think
we
should
experiment
with
and
evaluate
intelligent
agents
or
in
other
words,
what
are
the
requirements
to
build
an
experiment
environment
for
api
and.
C
A
C
All
right,
so
I
know
I
know
lucas
you're
much
more
of
an
expert
than
this
than
I
am
because
I
know
you're
working
on
this,
but
I
will
I'll
address
that
in
a
more
meta
answer.
If
I
can-
and
I
make
the
point
of
this
in
the
book
in
the
second
part
of
the
book-
I
think
the
way
we
are
thinking
about
machine
learning
benchmarks
today
is
not
how
we're
going
to
think
about
them
in
the
future,
and
I
make
the
analogy
to
computers.
C
When
computers
were
first
starting
to
be
built.
This
is
before
you
know
back
in
the
30s.
They
were
very
goal
oriented.
They
were
task-oriented
devices
like
we're
going
to
compile
the
computer
that
cracks
a
code
or
we're
going
to
build
a
computer
that
solves
you,
know,
build
these
mathematical
tables
and
how
good
is
it
at
doing
that?
But
we
move
towards
the
the
sort
of
general
purpose.
You
know
universal
turing,
machine
concept
where
we
ask
ourselves
well,
what's
a
computer,
we
no
longer
say
we
say:
is
it
a
universal
turing
machine?
C
Well,
does
it
have
the
following
attributes?
Does
it
have
a
memory
for
data
store?
Does
it
have
instruction
sets?
Does
it
have
a
cpu?
It
doesn't
have
an
input
and
output,
and
if
it
has
those
things,
we're
calling
a
general
purpose
computer
and
we
don't
longer
worry
about
what
it's
doing
I
can.
I
can
apply
that
to
running
my
toaster.
I
can
apply
it
to
running
my
car.
C
We
all
perceive
the
world,
we
all
learn
to
model
the
world
to
the
same
methods
and
that's
the
way
ai
has
to
get.
Eventually.
We
have
to
look
at
saying:
okay,
do
we
have
all
the
ingredients
for
being
an
intelligent
machine?
Yes,
well
then,
okay
and
then
there's
different
ones.
We
get
big
ones
and
small
ones,
and
fast
ones
and
slow
ones,
and
we
can
apply
them
to
different
problems
and
different
sensors.
C
So
that's
the
sort
of
meta
answer
to
your
question
in
the
meantime,
we're
not
ready
for
that.
Although
I
make
the
argument
for
that
in
the
book
we
have
to
come
up
with,
you
know,
robotic
or
embodied
benchmarks,
and-
and
I
know
you
know
more
better
than
I
do-
that-
there's
some
really
serious
attempts
in
this
area
right
now
and
people
are
doing
some
really
nice
work
in
this,
but
those
are
still
going
to
be
sort
of
task
oriented.
C
C
Wow
there's
so
much
in
that
it's
and
so
much
it's
you
know
unknown.
Let's,
just
let
me
take
a
couple
things
here.
First
of
all,
we
don't
know
enough
right
today
to
say:
okay,
here's,
here's,
the
the
schematic
diagram,
what
we
have
to
build,
here's
all
the
components.
We
know
how
to
do
everything,
let's
go,
do
it
what
we
we're
proposing
in
the
mentor
and
what
I
propose
in
this
book
is.
C
We
have
a
whole
new
framework
for
thinking
about
this
problem,
a
different
way
of
thinking
about
what
intelligence
is
and,
and
that
framework
is
important
and
some
of
the
pieces
of
that
framework.
We
understand
greatly
in
other
pieces,
we
don't
and
we
have
to
figure
them
out,
but
it's
a
framework
and
the
framework
then
provides
a
roadmap
of
the
things
we
don't
know
and
the
things
we
do
know
and
what
we
have
to
do
to
get
there.
So
could
we
start
building
a
true
agi
today
like
like?
C
I
can
specify
it
and
put
100
engineers
on
it
and
do
it
no
not
at
all,
but
it
is.
It
is
no
we're
not
just
wandering
around
in
the
darkness
and
having
no
idea
what
to
do.
We
can
actually
lay
out
what
we
need
to
do.
Even
if
we
don't
understand
all
of
the
components
like
we
could
say.
Well,
we
don't
really
understand
all
the
issues
about.
You
know
how
columns
coordinate
versus
with
motor
behavior.
C
C
C
I
don't
you
know
and
the
way
these
things
work.
Historically,
we
can
only
look
historically
for
precedent.
You
know
it's
like
an
exponential
curve.
It's
you,
it
looks
like
you're,
not
making
a
lot
of
progress
early
on.
You
know
incremental
small
things.
Oh
we
got
our.
You
know
like
we're
doing
the
sparsely
work
now
we're
showing
these
really
great
improvements
in
speed.
You
know:
50
100,
x,
performance
improvements
on
existing
neural
networks
in
speed,
but
they're
not
really
any
smarter
right.
So
are
those
agi?
C
No,
but
that's
one
step
along
the
way
and
at
some
point
you
know,
you'll
hit
this
inflection
point
and
all
things
will
take
off
the
same
thing
and
sort
of
happen
to
the
computers.
