►
From YouTube: COSYNE 2020 Recap Part 3 - Numenta Research, Mar 18
Description
Florian Fiebig discusses his attendance at COSYNE 2020.
Discussion at https://discourse.numenta.org/t/cosyne-2020-recap-numenta-research-mar-9/7268
Part 1: https://youtu.be/BLBPqIOyMgo
Part 2: https://youtu.be/JM5DE2BChT0
B
B
We
talked
about
the
the
the
talk
bad
with
moser
like
this
richard
gardner
work
about
this.
You
know
pinging
sweep
of
cells
and
so
what
I
want
to
turn
to
now,
but
because
we've
of
course,
been
thinking
a
lot
about
wider
uses
of
grid
cells
beyond
just
navigation.
B
B
So
so
this
talk
kind
of
started
off
with
a
bit
of
a
reminder
of
the
work
that
she
did
in
2017.
That's
this
killian
2017
paper
down
here
where
they
had
primates
freely
viewing
images
and
recording
from
media
entering
cortex
and
what
they
showed.
Is
that
obviously
on
any
particular
image
you
know
where
the
cards,
the
eye
movements
right
are
going
and
stopping
depends
a
lot
on
the
individual
image.
So
you
never
get
like
a
nice,
even
coverage
of
the
image.
B
But
if
you
have
lots
and
lots
of
images
to
explore-
and
you
know
these
monkeys
love
to
look
at
these
images,
so
these
are
like
tetra
recordings,
then
you
can
get
a
very
nice
covering
of
the
space
and
can
start
you
know
looking
for
signatures
of
grid
like
cells
in
in
you
know,
free
viewing
of
pictures.
B
So
you
don't
have
much
problem
motivating
these
monkeys
because
there's
lots
of
interesting
pictures
to
look
at
and,
as
you
can
see,
you
can
trace
out
the
the
cards
and
if
you
overlay
many
many
of
these
pictures-
and
you
look
at
the
firing
responses
in
media
entrepreneur
cortex,
then
you
do
find
good,
like
cells.
B
So
there
was
evidence
for
for
grids
and
even
like
for
for
something
like
boundary
vector
cells,
so
that
fire
only
when
the
when
the
monkey
is
looking
along
the
border
of
the
picture
and
one
of
the
one
of
the
questions
was
I'm
sorry,
you
have
a
yeah.
I
don't.
C
Understand
this,
yet
okay,
so
these
this
like
grid
like
pattern.
This
is
the
this,
this
red
dots
here,
that's
one
cell
they're
recording
from.
B
C
It
seems
like
it
would,
you
know,
grid
cells
re-anchor
in
different
different
environments,
and
so
you
would
think
that
they
might
re-anchor
in
different
images.
B
C
If
they
recorded
from
a
single
cell
as
a
rat
in
different
environments,
they
wouldn't
see
this.
It's
only
in
a
in
a
particular
environment.
Would
you
see
the
grid
like
pattern
so,
and
so
the
question
here
we're
seeing
multiple
images
and
and
now
we're
seeing
this
grid
like
sort
of
grid
like
pattern
to
it.
That
could
have
been
one
image
or
if
it's
multiple
images,
and
then
that
implies
that
the
grid
cells.
D
B
C
E
Sorry
can
I
interject
I'm
sure
they
addressed
this,
but
looking
at
these
responses
I
mean:
did
they
make
sure
that
these
aren't
just
like
those
locations
in
visual
space?
The
monkey
is
fixating
that
have
some
periodicity
and
obviously
there
tend
to
be
a
lot
of
them
at
the
borders.
B
E
B
B
Yeah
yeah,
I
guess
anyway,
so
the
point
of
the
talk
was
not
much
like
really
this
original
finding,
but
to
take
it
a
step
further
and
ask
I
mean
that
is
like
free
viewing,
but
is
that
also
true
in
when
with
with
covert
attention?
B
So
they
changed
the
task
a
little
bit
and
I
think
in
interesting
ways,
so
they
now
had
the
the
monkey
fixate
on
a
dot
and
the
task
was
quite
simply
to
release
a
lever
when
a
moving
dot
that
is
outside
of
the
fixation
changes
color
right,
so
you
have
a
fixation
onset.
B
Then
you,
you
start
like
some
some
moving
dot
that
moves
around
on,
like
you
can
see
these
these
outlines
here
and
at
some
point
between
700
and
2000
milliseconds,
that
that
moving
dot
will
change,
color
and
then
you're
supposed
to
have
a
700
millisecond
time
window
in
which
you
can
like
release
the
liver.
And
if
you
do
that
right,
you,
you
get
a
reward,
and
the
interesting
thing
about
that
is
of
course
this
is
now
covert
attention.
B
So
this
is
not
psychotic.
Eye
movement,
in
fact,
like
the
trials,
are
aborted
when
the
when
the
monkeys
move
the
ice
and
the
the
the
almost
all
of
the
failures
in
the
in
this
task
are,
as
a
consequence
of
monkeys
not
being
able
to
fixate,
rather
than
monkeys,
not
being
able
to
use,
convert
attention
to
follow
the
the
dot,
and
so
the
the
the
curves
that
these
that
the
dot
is
doing
is.
