►
From YouTube: Paper Reviews on Multiscale Representation and Representational Drift in the Neocortex -Jul 12, 2021
Description
We reviewed 2 papers in this research meeting. First, Numenta intern Jack Schenkman reviewed the paper “Multiscale representation of very large environments in the hippocampus of flying bats” by Eliav et al. The paper proposes a multiscale neuronal encoding scheme of place cells for spatial perception. The team then raised a few questions and discussed.
Next, our researcher Ben Cohen reviewed the paper “Representational drift in primary olfactory cortex” by Schoonover et al. The paper shows that single neuron firing rate responses to odor in the anterior piriform core are stable within a day, but continuously drift overtime. The team then discussed the notion of representational drift in the context of Numenta’s work.
“Multiscale representation of very large environments in the hippocampus of flying bats” by Eliav et al.: https://science.sciencemag.org/content/372/6545/eabg4020
“Representational drift in primary olfactory cortex” by Schoonover et al.: https://www.nature.com/articles/s41586-021-03628-7
Columns paper mentioned: https://www.frontiersin.org/articles/10.3389/fncir.2017.00081/full
- - - - -
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/
B
A
And
I
should
say
if
you
have
any
questions,
please
feel
free
to
interrupt
me
as
I'm
going
so
just
for
a
quick
overview
of
place
cells.
Basically,
they
work
to
represent
an
animal's
position
in
space,
there's
many
that
are
found
in
the
hippocampus
and
basically
they
tend
to
increase
in
firing
rate
when
an
animal
passes
over
a
specific
region
called
the
place
field
and
there's
a
large
number
of
them
found
in
the
hippocampus.
A
Of
course,
there's
many
more
details
than
what
I'm
sharing
here,
but
this
is
just
a
brief
overview
of
how
they
work.
So,
basically,
most
studies
of
place
cells
have
been
focused
on
relatively
small
environments
like
one
to
two
meters
in
size
and
there's
been
little
research
on
more
realistic
environments,
meaning
the
kilometers
that
animals
tend
to
traverse
as
they
go
about
their
daily
activities.
There's
been
a
few
studies
in
larger
environments,
but
these
environments
often
have
zigzags
or
can
be
compartmentalized
into
smaller
environments,
so
they're
effectively
just
a
conglomerate
of
many
small
environments.
A
And
so
the
this
research,
which
is
focused
on
small
environments,
is
in
pretty
stark
contrast
to
the
actual
behavior
of
animals
just
to
provide
some
examples.
Rats
can
navigate
more
than
a
kilometer
per
night
and
egyptian
fruit
bats
will
move
up
to
30
kilometers
per
night,
and
so
just
to
kind
of
motivate.
Why
exploring
multi-scale
representations
is
of
interest?
If
we
imagine
a
10
to
20
centimeter
field
size,
then
to
represent
the
distances
traversed
by
the
egyptian
fruit
bat,
you
would
need
about
10
to
the
13th
neurons.
A
The
way
that
they
arrive
at
this
number
in
the
paper
is
basically
they
say.
The
fruit
bat
is
flying
30
kilometers
and
it
has
a
width
that
it
flies
within
of
about
two
kilometers
and
a
height
a
variation
of
about
0.5
kilometers.
So
then,
if
you
take
that
volume
and
see
how
many
of
these
10
to
20
centimeter
diameter
spheres
would
fit
inside,
you
arrive
at
this
number.
A
B
B
B
C
A
Right
so
they
mentioned
a
little
bit
about
like
a
place
cell
firing
for
multiple
locations,
but
for
the
most
part,
the
assumption
is
that
a
place
cell
fires
for
like
a
single
location
or
a
single
neighborhood
and,
what's
really
being
varied,
is
how
sort
of
how
wide
the
neighborhood
is.
So
I
mean,
I
should
say,
like
a
given
place
cell
in
this
example,
may
fire
for
multiple
locations,
but
the
focus
of
the
paper
was
really
on
how
the
like
the
field,
the
size
of
each
location
varies.