If
you
go
back
and
look
at
the
history
computers,
you
know
people
functioning
around
the
late
30s
and
early
40s
trying
to
build
computers
and
they
were
all
different
types
and
they
were
kind
of
funky
and
different
work.
Didn't
work
very
well
and
there
wasn't
much
agreement.
C
We're
going
to
see
that
happen
here
too,
and
I
think
one
of
the
things
that
momentum
is
trying
to
do
is
to
be
a
catalyst
for
that
we're
trying
to
bring
about
like
showing
the
framework
showing
the
path
we
have
to
take
and
we're
going
to
implement
these
pieces
ourselves.
If
other
people
do
it
great
we're
not
going
to
wait
around
for
other
people
to
do
it,
so
you
know
how
long
will
this
take?
C
I
can
say
confidently
by
the
end
of
this
century.
You
know:
intelligent
machines
are
just
going
to
be.
You
know
as
big
as
computers
are
in
the
later
part
latter
part
of
the
20th
century.
Is
that
going
to
take
10
years
20
years
30
years?
I
don't
know
it's
hard
to
say
where
that
inflection
point
occurs
is
a
very
difficult
thing
to
predict.
A
C
Well,
yeah,
first
of
all,
you
know
we're
doing
this
we're
living
this
right
now
in
the
manta,
so
you
know
we're
doing
machine
learning,
we're
implementing
machining
techniques
based
on
the
brain
and
we're
using
the
brain
to
inspire
them,
but
we're
also
learning
from
the
machine
learning
techniques.
What's
working,
what
doesn't
work
so
that
sort
of
interaction
is
happening
right
now.
C
I
can
tell
you
what
I'm
working
on
and
I
think
I've
made
some
progress
on
I'm
working
on
the
whole
motor
modeling,
going
from
motor
behavior
to
modeling
reference
frames.
How
do
how
does
the
brain
learn?
These
reference
frames
based
on
motor
behavior,
and
I
could
talk
about
that,
but
I'm
doing
approaching
that
from
a
neuroscience
point
of
view
we
are
also
still
myself
and
marcus
and
others
are
working
on
the
issues
of
exactly
how
the
reference
frames
are
implemented.
Grid
cells
are
not
really
just
like
a
you
know:
cartesian
coordinate,
they're
complicated.
C
They
have
other
weird
things
going
on
that
we
don't
understand,
and
so
there's
a
lot
to
be
done.
There.
Understanding
exactly
how
the
reference
frames
are
done.
Another
thing:
that's
it's
and
I've
talked
about
this
in
the
book.
Is
we
don't
really
learn
where
some
sort
of
feature?
If
I'm
learning
a
coffee
cup
or
a
new
object?
I
don't
really
know
where
some
is
in
relative,
a
reference
frame.
C
What
we
think
is
going
on
is
you're
learning
where
objects
are
relative
to
other
objects,
and
so,
and
so
the
whole
thing
is
built
on
a
sort
of
compositional
structure
of
reference
frames
relative
to
reference
frames,
and
we
don't
understand
all
the
details
of
that
yet
either.
So
these
are
some
of
the
big
things
we're
working
on,
but,
as
I
want
to
point
out,
we
don't
have
to
wait
around
for
that.
We're
already
doing
the
sparse
work,
we're
doing
now,
work
on
integrating
dendrites.
C
We
didn't
talk
about
at
all,
but
dendrite,
processing
and
neurons,
which
we
think
allow
us
to
do.
Continuous
learning,
which
is
another
big
problem
with
existing
networks
today
and
we're
just
starting
to
think
about
how
to
implement
reference
frames
is
part
of
this
as
well.
If
and
super
tight.
If
I
missed
anything,
you
want
to
talk
about,
go
ahead
and
pipe
up,
but
but
you
know
we're
we're
going
down
this
path,
we're
not
waiting
around
how
quickly
it'll
process.
I
don't
know.
D
There's
some
sort
of
other
sort
of
big
issues.
We've
talked
about
in
past
research
meetings,
but
we
haven't
really
tackled
them.
You
mentioned
this
a
little
bit,
but
how
do
we
generate
behavior?
You
know.
We've
talked
a
lot
about
how
we,
you
know
how
we
might
take
behavior
as
an
input,
our
motor
commands
as
an
input,
but
what's
the
best
way
to
generate
behavior,
that's
something
we've
discussed
but
haven't
really
figured
out.
You
know
some
of
the
details.
D
C
C
One
a
little
bit
because
it's
particularly
interesting
to
me
it's
sort
of
like
transform
learning
right
I
mean
I
can.
I
can
learn
the
coffee
cup
by
touching
with
my
index
finger
and
then
I
can
infer
it
by
touching
it
with
my
my
ring
finger.
Well,
the
columns
that
are
getting
input
from
my
ring
finger.
Didn't
you
know
they
weren't
part
of
that
learning,
so
I
can't
rely
on
those
columns,
but
maybe
I
can,
and
so
so
how
does
that?
How
do
we
transfer?