B
I
think
it's
called
a
hamiltonian
or
something
there's
like
a
special
movement
pattern,
but
kind
of
make
sure
that
you
evenly
cover
the
space
that
there's
no
bias
with
regards
to
sort
of
attention
on
a
specific
part
of
the
picture,
and
so
so
that
that
should,
like.
B
You
know,
also
potentially
address
your
your
previous
concern
areas
so
like
they
made
sure
that
the
full
trials
always
cover
the
space
evenly,
as
you
can
see
like
from
these
snake
like
outlines
here
and
so,
and
they
did
find
even
on
a
covert
detention
with
no
psychotic
eye
movement.
Gridlock
firing,
as
you
can
see
from
like
the
this
one
of
this
example
cell.
Here,
you
can
see
in
red
the
the
individual
like
traces
here.
B
You
know
attempts
to
follow
these
snake
lines,
so
that's
like
starting
at
stopping
points
and
all
of
these
together
nicely
evenly
cover
the
space
and
what
you
see
with
the
black
dots
here
are
the
individual
spikes
on
one
trial
and
in
the
background
you
have
a
multi-trial
average,
and
so
what
you
see
is
that
there's
that
most
of
the
spikes
in
this
one
trial
are
occurring
in
the
hot
zones
of
the
multi-trial
average
heat
map,
so
that
just
goes
to
the
question
of
sort
of
what
is
the
consistency
of
this?
B
Well,
it's
good
enough,
at
least
for
I
think,
like
in
one
session.
They
do
I
mean
these
are
like
a
total
of
like
six
of
these
snakey
lines.
B
I
think,
and
then
I
do
like
a
set
of
five
of
those
like
two
two
to
seven,
I
think,
is
something
it's
a
in
the
paper,
and
so
that
is
then
what
you
get
for,
like
an
average
map,
and
you
see
in
the
autocorrelogram
that
okay,
these
are
like
grid-like
responses,
and
that
was
then
the
this
became
this
2018
paper,
android
cortex
receptive
fields
are
modulated
by
special
tension,
even
without
movement,
and
so
one
question
one
might
ask
okay.
B
If,
if
we're
going
to
have
grid
cells
now
does
that
mean
we
also
have
border
cells,
as
it
also
means
we
have
something
like
head
direction:
cells
like
the
direction
of
cards,
for
example,
in
the
freezer,
counting
task
or
the
direction
of
the
attention.
B
So
you
sell
some
of
that
here.
I
guess
these
are
yeah
here.
You
can
see
like
a
cell
right
that
fires
exclusively
on
the
left
border
here.
So
there
you
go
like
a
heat
map
here
which
actually
has
like
a
strong
heat
spot
there,
or
is
that
a
separate
cell?
B
I'm
not
sure
you
get
like
these
band-like
cells
together
with
these
grid
cells,
and
they
did
some
analysis
to
show
that
actually,
the
relation
like
the
prevalence
of
the
different
types
of
cell
types
is
the
is
the
the
ratio
that
you
would
expect
from
from
spatial
grid
cells
and
border
cells,
and
you
know,
vector
cells.
B
So
this
was
an
analysis.
They
did
on
cigar
direction
cells
and
it
turns
out
they
kind
of
neatly
divide
into
two
categories.
One
is
cells
that
fire
prison
card,
so
sort
of
upcoming,
like
when
you're
about
to
exploring.
C
B
C
Reviewing
you're
giving
a
talk
that
covers
multiple
papers
is
that
right.
B
Yes,
yeah
I
mean
the
topic
is
visual
grids.
I
guess
maybe
I
should
have
shown
this
one
first,
like
they've,
ordered
it
slightly
off.
C
In
the
covert
attention
task,
it
looks
like
the
grid
was
essentially
representing
screen
space.
I
mean
you
know
it's
just
like
it's
a
it's
a
meaningless
there's,
no
depth
to
this
thing.
There's
no
structure
to
it.
It's
just
the
xy
position
on
the
screen.
C
Yeah
we're
in
the
images
before
one
could
imagine
that
it
might
be
representing
something
else
it
might
be.
You
know
you
might
be
thinking
about
the
the
where
you
are
in
this
scene
right,
but
so
I'm
really
confused
as
to
what
what
the
what
grid
cells
are
representing
here.
Is
it
just
screen
space?
Is
it
something
more
than
that?
I.
B
Mean
like
in
the
in
the
cover
task,
there's
a
very
even
exploration
of
this
2d
screen
with
the
use
of
these
with
the
use
of
these.
You
know
random,
but
especially
unbiased,
trajectories.
C
B
C
D
So
one
thing
with
the
covert
attention
one
is
they
call
it
covert
attention?
You
could
put
a
totally
different
narrative
on
this.
It's
just
like
where,
let's
see
you
have
this
image
coming
into
the
monkey's
visual
field,
it
has
a
2d
variable
of
where
is
the
moving
stimulus,
and
basically
it's
just
it's.
It
almost
ignores
space
here.