A
C
C
E
A
E
C
B
C
We
do
know
that
that
place
cells
fire
in
multiple
locations
yeah,
especially
when
you
get
to
a
larger
environment,
so
I
think
the
default
assumption
all
along
was
hey
yeah
play
cell,
it
fires
in
one
location,
but
it'll,
actually
fire
in
many
locations.
If
the
if
the
room
gets
big
enough,
I
think
that's
been
the
default
assumption
so,
but
obviously
they're
going
to
show
something
else.
That's
still
going
to
be
true,
but
there.
C
E
G
F
E
F
C
D
And
if
you
have
a
place
cell,
responding
at
two
places
in
one
environment
put
the
animal
in
a
different
environment,
that
cell
might
no
longer
fire
at
two
places
and
that
and
that
new
one
like
there's,
no
there's
the
relative
firing
locations
are
not
systematic
in
any
way.
It's
not
just
that.
It's
not
periodic
they're,
not
systematic.
A
Awesome
so
just
for
a
quick
overview
of
the
setup
of
the
study,
so
they
were
using
egyptian
fruit,
bats
and
there's
a
sufficient
amount
of
light
to
be
able
to
basically
recognize
potential
landmarks
within
this
tunnel.
That's
about
200
meters,
long
and
the
tunnel
actually
has.
It
actually
has
a
curve
of
about
43
degrees.
So
it's
like
one
piece
is
about
140
meters
and
the
other
piece
is
roughly
50
meters
and
then
there
they
meet
at
this
43
degree
curve,
basically
and
they're.
A
A
A
C
C
Well,
you
know
place
fields
if
I,
if
I
recall
they
in
you
know,
if
you
put
a
rat
on
a
on
a
linear
track,
then
the
forward
and
backwards
directions
have
different
different
coatings.
But
if
they're,
if
it's
in,
if
it's
just
in
a
room,
then
then
it's
the
same
cell
virus,
regardless
of
which
direction
you're
going.
A
C
But
I'm
talking
about
sideways,
I'm
just
saying
that
that
on
a
linear
track
a
rat,
will
a
cell
will
encode
at
one
location,
going
one
direction
and
it
will
not
fire
at
that
same
location
going
the
other
direction,
but
if
it
was
in
a
two-dimensional
space,
the
rat
is
traveling
through
that
spot
and
backwards
with
that
spot.
The
same
cell
fires
in
both
cases.
C
So
it's
it's
almost
it's
I've
always
interpreted
that
as
when
you're
on
this
linear
track
and
the
animal
perceives
it
like
almost
like
a
it's.
It's
not
just
I'm
in
a
room,
I'm
on
this
track
and
it's
sort
of
a
different
encoding
going
one
way
coming
back,
but
if
you're
in
a
room,
it's
the
same
encoder,
no
matter
which
way
you're
where
you
are
so
this.
It's
not!
It's
not
like
you're,
going
orthogonal
to
the
track.
Yeah.
A
A
So
basically
what
they
were
able
to
conclude
from
this
is
that
there's
a
very
wide
variety
in
the
size
of
the
place
fields.
So
a
single
cell
would
have
up
to
20
times
differences
in
the
size
of
its
place
field
so
going
from
about.
You
know
well
under
a
meter
to
like
roughly
15
to
20
meters,
so
it's
quite
a
large
range
in
the
size
of
the
place
fields
for
a
given
cell.
A
A
B
C
But
I
don't
stand
because
if
you
didn't
change
the
landmarks-
oh
okay,
all
right
so,
but
a
cell
is
going
to
respond
to
the
same
location
and
regardless
right
it's
and
they
didn't
it's
and
they
all
right.
So
maybe
maybe
is
this
correct
to
say,
like
hey,
you've
got
these
landmarks
and
as
this
and
the
size
of
the
or
the
peak
and
the
where
the
cell
fired
and
the
size
of
a
trip
that
was
its
field
wasn't
correlated
to
how
far
away
is
from
a
landmark.