C
A
Yeah,
I'm
excited
with
for
discovering
all
this,
and
I
have
one
final
question
before
we
move
on
to
the
questions
from
the
audience
directly
and
I
think
that's
a
personal
question
for
you.
So
given
let
you
just
talk
about
that,
there
is
a
long
road
ahead
and
there
are
things
to
figure
out
and
we
are
not
at
the
point
we
can
just
put
100
engineers
into
it.
So
for
you,
is
it
about
the
journey
figuring
it
out
or
is
it
about
the
destination
you
want
to
you
want
to
get
there?
C
C
Excuse
me
one
second,
my
wife
is
running
around
the
kitchen
here,
my
motivations,
always
the
thing
that
got
me
going
in
this
whole
field
was
not
agi.
It
was.
I
want
to
understand
how
my
brain
worked.
I
just
felt
like,
oh
my
god,
I'm
this
human
I'm
on
this
planet.
What
a
crazy
world
this
is
we
evolved
to
get
this
point.
You
know
we're
a
bunch
of
molecules
running
around
and-
and
now
I
understand
what's
going
on
in
the
world
and
but
what
the
hell's
going
on
my
head.
How
does
this
work?
C
You
know
who
am
I
I
want
to
have
just
a
general
idea
of
what
it
means
to
have
a
brain
that
was,
it
wasn't
like.
I
needed
everything
I
just
think
just
when
I
got
into
this,
like
you,
couldn't
there's
nothing
that
you
explain
how
your
brain
worked,
and
so
I
feel
in
some
sense
I've
achieved
that
goal.
I
have
this
the
framework
of
the
thousand
brain
theory
sort
of
like
okay.
I
have
a
pretty
good
idea
what
the
basic
thing
is
going
on
in
our
head,
so
that
was
personally.
C
C
Most
people
didn't
believe
that
most
people
believed
most
ai
researchers
were
going
back
now
into
the
80s
right,
mostly
introduced,
then,
and
even
today,
think
you
probably
can
ignore
how
the
brain
works.
It
doesn't
really
matter
we'll
figure
it
out.
We
don't
need
to
study
the
brain
to
do
that.
We'll
figure
out
how
to
do
this
because
we're
good
smart
engineers
and
scientists.
C
I've
always
felt
that
that
was
unlikely
to
be
successful.
I
felt
that
the
brain
we
knew
it
was
so
much
so
different
than
the
way
people
were
thinking
about
ai
that
we
would
probably
not
get
there
if
we
weren't
studying
the
brain,
and
but
I
couldn't
prove
that,
and
I
didn't-
I
never
tried
to
convince
other
people
of
it.
I
just
said
we'll
see
I'm
going
to
study
the
brain.
You
can
go,
do
your
own
thing
and
so
far
my
predictions
have
turned
out.
True.
C
You
know
we're
not
really
close
to
having
agi
with
true
ai,
but
we
are
getting
close
to
understanding
how
the
brain
works
so
and
we
started
having
this
roadmap
how
to
get
there.
So
I
think
that's
cool.
I
of
course,
I'd
love
to
see
these
things
come
about,
but
I'm
also
realistic
that
you
know
in
my
lifetime.
I
will
not
see
you
know
the
ultimate
impacts
of
true
ai,
because
it's
going
to
be
playing
out
for
many
many
decades
into
the
future
and
that's
some
reason.
C
One
of
the
reasons
I
wrote
the
last
part
of
the
book-
and
maybe
I
should
talk
about
that,
but
the
second
part
of
the
book
is
all
about
ai,
but
the
last
part
of
the
book
is
about
humanity
and-
and
I
said
to
myself,
what
are
the
implications
of
having
intelligent
machines?
Truly
intelligent
machines
are
smarter
than
we
are
that
are
can
have
different
embodiments
that
you
know
have
different
parameters
and
attributes
associated
with
them.
C
We
think
going
forward
into
the
into
the
not
even
the
distant
future,
but
just
even
a
few
hundred
years.
You
know
our
world's
gonna
be
really
different
in
a
few
hundred
years,
even
a
hundred
years.
It's
even
at
the
end
of
this
century.
It's
gonna
be
really
different
with
all
these
intelligent
machines
and
how
should
we
think
about
ourselves
as
humans
and
how
should
we
think
about
humanity
in
that
world
in
that
world?
C
A
B
A
A
B
B
E
Okay,
I
was
having
a
conversation
with
a
octopus
and
he
pointed
out
to-
or
she
pointed
out
to
me
that
not
all
creatures
think
the
same.
I
thought
about
it
for
a
while
and
said
gee,
that's
very,
very,
very
interesting.
I'm
descended
from
a
squiggly
little
proto-chimp
and
she's
descended
from
a
clam
or
a
proto-clam.
C
C
Let's
talk
about
birds,
birds
are
smart
too,
and
they
don't
have
a
proper
near
cortex.
Now,
as
we
got
into
understanding
the
thousand
brains,
theory
and
understanding
the
role
of
reference
frames
and
movement,
it
became
clear
that
this
is
not
just
a
way
of
understanding
the
world.
It
is
probably
the
way
of
understanding
the
world
that
it's
hard
to
imagine.