Think
almost
like
the
stretchy
bird
thing,
you
have
a
two-dimensional
variable,
that's
being
encoded
using
a
grid-like
code,
and
I
don't
know
the
word.
Attention
might
not
even
be
needed
to
explain
this
right.
F
F
C
Well,
my
assumption
you
didn't
say
this,
but
my
assumption
was
usually
in
these
kind
of
tasks
like
this.
They
make
it
so
that
the
animal
would
not
see
the
change
unless
they
were
attending
to
it.
That
is
the
change
is
small
enough
that
it
would
be
right
in
it'd,
be
very
difficult
to
see.
C
That
I've
done
these
brain
exercises,
where
this
is
what
they
do
they.
Basically
you
have
to
be.
You
have
to
attend
corporate
attention.
Otherwise
you
wouldn't
see
the
thing.
It's
too
subtle.
So
if
it's
a
big
color
change
here,
you
wouldn't
have
to
do
that
at
all,
but
I
just
assumed
that
there
were
many.
F
D
Or
even
if
attention
is
needed,
this
might
be
something
that's
far
downstream
from
that.
It's
it's
like
this
is
enter
angle
cortex,
it's
like.
Sometimes
it
represents
the
neck,
the
length
of
the
neck
of
a
bird,
and
in
this
case
it's
representing
a
screen
location
but
like
attention,
is
just
like
an
implementation
detail
that
led
to
internal
cortex
being
fed
this
information.
B
All
right,
I
mean
at
least
it's
interesting
that
well
first,
this
is
not.
This
is
not
navigation
and,
yes,
it
is
space,
but
it's
not
even
moving
in
the
space
right.
The
eyes
are
not
moving
and
that's.
I
think
what
makes
this
task
interesting.
C
It's
also
what
makes
it
very
very
hard
to
interpret
as
to
what
the
hell
is
going
on
here.
So
just
be
honest,
it's
it's!
You
know,
at
least
when
we
talk
about
iraq
moving
in
an
environment,
and
we
have
very
we
understand.
There's
a
very
lot
of
literature.
Explain
what's
going
on
here
here.
I
I'm
confused
as
to
what's
actually
being.
B
Where's
the
confusion
I
mean,
I,
I
think
this
is
pretty
straightforward.
It's
just.
C
Not
you
know,
I'm
trying
what
the
animal
is.
What
are
the
grit
cells
mapping
right?
What
is
the
what
what
this
is
the
grits
like
a
metric
space
they're
used
to
map
something,
but
what
is
it
is
it
the
screen?
Is
it
the
scene
is
the
thing
in
the
scene
and
the
covert
attention
part?
You
can
see
we're
all
kind
of
confused
by
that,
so
it
it
may
be
clear
with
the
authors,
but
it's
not
real
clear
to
me
what
you
know.
C
What's
what
is
being
it's
not
is
it
it's
is
it
where
the
dot
is,
and
why
would
that?
Why
does
that
require
covert
attention?
Why
would
could
it
be
done
if
they
had
animals
tracking
with
their
eyes?
Would
you
get
the
same
results.
B
Yeah
I
mean
that
was
the
point
right
of
the
of
the
previous
papers,
where
you
had.
The
monkeys
like
free,
be
free
being
able
to
the
card,
but
the
problem
was
that
it
you,
you
need
to
like
kind
of
disentangle,
then
the
the
movement
of
the
eyes
and
that's
tricky
right.
If
you
want
to
see
if,
if
the
screen
space
is
coded,
irrespective
of
the
movement,
then
you
would
do
exactly
the
the.
C
C
I
understand
that,
but
it's
not
I
couldn't
understand
you
know.
My
first
question
was:
is
it
is
it
independent
or
what
the
image
is
where
the
covert
attention
test
implies
that
it's
independent?
It's
a
screen
space.
So
maybe
that's
the
assumption
here.
They're,
just
mapping
screen
space
xy
locations.
B
Yeah,
that
is,
that
is
the
assumption
that
this
is
just
the
metric
for
the
space
and
there's
nothing
to
do,
in
particular
with
the
particulars
of
a
image,
and
it
has
nothing
to
do
with
the.
C
That's
not
a
way.
I
think
that
bristles
would
be
very
useful
when
I
look
out
at
the
world
and
my
eyes
are
moving
around.
I
wouldn't
expect
grid
cells
to
map
out
the
xy
coordinates
of
my
visual
scene.
I
mean
the
point
of
the
they.
As
we
know,
if
you
watch
a
video
game,
they
map
out
the
location
in
the
in
the
video
game,
not
the
screen
space.
C
Yeah,
but
that's
true
in
a
video
game,
it's
you
know
if
you
want
so
so
in
the
in
the
video
game
example
with
humans,
they
they
they've
shown
this
grin
cell,
like
responses
representing
where
the
ant,
where
the
person
is
mentally
moving
through
the
space
rep,
you
know
not
the
screen
space,
but
the
space
of
the
game.
C
B
C
Yeah,
okay,
so
when
I
that's
that
my
point
is
when
I
saw
the
first
image
is
I'm
assuming
that
the
grid
cells
would
be
representing
the
space
of
the
image?