Is
that
what
you're
saying.
C
B
But
I
guess
that's
what
you
were
saying:
jeff,
that
it's
they're
looking
at
the
distance
of
the
peak
to
each
landmark
and
yeah
and
then
seeing
whether
that
there's
any
correlation.
And
if
you
then,
if
you
shuffle
the
locations
of
the
landmarks,
virtually
it
doesn't
make
any
difference
to
the
average
distance
to
each
peak.
B
A
Yes,
so
then
they
also
wanted
to
assess
how
the
behavior
was
different
in
large
and
small
environments.
So
in
these
small
environments
they
noted
that
a
much
smaller
percentage
of
the
neurons
that
they
measured
were
were
active.
Basically,
one
thing
that
I
was
as
I
was
reading.
The
paper
I
was
struggling
to
understand
was
why
there
was
such
a
large
difference
in
the
number
of
neurons
that
were
recorded
in
the
original
large
environment
and
this
smaller
six
meter
environment.
A
C
C
Is
this
for
the
same
cells
like?
Is
it
like
it's
interesting,
because
you
know,
I
don't
think
people
have
reported
these
changes
in
size
of
the
receptive
fields
before
at
least
nothing
of
this
scale,
and
so
the
question
is
like
a
single
place
cell
in
small
environments
always
represents
sort
of
a
similar
size
place.
But
when
you
put
in
a
big
environment
it
it,
it
might
start
representing
big
places
is
that
is
that
right.
E
C
C
B
B
A
A
This
may
just
be
because
there's
the
size
of
the
place
field
will
only
get
so
small
and
it
can
only
be
so
big
in
the
smaller
environment,
but
in
the
larger
environment.
This
maximum
size
will
be
larger.
So,
like
one
one
question,
I
guess
for
me
that
emerges
is:
is
this
increase
due
to
the
maximum
place
field
size
simply
increasing
in
the
larger
environment?
Or
is
there
also
a
counteracting
effect
of
the
minimum
place
field
size
decreasing
in
the
smaller
environment,
to
represent
finer
tuned
areas?.
B
B
C
A
A
So
one
other
point
that
there
is
in
the
paper
is
like-
and
this
is
something
I
guess
we
kind
of
touched
on
as
we're
discussing,
is
the
multi-uh
scale
coding
something
that's
learned
or
something
that's
innate
to
the
bats.
So
all
the
data
that
I've
presented
so
far
was
from
bats
that
were
captured
in
the
wild,
so
so
these
bats
are
used
to
existing
in
very
large
environments.
A
So
basically,
this
is
going
to
this
point.
So
what
they
found
is
when
they
take
these
laboratory-borne
bats
and
place
them
in
the
200-meter
long
tunnel
and
let
them
train
for
a
few
weeks.
Then
they
basically
show
the
same
multi-skill
properties
as
the
wild
born
bats.
Although
the
wild
born
bats
did
have
slightly
larger
place
fields,
so.
A
Well,
so
that
was
my
question
because
in
the
bullet
here,
basically,
they
tested
whether
this
point
was
using
bats
from
the
wild,
so
bats
that
are
used
to
large
environments
and
basically
they're
saying
bats
that
are
used
to
large
environments.
Have
this
this
multi-skill
encoding
from
day
one
onwards,
although
there
is
more
variability
on
day
one
and
the
variability
kind
of
decreases,
but
it's
possible
that
in
the
laboratory
bats
that
it's
during
these
first
few
weeks
when
they're
training
that
they're
adapting
to
this
multi-skill
coating.
A
E
You
know
that's
an
important
data
point
to
know.
You
know
because
it's
it's
it's
a
con.
You
know
it's
just
it's.
I
don't
think
unless
marcus
knows
of
anything
with
super
tired
or
anyone
else,
I
don't
know.