Maybe
it's
a
failure
of
imagination,
but
it's
hard
to
imagine.
C
You
can
understand
the
world
without
helping
to
model
the
world
it's
hard
to
have
imagine
having
a
model
of
the
world
if
you
don't
move
and
build
a
model
of
the
world
in
some
reference
frame,
and
so
what's
interesting.
If
you
look
at
birds,
they
don't
have
a
neocortex,
but
they
have
these
two
areas.
I
don't
remember
the
correct
name
from
them,
but
they're
called
blobs.
C
They
have
some
some
such
blob
and
another
blob,
and
it
turns
out
that
they
are
looking
there's
more
and
more
evidence
that
those
two
blobs
are
some
sense
doing
the
same
function
using
the
same
type
of
reference
frame
calculation
as
mammals
do.
That
is,
that
mechanism,
whether
whether
we
we
have
a
common
evolutionary
origin
for
it
and
what
is
separately
evolutionary
derived,
but
it's
looking
more
and
more,
like
birds
have
been
equivalent
to
a
neocortex.
C
But
it's
not
structured
like
a
near
cortex,
it's
structured
in
two
blobs,
as
opposed
to
a
layer
of
neural
tissue.
But
this
you
see
the
same
types
of
cells
and
the
same
types
of
interactions
with
them,
and
so
I
would
imagine
that
birds
actually
do
perceive
the
world
in
some
ways
like
we
do
the
octopus.
I
think
there
was
some
evidence
of
the
same
thing
for
the
octaves,
but
I
can't
remember
it
now,
so
I
I
have
to
be
careful
about
that.
C
I
know
this
evidence
exists
for
birds
and
I
think
it
also
exists
for
octopus,
but
an
octopus
clearly
can
move
its
limbs
to
oh.
You
know
we've
seen
the
pictures
of
them
opening
jars
and
solving
problems,
they
have
a
model
of
the
world
and
they
have
you
know
reference
frames
going
on
in
that
head.
So
I
actually
think
they
we
do
understand
the
world
in
a
similar
way.
There's
a
famous
paper.
It's
called
something
about.
It
was
about
bats.
It
was
called
like
I
didn't
remember
the
name
of
it.
G
C
People
are
arguing
that
that
you
know
bats
must
perceive
the
world
completely
different
than
we
do
because
they
use
echolocation.
But
there's
another
argument
that
says:
no:
they
don't
here's
a
here's,
an
exam
analogy.
I
can
learn
the
world
my
world
through
touch
right.
If
I
was
blind,
I
could
I'm
and
there
are
blind
people
and
they
learn
a
model.
C
Learn
the
model
of
the
world
and
I
have
both
of
those
senses.
I
have
a
touch
sense
and
a
vision.
Sense.
I
can.
I
can
not
touch
things
look
and
I
can
touch
and
look
and
not
touch
things,
and
yet
I
build
a
very
similar
model
of
the
world.
Yes
they're,
different
attributes.
If
I'm
blind,
I
don't
see
the
colors,
I
can't
see
a
distance,
but
I
still
model
have
a
similar
model
to
the
world.
C
Helen
keller,
who
was
deaf
and
blind,
had
she
could
converse
and
understood
the
world
just
like
you,
and
I
did
so
a
a
bat
or
an
octopus
or
bird
has
different
senses,
but
in
the
end
they're
going
to
build
a
model
of
the
world
and
they
live
in
the
same
world.
We
do
and
that's
at
least
the
extent
that
the
part
of
the
world
they
live
in
is
the
same
as
we
do.
I
think
they
actually
would.
They
would
think
like
we
do,
they
would
perceive
the
world
just
like.
C
B
E
B
H
I
B
H
Great
so
yeah
in
the
book,
the
concepts
well
at
least
the
way
the
concepts
were
explained
I
at
least
found
them
quite
accessible.
But
then,
when
you
sit
down
to
write
an
implementation
of
it,
you
realize
it
gets
complex,
really
really
quickly,
and
so
my
question
was:
is
there
any
way
to
experiment
with
the
tbt
theories,
to
get
a
deeper
understanding
of
the
mechanisms,
or
is
there
any
opportunity
to
help
build
such
a
framework
if
it
doesn't
exist?
D
Matthew
when
you
said
experiment
with
it,
did
you
mean
running
neuroscience:
experiments
to
test
different
parts
of
the
theory,
because
there's
a
lot
of
stuff
that
could
be
done
there
or
do
you
mean
sort
of
from
a
coding
implementation
standpoint?
Do
we
have
software
frameworks
or
or
things
like
that,
to
experiment
with.
H
D
D
So
you
could
definitely
take
that
code
base
as
a
thing
and
try
to
expand
it
and
work
with
it
and
try
to
see
if
there's
different
ways
to
you
know,
expand
the
theory
in
different
ways
and
try
to
do
ad
goal,
oriented
behavior
or
some
of
these
other
things
we've
talked
about.
We
could
definitely
try
doing
that.
However,
that
stuff
was
not
designed
to
be
practically
useful.
You
know,
I
think
it's
going
to
be
very
hard
to
take
real
world
input
and
try
to
get
real
systems
working.