When
you
see
the
beach
and
you
look
at
the
the
horse
is
further
away,
then
you
look
at
something
close
away,
just
like
a
video
game
yeah,
but
that's
not
what's
being
claimed
here.
It's
been
claimed
here,
you're
saying
it's
just
an
xy
screen
space.
So
it's
just
confusing
to
me.
B
B
C
Yeah
again,
but
that
would
be
that's
why
my
other
question
was
that
would
be
inconsistent
with
the
idea
that
grid
cells
remap
to
learned
environments
now
maybe
marcus's
point
was
that
maybe
these
aren't
learned.
But
if
these
were
learned
environments
then-
and
I
averaged
out
over
multiple
different
environments-
I
would
not
expect
to
see
any
bridge
cell
like.
B
I
mean
it
seems
that
you
have
questions
that
would
actually
like
require
the
you
know
like
a
proper
review
of
the
paper.
We
can
do
that
if
you
want.
E
So
either
I
mean
in
the
case
of
the
in
the
case
of
the
images,
it
could
be
something
like
what
you
said
that
just
based
on
the
statistics
of
the
kind
of
images
that
they
showed,
that
there
tended
to
be
areas
of
interest
there
that
just
showed
up
in
the
average.
The
other
thing
is
versus.
C
I
think
you're
right
harry
so
in
this
in
this
image
here
with
the
covert
attention.
Okay,
that
makes
sense
to
me
because
it
is
a
different
space
and
it's
it's
it's
it's
one.
You
can't
even
say
it's
learned
because
there's
no
features
in
it,
so
maybe
this
is
the
generic
space
of
screens
and
and
so
okay
great.
I
get
that
so
that
I
understand
it's
the
other
one
which
is
confusing
to
me
when
you're,
showing
horses.
E
C
B
Right
so
the
the
last
point.
B
I
just
wanted
to
to
to
point
out
that
in
in
the
task
where
there
are
the
cards
right
there,
you
can
apparently
find
cells
that
are
tuned
towards
the
direction
either
of
the
upcoming
task
or
to
the
just
happened,
sakat,
so
presa
card
and
poster
card
tuning,
as
you
can
see
here
in
gray
and
in
light
blue
and
some
of
those
tend
to
be
very
sharp
as
you
can
see,
and
they
cover
all
possible
angles,
as
you
can
see
from
this,
this
compass
here
on
the
right.
E
B
C
B
All
right,
I
would
like
to
move
on
to
something
that
left
me
quite
intrigued,
which
was
the
work
of
cliff
cantros
on
the
interaction
between
play
cells
and
grid
cells,
and
something
that
he
calls
artificial
remapping.
B
B
In
particular,
how
does
the
firing
of
grid
cells
influence,
placeholders
and,
of
course,
the
the
idea
that
play
cells
come
from
grid
cells
like
in
the
way
that
theoretical
models
of
lafitte,
for
example,
might
would
have
been
a
round
for
a
long
time?
It's
part
of
the
canonical
view,
but
there's
some
problems
with
it
like,
for
example,
in
novel
environments.
It's
the
place
cells
that
tend
to
stabilize
first
before
the
grid
cells,
stabilize
also
with
interventions
that
you
can
do
to
it.
You're
not
like
the
the
code
is
not
really
like.
B
B
You
can
do
all
kinds
of
interventions
to
mess
with
grid
cells
and
that
does
not
always
produce
changes
in
the
place
cells
like
so.
For
example,
if
you,
if
you
block
input
from
the
medial
septum
like
you're,
gonna,
obliterate,
theta
and
the
hexagonality
of
the
grid,
is
gonna
go
away,
but
most
of
the
place
cells
still
stay
place
cells
and
they
mostly
stay
in
the
same
place
too.
B
The
same
goes
for
different
kinds
of
pharmacological
or
optogenetic
interventions
that
people
have
tried,
lesion
studies
and
whatnot,
and
so
it's
not
always
like
so
clear
whether
like
what
role
do
the
grid
cell
inputs
play,
and
so
one
way
of
like
sort
of
doing
this
more
directly,
is
to
like
try
to
design
an
intervention
that
is
very
specific
to
meteor,
enter
rhino
cortex
layer,
2
stellate
cells,
which
is
of
course
a
sub-population
of
a
sub-population
of
a
sub-population
of
the
brain,
and
so
you
would
need
something
very
specific,
and
so,
in
order
to
get
that
kind
of
specificity
they
came
up
with
with
there
was
like
proceeding
work.
B
I
think
like
a
development
in
2015
or
16,
or
something
that
came
up
with
what
they
call
dread,
so
designer
receptors
exclusively
activated
by
designer
drugs,
so
genetic
toolkit,
which
allows
you
to
insert
specific
receptors
into
very
specific
cell
types
in
specific
areas
that
will
not
activate
anything
else,
because
they
only
activate
on
like
this
specific
designer
drug
which
doesn't
do
anything
in
control
animals
because
they
don't
have
receptors
for
that.
So
it's
a
designer
receptor
for
designer
drug.
B
So
that
gives
you
a
lot
of
control,
and
so
apparently
they
came
up
with
two
different
versions
that
would
either
excite
or
inhibit
stellar
cells
in
middle
anterior
and
the
cortex
layer.