I
think
this
is
a
very
surprising
result.
So
it's
trying
to
understand
how
it
could
have
come
about.
G
Well,
it
could
also
come
about
because,
when
they're,
when
they're
no
longer
a
confined
environment,
they're
spending
some
time
I
mean
they
said
they
have
to
be
trained-
they're,
probably
spending
some
time
worrying
to
traveling
larger
distances
and
then
associating
that
with
finding
food,
whereas
the
food
has
always
been
kind
of
hand
delivered
to
them.
So
you
know
it's
it's
a
confounding
factor
that
you
know
that
they
couldn't
get
the
experiment
set
up
right.
C
E
C
E
B
I
think
any
again,
if
you
think
about
it
from
a
coding
standpoint
like
us,
think
about
like
a
spatial
cooler,
that's
trying
to
represent
its
inputs.
If
you're,
given
a
very
restricted
set
of
inputs,
it's
going
to
have
very
restrict
you
know,
restricted
responses,
and
if
you
take
the
exact
same
spatial
cooler
and
train
it
on
a
much
larger
set
of
inputs,
it's
going
to
get
much
more
diverse
and
yeah.
B
C
G
C
E
C
Of
space-
and
so
this
research
breaks,
two
assumptions
I
would
have
had
is
that
place
fields
would
have
had
similar
receptor
field
sizes.
They
could
have
scaled
up
and
down,
like
you,
said,
super
tight,
but
I
would
expect
them
to
have
similar
separate
field
sizes.
It
also
violates
something
else.
I've
argued
for
in
the
past,
which
I've
argued
for
that
when
we're
trying
to
to
read
out
a
location.
E
H
I
think
one
of
my
main
take
of
this
paper
is
besides.
The
computational
aspect
is
from
the
biological
aspect.
How
much
of
the
behavior
and
computation
that
we
observe
is
actually
born
from
mice
and
animals
being
in
artificial
environments
and
then
how
much
of
that
of
those
cognitive
and
and
computational
computation
that
we
observe
and
we
want
to
interpret-
are
actually
artificially
born
from
from
them
not
being
into
their
ecological
environment
yeah.
So.
B
B
H
C
I
mean
this
is
this:
is
a
general
problem
of
interpreting
any
kind
of
experimental
data,
because
you've
got
the
age
of
the
animal.
You
know
with
the
preparation.
You
know
in
general,
we're
talking
about
all
different
types
of
empirical
experiments,
and
sometimes,
if
it's,
you
know,
you
know,
the
different
preps
are
different
techniques.
They
all
have
their
biases
right.
They
all
end
up
with
different
results.
So
now
we
just
have
to
add
another
one.
Here
we
have
to
think
about
all
these
good.
C
H
E
A
B
H
B
A
All
right
and
to
then
to
some
of
the
computation
type
of
stuff,
so
they
just
basically
consider
the
impact
of
having
single
scale
versus
multi-scale
coding,
so
they
present
a
few
different
schemes
in
this
one.
Basically,
each
neuron
has
a
single
small
place
field.
This
one.
Each
neuron
is
a
single
large
place
field.
This
one
each
neuron
has
a
single
place
field,
but
the
place
fields
are
of
different
sizes
between
neurons
and
increase.
A
A
The
place
fields
for
a
given
neuron
are
all
of
the
same
sizes,
but
different
neurons
have
different
place
field
sizes
and
then
in
scheme
six,
it's
a
hodgepodge
of
everything.
I
guess
each
neuron
has
multiple
place
fields
all
of
different
sizes
so
and
basically
scheme.
Six
is
what
aligns
most
with
the
empirical
observations.
A
A
So
they
measure
performance
basically
by
in
a
few
different
ways.
One
is
how
many
neurons
do
you
need
to
accurately
keep
track
of
the
the
position
in
this
simulation
and
they
define
accuracy
as
being
less
than
two
meters
from
the
actual
position
and
then
also
just
kind
of
the
the
average
of
the
error
in
terms
of
from
the
ground.