It
was
more.
D
These
are
models
of
the
neuroscience,
and
so
you
know
if
you
want
to
expand
on
the
neuroscience
theories,
I
think
you
could
probably
use
that
code
base
on
the
practical
side.
That's
a
good
chunk
of
our
research
agenda.
Right
now
is,
as
jeff
you
know
alluded
to
earlier,
is
taking
different
aspects
of
thousand
brains.
Theory
incorporating
them
into
practical
systems
and
then
sort
of
showing
how
we
can
you
know,
get
properties
that
are
difficult
to
get
otherwise,
and
so
for
that
you
can
go
to
newpic.research
and
follow
along.
You
can
look
at
our
research.
D
C
D
Those
are
available
in
newpick.research,
available
in
open
source.
A
lot
of
the
you
know,
some
of
the
stuff
we've
done
with
partners
is
not
available,
but
the
four
ideas
that
that
were
in
house
events
in
the
house
of
dance
paper.
How
do
you,
how
you
might
you
create
deep
learning
systems
that
are
extremely
sparse,
that
show
properties
of
robustness
and
try
to
get
you
know,
models
that
are
as
sparse
as
possible.
We
have
code
for
that.
We
have
our
code
for
incorporating
dendrites
and
working
continue.
D
C
C
I
don't
want
to
put
words
in
your
mouth,
but
you
know
because,
like
the
code
that
super
tight
was
just
describing
this,
this
code
we're
doing
right
now
on
sparsity
and
dendrite
learning,
that's
really
deep,
technical
stuff
that
we're
trying
to
be
practical
with
you
know
it's
not.
I
wouldn't
call
it
a
platform
for
experimenting.
C
It's
like
it
takes
a
team
of
you
know
dedicated
machine
learning
people
to
do
it.
You
know,
but
it's
an
interesting
idea
that
that
you
know
sort
of
a
some
sort
of
intermediate
modeling.
That
would
illustrate
the
properties
and
let
people
experiment
with
it
that
maybe
not
yet
not
yet,
like
you
know,
a
world-class
machine
learning,
you
know
benchmark
performing
level,
but
would
be
more
of
illustrating
the
principles
and
allow
people
to
to
explore
those
principles.
Different
ways
of
implementing
would
be
very
kind
of
really
cool
thing
to
create.
C
H
C
D
H
G
G
G
During
this
session
we
have
approximately
150
000
cortical
columns
in
the
in
the
neocortex
and
whenever
we
have
a
learning
phase,
where
we,
for
example,
learn
about
a
new
object
like
a
cup,
so
we
build
an
allocentric
model
of
that
cup
by
maybe
movement
and
touch
a
certain
subset
of
let's
say,
1500
of
those
column.
Cortical
columns
will
be
active
during
that
learning
phase
and
we
and
each
one
of
them
builds
an
individual
model.
They
have
a
slightly
different
inputs
because
they
probably
have
different
sensory
origins.
G
G
G
Otherwise
they
would
be
looking
or
maybe
I
missed
something
in
my
understanding,
but
in
my
opinion,
out
of
the
out
of
the
over
100
models
each
one
of
them
has.
They
have
to
all
turn
to
the
same
page,
basically
the
same
map,
because
they
want
to
they're
trying
to
identify
or
disambiguate
which
part
of
the
cup
they're
actually
touching,
but
first
they
have
to
select
that
subset,
otherwise
they
would
be
searching
in
a
huge
search
space.
G
C
During
the
learning
process
so
joe,
let
me
make
sure
I
understand
your
question.
The
question
is:
are
you
asking
if
I
look
at
an
individual
column,
it's
going
to
be
sensing
some
small
part
of
something,
and
how
does
it
know,
are
you
asking?
How
does
it
know
that
that's
a
part
of
a
cup
because
it
could
be
part
of
lots
of
things.
G
G
In
the
general
idea,
yes,
they
do
yeah,
right
and
and
in
the
in
the
in
the
paradigm
analogy
that
you
explained
in
the
in
the
first
part
of
the
book,
you
were
explaining
how,
for
example,
each
one,
for
example,
each
finger.
If
it's
one
each
finger
in
the
that
is
touching,
the
cup
is
trying
to
decide
which
part
of
the
cup
they're
actually
touching,
they're
trying
to
identify.
So
I'm
envisioning,
all
1500
or
so
active
columns.
G
G
C
D
Let
me
see,
maybe
joe,
because
we
have
limited
time.
Let
me
see
if
I
can
rephrase
it
in
some
ways,
and
you
can
it's
sort
of
you
know
you
have
each
cortical
column
is
getting
different
sensory
input
but
they're
sensing
from
the
same
object,
they're
getting
sensory
features
and
there's
sort
of
coordination
among
columns
to
figure
out
what
the
object
is,
but
then
knowing
what
the
object
is
should
help
each
cortical
column
know
where
they
are
on
that
object
on
that
coffee
cup
and
how
does
that
coordination
and
locations
happen?
You
know.
D
C
So,
thank
you,
so
you
don't
you
don't
have
a
question:
how
a
single
column
can
do
this?