Two.
So
you
can
see
this
here
very
nicely
how
how
these
map
out
exclusively
in
media
and
rhino
cortex
in
this
standing
here
and
that
they're
pretty
much
exclusively
layer.
Two.
They
don't
really
target
layer
three
and
that
they
even
leave
out
like
these
these
islands,
these
carbondine
positive
islands.
B
So
there's
like
these
are
like
the
the
reelin
ones
right,
which
are
like
the
the
majority
of
it
and
it's
the
stellate
cells,
which
are
the
major
like
the
vast
majority
of
the
of
the
grid
cells,
and
so
you
now
have
a
very
specific
target
and
with
a
designer
drug
you
can
now
temporarily
make
these
cells
more
excited
or
more
inhibited,
so
depolarizing
hyper
polarization,
and
with
that
amount
of
control,
now
you
can
take
a
look
okay.
B
What
how
are
the
place
cells
going
to
respond
to
this
and
what
they
found
is
that
they
get
they
get
very
interesting
remapping,
so
they
they
take
these
mice
out
of
their
their
environments,
apply
this
transient,
depolarization
or
hyperpolarization,
which
now
makes
the
middle
and
toronto
cortex
layer
tools,
delayed
cells,
transiently,
more
excitable
or
less
excitable,
and
when
you
are
hyperpolarized,
nothing
happens.
B
The
place
cells
that
you
record
stay
the
play
cells,
the
firing
obviously
of
the
grid
cells
goes
down,
but
when,
instead,
you
depolarize,
you
get
a
complete
remapping
of
the
place
cells,
so
you
can
see
some
of
this
here.
So
this
is
what
he
calls
artificial
remapping
of
ca1
placeholds.
Following
depolarization,
so
cno
is
this
drug
that
they
that
they
apply
to
to
do
this
remapping
and,
as
you
can
see,
the
the
firing
fields,
change
they're,
really
like
not
the
same
anymore
in
very
rare
cases
right
they
do
stay
in
the
same
spot.
B
So
and
when
you
look
at
what's
happening
in
the
middle
and
trying
to
cortex,
where
that
you
know
drug
is
applied
where
that
drug
acts,
you
can
see
these
different
effects,
so
they
have.
This
hm3
type.
Is
the
type
that
depolarizes
upon
this
cno
drug.
So
you
see
that
the
firing
rate
goes
up.
The
firing
fields
stay
in
the
same
place
right
from
here
to
here
right,
there's,
some
like
in
the
different
types.
You
know
the
the
one
that
hyperpolarizes
and
the
one
that
depolarize
there's
different
reactions.
B
B
This
of
course,
flies
a
little
bit
in
the
face
of
the
of
the
canonical
model,
which
puts
a
lot
of
importance
onto
this
idea
of
that,
the
that
the
places
are
composed
of
the
fields
of
the
grid
cells.
So
it
seems,
like
the
absolute
rate
and
the
rate
relationships
between
griff
and
grid
cells
are
actually
way
more
important
for
the
coding.
B
This
artificial
remapping
that
they
see
typically
holds
as
long
as
the
mice
then
like
stay
in
that
environment,
and
you
can
transiently
take
them
out
and
put
them
back,
they're
still
going
to
remain
in
that
new
map,
so
the
place
cells
like
changing
in
that
in
that
place,
unless
you
take
them
out
for
quite
a
long
time.
So
it's
not
just
the
the
the
odometry
so
to
speak
right,
the
tracking
of
head
direction
and
whatnot.
That,
like
is,
is
part
of
that.
So
there's
some
actual.
B
Like
change
of
the
map
and
that
takes
a
while
to
stabilize
this-
is
like
something
like
15
minutes
and
returns
back
to
baseline
when
you
remove
them
from
the
environment
for
a
longer
time,
which
is
which
is
quite
interesting,
and
that
is
like
that
is
not
just
reversible
but
also
repeatable.
So
the
kinds
of
cells
that
would
remap
also
remap
a
second
time
when
you
do
that,
and
they
that
also
you
know,
returns
to
baseline.
B
A
second
time
which,
which
I
found
quite
you
know,
puzzling,
and
I
think,
there's
something
to
be
learned
here.
B
The
the
fact
that
it
tends
to
be
the
same
cells
that
react
that
remap
and
that
they
remap
not
just
to
any
space
but
to
like
a
rather
particular
space,
could
be
seen
as
a
sign
that
there's
a
hidden,
attractor
landscape
of
sorts,
so
that
so
it's
kind
of
predetermined
by
the
recurrent
connectivity.
I
guess
where
these
cells
can
go,
which
is
in
some
sense.
B
G
Florin,
how
did
they
come
to
the
conclusion
that
the
natural
remapping
was
instantaneous.
B
Oh,
that
is
well
known
like
when
you
have
maps
in
two
different
environments,
they're
familiar
with
both
put
them
from
one
into
the
next
there's
immediately.
The
new
map
is
fully
there.
It
needs
no
time
to
stabilize
it's
immediately
fully
active.
B
So
all
the
hexagons
you
know,
hexagonal
patterns
are
full
and
the
and
the
placers
are
all
in
the
place
where
they
are,
and
they
don't
change,
as
opposed
to
this,
where
it
like
takes
15
minutes
for
this
new
map
to
to
to
form.
So
that's
a
difference
between
like
the
the
natural
remapping
and
the
and
this
artificial
remapping.
B
So
right,
okay,
so
I
guess
no
question
so
for
I
I
found
this
very,
very,
very
puzzling,
because
it
kind
of
needs
different
models
to
to
explain
this.
The
big
takeaway
for
me
was
like
we.
We
need
to
pay
attention
to
the
information
content
that
is
in
the
differential
or
grid
zone
firing
rates,
and
so
firing
fields
are
not
binary
on
off.
D
So
I
guess
I
I'll
point
out
that,
like
this,
this
does
align
with
other
things.
We've
seen
the
one
classic
thing
we
bring
up.
All
the
time
is
the
is
the
david
tank
paper
with
the
missed
fields,
that's
kind
of
similar
to
like
rates
being
different
in
different
firing
fields,
there's
also
something
from
kate,
jeffrey
dori.
D
Der
dickman
involving
the
paper,
is
called
something
like
grid
cells,
encode
local
position
info
and
that's
just
showing
that
consistently
the
firing
rate
and
different
firing
fields
of
a
good
cell
is
different,
such
that
you
can
actually
get
more
information
about
where
the
animal
where
the
ra
is
than
just
yeah,
then
just
which
cells
firing.
You
can
also
use
the
right
so
so
there
this
does
align
with
other
things.
We've
seen
right.
B
C
B
Yeah,
it's
all
good
so
so
to
just
be
be
clear
right,
so
the
cells
that
remap
are
most
often
cells
with
the
second
firing
field
and
and
and
so
which
again
like
supports
this
idea
that
there's
some
hidden
attractor
landscape,
which
is
formed
by
recurrent
connectivity.
C
Do
me
a
favor:
this
is
a
complex
paper
and
a
lot
of
biology
and
experimental
technique
involved.
If
you
think
it's
really
interesting,
I'd
like
to
read
it,
so
maybe
I
can
just
get
it
right
off
the
screen
here
is
this:
is
this
a
publish
paper?
Is
that
the
name
of
the
paper.
B
All
right
cool
next,
what
we
might
want
to
do.
I
mean
like
in
this
context.
It
may
be
interesting
to
take
a
look
at
new
map
permission.
I
don't
know
how
we're
doing
on
time.
B
So
there's
also
a
presentation
by
by
julia
krupitch,
who
is
talking
about
the
the
the
real-time
formation
of
maps
because
of
course,
like
most
of
the
time,
we're
talking
about
like
these
stable
maps
all
the
time,
but
whenever
we
put
a
rodent
into
a
place
like
these
maps
are
forming,
and
can
we
track
that
formation
in
real
time?
If
we
really
think
of
that
whole
hippocampal,
android
or
cortex
system
of
doing
like
simultaneous
localization
and
mapping
with
the
robot
assist,
call
it
right,
then
we
should.
B
B
So
you
give
them
like
segments
of
of
of
learning
time.
So
when
they
go
into
environment
for
like
three
hours
and
then
you
can
break
that
apart
into
like
20
minute
segments
to
to
track
sort
of.
How
does
the
map
change
over
over
time?
B
Cool
thanks,
so
I
guess
should
be
enough
to
go
through
this,
at
least
so.
What
I
mentioned
before
right,
as
as
cliff
centros
mentioned,
as
one
of
the
things
that
is
sort
of
a
little
bit
critical
of
this
idea,
is
that
somehow
the
playsets
come
from
with
cells
is
that
the
play
cells
are
typically
present
much
earlier,
but
stabilize
rather
slowly,
so
they
still
increase
their
firing
rate
and
they
tend
to
move
around
a
little
bit
until
they
sort
of
are
always
reliable
in
the
same
fixed
spot.
Right,
keep
in
mind.
B
This
is
so
over
the
course
of
like
three
hours
right,
I
don't
know
the
exact.
I
don't
remember
the
the
you
know
exact
time
course
of
this,
but
the
the
point
was
to
say
you
you
get
to
the
stable
position
and
stable
rate
quite
slowly,
but
they're
there
earliest
and
then
so
with
with
the
grid
cells.
That
tends
to
be
a
bit
different
in
that
there
tends
to
be
like
one
of
the
fields
you
know
of
like.
Let's
say
that
grid
cell
has
like
four
or
five
firing
fields
in
this
arena.
B
It
tends
to
be
one
that
is
first
and
then
it
tends
to
fill
on
like
the
hexagon
around
it
over
time.
So
there's
like
one
of
the
other
that
kind
of
like
pops
up
at
the
at
the
right
distance.
So
there's
a
there's,
a
filling
in
of
sorts,
so
so
that
the
change
of
that
of
those
fields
is
also
like
slowly,
let's
see,
yeah,
and
so
the
grid
cells
tend
to
change,
monotonically,
meaning
you.