Okay
got
it.
You
know
I
maybe
we'll
try
to
leave
it
this
good.
I
don't
spend
too
much
time
in
this.
I
made
a
brief
mention
of
this
in
the
book,
and
this
is
a
an
idea.
First
brought
up
by
marcus
lewis.
Is
that
the
there
is
a
coordination
between
the
columns
it
appears
to
be
imagine
I
have
five
fingers
touching
the
cup
you
could
say:
oh
those
five
fingers
are
completely
independent.
C
They
don't
know
where
each
other
is,
and
so
each
one
on
its
own
is,
is
there's
no
coordination
between
their
locations,
they're
trying
to
vote
on
the
object,
but
they
they
don't
know
where
the
other
ones
are.
We
don't
believe.
That's
actually
the
case.
What
we
believe
it
the
case
is
that
they
do
know
their
relative
positions
to
each
other,
and
I
gave
this
example
in
the
book
when
I
was
talking
about
five
people
landing
in
a
town
at
once
and
and
trying
to
figure
out
where
they
are.
C
If
they
just
look
at
the
features,
it's
like
a
bag
of
features
problem,
but
if
they
know
where
their
relative
position
to
each
other
is
they
can
very
quickly
infer
what
what
town
they're
in
so
we
believe
that
there
is
this
coordination,
that's
going
on
between
columns
where
they
know
their
relative
positions
to
each
other.
We
do
not
have
a
good
model
for
how
that's
happening.
C
There's
some
candidate
projections
in
the
cortex
in
the
brain
that
might
be
playing
a
role
in
this,
but
we
don't
know
yet
how
that's
happening
so
it
works
on
a
single
column
basis.
It
works
on
a
multi-column
basis
if
you,
but
you
have
some
limitations,
but
if
you
can,
if
we
can
get
this
sort
of
relative
position,
they
can
they,
then
then
they
don't.
They
don't
really
have
to
vote
on
their
locations.
They
just
have
to
know
their
relative
positions
and
that
itself
is
sufficient
to
resolve
the
ambiguities.
G
C
But
here's
what
we
can
do
there
are
researchers
are
interested
in
our
theories
and
and
first
and
foremost
they
say.
Oh
I've
got
some
data
that
already
supports
this
or
contradicts
it.
So
there's
a
lot
of
this
unassimilated
data
that
we
can
mine,
we
can
mine
on
our
own.
We
can
mine
it
with
their
help.
Also,
there
are
ongoing
experiments
which
just
happen
to
be.
You
know
supportive
or
not
supportive,
so
we-
I
don't
know
if
they're
doing
it
based
on
our
work,
because
these
experiments
can
take
years
to
do
you
know.
C
So
it's
a
combination
of
hey
they're,
interested
in
helping
us
b.
They
got
some
data,
they
think
that
would
be
useful
or
c
they
may
they
may
plan
some
future
experiments.
So
we
have
some
people
doing
that
too.
But
it's
a
it's
a
very
complex.
It's
not
so
simple.
Like
here's,
a
theory,
please
test
it
that
doesn't
work.
B
C
Supertime
I
just
spoke
at
this
neuromorphic
computing
hardware
conference
recently
and
we
were
invited
to
speak
there
all
the
time,
because
this
is
a
big
question
right.
You
know
what
is
it?
What
is
the
physical
substrate
underlying
all
this
stuff
today,
I'll
give
a
simple
answer
then,
and
someone
else
can
chime
in
you
know
today.
The
today's
gpus,
for
example,
are
not
really
well
suited
for
doing
the
kind
of
sparse
computations
and
the
complex
computations
that
we
see
in
the
thousand
brains
theory.
C
C
You
can
also
do
this
on
gpus,
but
they're
not
ideal
either
so,
and
then
I
think
so.
Those
are
the
options
today
really
like.
You've
got
gpus,
cpus
and
fpgas,
but
in
the
future
there's
going
to
be
new
architectures
or
derivatives
of
those
three
and
although
we're
in
a
lot
of
discussions
with
people
about
that,
it's
not
clear
what
the
right,
the
ultimate
answer
is
going
to
be.
You
asked
what
are
the
bottlenecks?
D
Well,
you
know,
instead
of
for
us
everything
sort
of
relies
on
sparse
structures,
like
you
said,
and
so
having
really
efficient
circuits
that
can
handle
really
unstructured,
sparse
systems
seems
to
be
a
real
impediment
right
now
and
and
what
those
circuits
are
not
always
clear.
Memory
is
always
a
bottleneck.
D
You
know
when
you're
doing
really
sparse,
sparse
computations.
You
want
to
be
able
to
take
just
the
non-zero
pieces
of
one
thing
and
then
non-zero
piece
of
the
other
thing
and
then
bring
them
together,
and
today's
kind
of
dense
matrix
multiplication
focus
systems
are
just
not
oriented
for
that,
and
it
turns
out.
You
actually
need
very
you
know
specific
types
of
memory
channels
and
memory
bandwidths
too,
to
make
that
efficient.
So
there's
a
lot
of
different
challenges
in
that,
but
you
know
sparse.