B
You
have
like
one
field
and
then
you
fill
in
the
other
ones,
whereas
the
place
cells
sometimes
jump
before
they
settle
into
their
final
point
and
then
build
out
the
right
firing
rate.
B
B
I
don't
remember
that,
but
I'm
I'm
sure
she
said
something
like
that.
C
I
mean
the
reason
I
ask
you
is
that
you
know
I
have
a
model
in
my
head
that
we've
discussed
a
lot
about
how
good
cells
and
place
cells
are
interacting
and,
and
I'm
always
testing
these
papers
against
that
model
to
see
if
that
model
is
wrong,
and
so
some
of
these
things,
I'm
I
was
just
imagining
here,
like
imagine
yourself-
being
dropped
into
some
big
sort
of
box-
that
there's
no
real
landmarks
to
to
judge
where
you
are.
C
Your
grid
cells
are
going
to
be
very
inaccurate,
initially
you're
going
to
be
relying
almost
completely
on
path,
integration,
and-
and
so
you
just
can't
expect.
C
Right,
there's
no
landmarks
to
really
you
know.
It's
a
you
know.
Grid
says
they're
not
going
to
work
they're,
just
not
going
to
do
their
job
in
a
situation
where
you
can't
concrete
landmarks
to
know
where
you
are
and
then
place
fields,
I
always
think
of
them,
as
is
primarily
being
sensory,
driven
and
so
they're
going
to
be
there
from
the
start.
There's
going
to
be
some
century
landmarks,
something
otherwise
you
wouldn't
have
any
place
fields.
C
I
put
you
in
a
white
field
of
no
nothing.
You
wouldn't
have
any
places.
So
so,
as
I'm
just
I'm
just
pointing
out
that
when
I,
when
I
look
at
these
kind
of
points
that
you're
making
here,
I'm
just
testing
our
model
of
what
we
think
is
going
on
and
and
if
I
don't
find
a
contrary
evidence,
then
I
say:
okay,
it's
fine,
but
these
seem
consistent
with
what
my
model
in
my
head
about
how
place
cells
and
grid
cells
interact
and
how
they
form
how
an
animal
learns
an
environment.
G
Lauren
did
was
there
any
hypothesis
as
to
what
the
mechanism
was
for
the
place
cells
jumping
suddenly.
B
Yeah,
well,
I
mean
like
I
could
answer
this
with
a
pointing
to
the
talk
by
cliff
cantos
right.
The
fact
that
the
when
the
when
the
grid
cells,
like
start
filling
in
and
like
they
typically
have
an
anchor
somewhere
right,
there
will
be
changes
in
the
rate
relationships
right
between
the
grid
cells,
and
it
is
as
clear
control
showed,
is
exactly
these
changes.
These
differences
in
the
rate
relationships
of
the
grid
cells.
B
Even
if
there's
no
big
shift
in
space
right,
it's
only
a
quantitative
change
that
leads
to
a
qualitative
change
in
the
in
the
place
cell
map
right.
That
was.
That
was
exactly
the
the
point
that
cliff
centros
was
was
making
that
from
quantitative
changes
in
grid
cell
firing.
You
get
qualitative
tension
in
ci1.
B
That
is
obviously
only
like
a
correlation
that
doesn't
answer
your
question
as
opposed
to
mechanistically
what
happens
there,
but
I
guess
the
the
the
answer
is
that
you
would
hypothesize
that
there
is
some
hidden
attractor
landscape
and
somehow,
by
having
different
stimulating
rates,
you
fall
into
different
attractors,
which
is
why
the
placer
then
flips
to
a
different
possible
field
and
either
stabilizes
this
one
or
the
other
one.
But
it's
not
going
to
be
free
to
choose
any
point
in
space.
G
Yeah,
I
was,
I
was
kind
of
the
combination
of
those
things
I
was
in
in
in
my
universe.
I
would
call
it
mode
locking
or
a
number
of
other
things
where
all
of
a
sudden
there's
a
phase
relationship
that
reinforces
so
that
was
why
I
was
just
wondering
if
she
had
speculated
whether
this
is
part
of
the
attractor
model
or
part
of
the
you
know
some
of
the
other
models
we've
looked
at.
B
Like
people
have
been,
you
know
thinking
about
how
this
possibly
what
would
happen.
I've
been
piecing
this
together,
but
I
I
think
this
was
the
very
first,
like
you
know,
conclusive
demonstration
with
like
these
high
density
probes
of
following
this
whole
map
formation
process,
while
it
was
going
on
so
even
though
none
of
this,
like
really
contradicts
our
our
current
understanding,
I
think
it
was
very
important
to
to
to
demonstrate
that
rather
clearly,
and
it
seems
like
that
they
have
done
that.
C
You
can
imagine
easily
how
just
again
imagine
yourself
in
these
environments
and
how
it
would
be
easily
difficult
to
know
where
you
are
and
the
more
time
you
spent
in
sort
of
a
pretty
much
of
an
environment,
a
door
that
looks
the
same
everywhere
and
it's
devoid
of
almost
any
interesting
things
in
it.