C
It
reminds
me
of
again
I
like
to
think
of
the
history
of
computing.
You
know
in
the
beginning
it
was
you
could
ask
john
von
neumann
what
are
the
bottlenecks
of
building
computers
and
his
answer.
I'm
sure
you
give
me
an
answer,
but
but
but
the
you
know,
ultimate
inventions,
such
as
disc
storage
and
solid-state
memory,
or
just
even
the
transistor.
C
You
know
which
just
leveled
in
the
integrated
circuit
completely
changed,
blew
everything
out
of
the
water.
These
guys
couldn't
anticipate
any
of
that.
I
mentioned
this
because
I
think
we're
going
to
see
similar
types
of
innovations
occurring
in
this
field
that
we
can't
anticipate
yet
just
to
pick
one
thing
I
think
we
saw
this
company
a
meeting
with
a
company
called
rain.
Is
it
rain,
neuromorphics
or
rain?
I
forget,
and
they
have
a
completely
new
way
of
thinking
about
how
to
build
neuromorphic
computers.
Is
it
the
right
way?
C
I
don't
know
it's
pretty
cool,
it's
really
smart,
but
you
know,
I
think,
there's
gonna
be
innovations
which
we
can't
anticipate
today,
but
in
today's
world
a
asuba
type
pointed
out,
it's
very
difficult
to
do.
Sparse
computations,
especially
on
gpus
and
and
so
we're
working
with
more
of
the
fpga
and
cpu
side
of
things.
C
I
didn't
talk
about
sparsity
at
all
in
the
book.
In
case
you
were
wondering
about
that.
I
just
felt
it
wasn't
necessary
for
the
audience
intended
audience
in
the
book.
It
was
hard
enough
for
many
people
to
get
through
it
anyway.
So,
but
it's
underlying
everything
pretty
much.
I
can't
think
of
a
single
thing,
we're
thinking
about
doing
or
doing
that
doesn't
rely
on
sparsity
in
one
way
or
the
other.
C
As
I
said
earlier,
it's
not
all
the
classic.
You
know
very
highly
sparse
population
codes.
There
are,
if
you
think,
about
grid
cells.
For
example,
grid
cells
are
very
complicated
beasts
and
they
rely
on
other
things.
They
do
rely
on
on
scalar
activation.
They
also
rely
on
on
oscillations
and
and
and
where
cells
are
spiking
in
different
points
of
the
different
phases
of
the
theta
cycle.
C
So
there
are
other
things
that
are
going
on
that
are
that
are
much
more
complicated,
but
so
it's
not
all
everything
is
just
a
bunch
of
sdrs,
but
the
sdr
is
still
the
the
primary
underlying
data
structure.
I
think
we
still
believe
that
for
everything,
pretty
much
there's
other
things
going
on
too,
but.
D
C
But
if
you
look
at
grid
cells
per
se,
they're,
not
as
sparse
as
we
talk
about
in
sdrs
grid
cells
themselves,
are
are
still
sparse,
but
they're
not
as
sparse
and
they
are.
They
are
modulated
by
frequency
phase
and
by
frequency
of
firing.
So
there
are
different
parts
of
the
theory
that
aren't
purely
purely
is
these
very,
very
sparse
sdrs,
but
they're
still
sparks
grid
cells
are
still
sparse,
they're,
just
not
as
far
as.
A
B
I
Okay,
first
of
all,
it's
an
honor
to
meet
you
all
and
thank
you
so
much
for
organizing
this
event.
So,
while
reading
the
book
you
mentioned
briefly
about
how
the
tpt
could
extend
to
higher
level
of
cognition
such
as
language,
so
I
was
wondering
if
you
could
elaborate
on
how
language,
comprehension
and
language
production
could
equate
to
movement
as
implied
by
the
dvd
and
how
these
ideas
might
extend
to
current
state
of
the
art
in
natural
language
processing.
C
Okay,
I
was,
I
was
with
you
to
the
very
last
part
of
your
question,
because
so
we'll
get
there.
First
of
all,
this
issue
of
the
common
cortical
algorithm
that
the
same
algorithm
is
applying
everywhere,
including
the
language
areas
of
the
brain.
It's
just
an
empirical
observation
and
we
we
accept
it.
It's
like
okay,
the
neural
circuitry
looks
almost
the
same
everywhere,
and
that
was
the
mount
castle's
theory.
So
and
honestly,
when
you
first
start
thinking
about
it's
really
hard
to
imagine.
Well,
what
is
the
thousand
brains?
C
How
do
these
modeling,
these
sensory
motor
modeling
systems?
How
do
they
model
concepts
and
language,
and
things
like
that?
So,
as
you
know,
there's
an
entire
chapter
in
the
book
about
this
and
I
did
the
best
job.
I
could
to
explain
this
and
and
talk
stepping
back
from
language.
I
talked
about
mathematics
and
politics
and
how
you
can
think
about
conceptual
knowledge
in
a
framework
and
a
reference
frame
and
how
the
equivalent
for
what
a
motor
behavior
a
movement
behavior
is.
C
It
doesn't
have
to
be
a
physical
movement,
but
it's
still
moving
from
location,
the
location
of
a
reference
frame
and
the
same
process.