That
you
would
your
perception
of
where
you
would
initially
would
be
confused
a
lot
and
and
over
time,
as
you
spend
more
and
more
time
in
the
environment.
C
You'd
pick
up
on
subtle
clues
and
you
would
get
better
and
better
at
knowing
where
you
are
just
it's
just
sort
of
naturally
occurred
so
to
me
that
that
you
know
given
such
these
are
such
impoverished
environments.
B
Although
it's
interesting
also
that,
like
the
the
grids,
really
become
like
you
know,
canonical
after
after
a
longer
time
right,
so
they
say
that
you
know,
like
you,
get
sort
of
the
full
convergence
of
like
the
stability
of
the
rates
at
the
different
fields
that
are
mapped
out
and
the
full
hexagonality
of
it.
You
you
get
after
a
week,
which
kind
of
suggests
that
there
might
be
some
offline
processes
that
also
play
a
role
here
like
whether
I'm
I'm
not
sure
about
that.
B
But
I
would
hypothesize
right
that
there's
like
offline
processing
that
also
helps
with
like
making
these
grids
very
nice
like
what
you
would
get
doing.
You
know
from
from
from
replay,
if,
indeed
the
grid
cells
are.
You
know
tagging
along
during
that
replay,
then
that
also
is
another
opportunity
to
refine
the
the
the
the
grid.
C
Replay
may
make
an
important
part
of
it,
but
I
mean
it's
like
any
task.
You
learn
right
if
you're
learning
how
to
play
a
new
piece
on
the
piano
or
something
it
takes
days,
you
get
continually
better
at
it,
so
that,
as
a
general
function
of
learning
in
the
brain,
that's
just
not
unexpected,
but
you
know
it's
an
interesting
question.
Whether
replay
is
an
important
part
of
it.
Probably
is.
B
Right
and
what
we
might
get
like
from
from
the
fact
that
the
the
grid
surfing
tends
to
like
have
sort
of
an
early
anchor
from
which
it
starts
filling
out
the
neighboring
fields,
and
that
seems
to
be
there,
irrespective
of
the
fact
that
you
don't
have
me
like
an
environmental
anchor
yet
because
the
environmental
anchors
right,
your
your
place,
cells
they're,
like
jumping
around,
so
you
kind
of
that
that
goes
well
with
the
with
the
attractor
idea
that
somehow
there
is
already
some
some
fixed
point
right,
a
bump,
and
that
you
can.
B
Then
you
know,
use
that
because
you
have
surrounding
inhibition,
and
that
opens
up
the
possible
fields
outside
of
that
yeah
anyways.
I
think
we
should.
You
know,
probably
concluded
that
this
unless
there's
you
know
like
more
specific
questions,.
C
No,
I
just
want
to.
I
want
to
read
that
paper
that
you
you
mentioned.
C
C
B
C
We'll
be
back
in
45
minutes
or
something
like
that.
B
C
A
Okay,
thanks
for
watching
my
name
is
matt
taylor.
I
work
at
dementa
and
during.
I
just
wanted
to
note
that
we're
going
to
continue
trying
to
live
stream,
our
research
meetings,
while
everybody's
sort
of
down
working
from
home
sheltering
in
place
whatever
in.
A
At
it's
a
big
deal,
there's
a
whole.
Everything
is
shutting
down
and
we're
all
working
from
home.
So
that's
why
you
see
the
the
zoom
meetings
instead
of
seeing
this
meeting
being
in
the
office,
but
we're
I'm
still.
I
have
the
capabilities
to
continue
to
live
stream.
All
this
stuff,
I've
already
set
it
up.
A
I've
been
doing
it
for
almost
a
year
now,
so
we're
going
to
continue
to
share
and
have
these
research
meetings
and
and
probably
more
stuff
over
the
next
month
since
everybody's
going
to
be
working
from
home,
and
I
could
do
it
and
let
me
maybe
we'll
have
a
couple
virtual
talks.
I
know
we're
going
to
have
a
big
htm
hackers
hangout
april
8th.
If
anybody
wants
to
join
that
go
to
our
forum,
I'll
put
in
it's
a
discourse.numenta.org.
A
I
just
threw
that
into
chat.
I
hope
I
spelled
it
right
or
just
search
for
htm
forum
and
you
can
find
out
there
about
the
htm
hackers
hangout,
I'm
gonna
be
have
a
big
big
zoom
meeting,
get
anybody
as
many
people
as
possible.
Apparently
we
can
fit
a
hundred
people
in
a
zoom
meeting
under
our
plan,
so
we're
going
to
do
that
april.
8Th
4
p.m,
pacific
daylight
savings
time.
C
A
So
I
am
looking
forward
to
seeing
all
of
you
guys
there
all
your
familiar
faces
and
chat
over
there.
I
appreciate
you
guys
watching.
Please
spread
the
news
that
we're
still
streaming
and
share
these
videos
with
your
friends
like
this
video.
Please
give
it
a
thumbs
up.
It's
very
helpful
and
subscribe
to
our
channel,
as
you
saw
somebody
earlier
subscribed
and
that
brain
pops
up
so
that's
fun
so
signing
off.
I
will
see
you
guys
with
another
stream
on
monday.