Properties
of
metric
spaces
apply.
So
there's
a
whole
chapter
on
that.
I'm
not
going
to
review
that.
I
think
the
question
of
language
is
particularly
tricky
and-
and
I
said
that
in
the
book-
and
so
I
could
only
address
one
small
part
of
language
at
the
time.
I
was
writing
this,
and
that
was
the
idea
of
of
a
recursive
structure,
which
is
a
very
a
key
component
of
language
and
linguists.
C
We'll
talk
about
that,
and
so
the
thousand
brain
theory
does
explain.
At
least
it
gives
a
framework
for
understanding
how
neurons
create
recursive
structure
in
the
world,
but
then
you
said
well,
how
can
we
apply
this
and
prove
the
current
state-of-the-art?
You
know
language
things
well,
we
are
currently
working
on
language
models
and
using
sparsity
and
some
of
the
other
principles
to
improve
those
existing
language
models,
but
we're
not
anywhere
close
to
building
machines
that
understand
language
the
way
humans
do.
I
think
we
have
a
lot
to
go
there.
C
Yet
we
don't
understand
that
yet,
and
so
I
think
in
the
short
term,
can
we
improve.
You
know
gpt
3
and
you
know
different
language
models
while
we're
doing
that,
we're
showing
you
we
can
do
that
already,
but
going
to
the
whole
thousand
brains,
theory
aspect
and
showing
how
you
can
make
a
machine
that
really
truly
understands
language,
we're
just
not
there.
Yet
we
have.
We
have
a
ways
to
go,
so
I
wish
I
had
a
better
answer.
B
J
I
have
one
general
question
about
how
up-to-date
are
your
ideas
in
your
new
book,
and
by
this
I
mean
that
I
understand
that
there
is
a
very
long
process
between
the
writing
of
the
last
page
and
the
publication
of
the
book.
Maybe
up
to
one
year,
I
guess
yeah.
So
are
there
any
recent
ideas
or
corrections
that
you
wish
to
have
included
in
your
new
book.
C
For
my
personal
side,
what
I've
been
working
on
is
understanding
how
reference
frame
transformations
occur
in
the
cortex
using
the
thalamus.
I've
also
been
working
on.
I
think
I
have
a
really
clear
idea
what's
going
on
there,
and
I
also
have
a
very
clear
idea
how
many
columns
in
the
cortex
represent
dimensions
in
space.
C
C
That's
what
I've
been
working
on
other
things
we've
done
is:
we've
done
a
lot
of
work,
as
we
already
said
in
the
machine
learning
side
of
the
world
in
terms
of
improving
performance
on
neural
networks
and
implementing
you
know
sparsity
and
other
aspects
of
it
and
that's
not
in
the
book.
I
could
have
talked
about
that.
So
those
are,
I
think,
and
then
I
think
in
the
super.
Tired
lucas
wanted
to
say
something
else
that
I
forgot.
I
D
No,
I
mean
there's
a
lot,
there's
a
long
list
of
stuff
that
we're
working
on
in
the
in
terms
of
incorporating
machine
learning,
one
of
the
things
that
comes
up
in
some
of
our
recent
work,
that
we
is
sort
of
exactly
what
the
learning
rules
should
be
and
what
the
how
learning
should
happen,
particularly
when
we're
doing
continual
learning.
We
talked
we've
done
that
a
little
bit
of
that
in
the
htm
world.
But
you
know
how
unsupervised
or
self-supervised
learning
should
happen
in
a
way.
So
you
can
create
powerful
representations
in
a
way.
D
C
On
yeah,
I
think
I
think,
since
our
last
question,
it's
a
good
thing
to
point
out
that
this
is
an
ongoing
concern.
We're
trying
to
be
very
open
about
everything,
we're
doing
we're,
trying
to
be
honest
about
what
we
don't
understand
and
you
point
out.
The
book
is
a
point
in
time.
It
is
over.
You
know
I
stopped
writing
it.
A
year
ago
over,
you
know
a
little
bit
over
a
year
ago,
so
we're
and
we're
you
know
we're
just
doing
our
best
to
try
to
progress
the
field.
C
A
B
A
We
still
have,
we
still
have
a
lot
of
questions.
Those
are
really
good
questions
and
both
in
slido
and
zoom
qa
in
chat
and
htm
forums.
So
what
we
can
do,
what
we
do
in
every
brains
at
base
session
is
that
we
collect
all
those
questions
in
a
word
document
and
we
provide
the
link
to
both
the
attendees
and
the
penalties,
and
the
penalties
can
answer
their
question
directly
in
the
word
document
later.
So
would
you
be
willing
to
do
that
as
well?
Jeff.
D
Yeah
and
something
charmaine's
been
actually
been
doing
recently
is,
is
doing
our
research
meeting
at
the
beginning.
Sometimes
she
will
ask
some
of
the
questions
from
the
community.
Maybe
we
can
just
answer
them
a
few
of
them
live.
We
can
pick
a
few
of
them
and
then
answer
them
live
in
our
research
meetings
and
that
would
be
pretty
cool.