Truth
position
as
and
both
of
these
are
as
a
function
of
the
size
of
the
environment.
A
A
That
is
that
is
seen,
or
is
it
that
the
neurons
are
actually
switching
when
they
get
into
the
larger
environment,
to
use
multi-skill
encoding,
and
the
authors
seem
to
prefer
to
seem
to
prefer
the
second
approach
that
the
neurons
are
switching
from
single
scale
encodings
to
multiscale,
because
they
argue
if
it
were
multiscale
encodings.
Then
in
the
small
environments,
you'd
see
some
neurons
that
are
just
always
on,
because
their
place
field
would
encompass
the
entire
length
of
the
small
environment.
A
E
A
A
B
C
D
I
mean
it's
a
I.
I
would
say
that
I've
never
had
like
a
compelling
theory
of
place
cells.
I
mean
one
answer
would
be.
I
mean
right
now,
I'm
looking
at
them
through
the
lens
of
being
graph
nodes,
and,
and
so
it
could
be
the
case
that,
as
a
as
a
bat
is
flying
through
like
a
long
tunnel,
it
has
these
memory
graph
nodes
associated
with
different
parts
of
it
and
the
way
that
those
nodes
are
allocated
as
what
we
would
call
an
sdr.
D
You
just
have
some
random
code,
a
random
code,
and
sometimes
these
are
going
to
be
associated
with
like
far
off
landmarks.
That
can
be
seen
for
a
while
other
times,
they're
going
to
be
associated
with
like
quick
things
that
are
seen
in
passing,
and
so
one
point
of
view
would
be
that
you
just
have
this
general,
this
general
neural
mechanism
for
representing
nodes,
and
you
associate
things
with
them,
and
sometimes
you
associate
things
that
you
see
for
long
skills,
and
sometimes
you
associate
things
you
see
for
small
skills.
D
C
D
C
C
Because
that
would
that
would
be
the
thing
that
would
really
mess
me
up
right.
You
know,
because
we,
you
know,
there's
all
these
proposed
mechanisms
for
path,
integration
and
and
for
the
procession
of
activities
and
all
this
stuff,
and
that,
I
think,
would
have
to
make
me
recent
everything
I
thought
about
grid
cells,
at
least
a
good
portion
of
it
if
they
were
the
3d
as
well
right.
The
maps.
E
D
Yeah
so
so
I
have
seen
olinovsky
speak
on
on
this
and
on
3d
grid
cells,
and
they
came
to
the
conclusion
that
they're
that
the
that
the
cells
distance,
the
fields
have
the
same
r
like
rough
distance,
apart
from
each
other.
So
with
the
same
rough
distance,
although
I
should
mention
it's
not
very
regular,
it's
not
a
nice
clean
3d
lattice,
but
it
is
3d.
But
the
point
is,
if
they're
roughly
evenly
spaced,
if
the
average
distance
between
them
is
roughly
the
same,
you
would
expect
their
scale.
E
D
The
fact
that
we
describe
these
cartoonish
place
cells
like
these
cells
that
just
represent
locations
has
like
almost
nobody,
believes.
That's
what
place
cells
are.
That's
just
a
way
to
think
about
them
and
I
feel
like
I
feel,
like
that.
Cartoonish
initial
definition
has
led
to
now
these.
I
I
I
think
that
the
way
subutai
characterized
this
as
like
having
a
population
code
sort
of
like
how
we'd
use
sdrs
is,
is
a
more
accurate
way
of
thinking
about
the
place
cells
and
go
on.
C
C
No,
and
I
I
think,
you're
right
too
markus
you
know
not.
We
never
really
thought
too
much
about
play
cells
that
we
think
a
lot
about
grid
selves,
but
it's
still
surprising,
I
guess
just
it
doesn't
it
doesn't
throw
anything
out
the
window.
Yet
it
just
says.
Oh,
like
that's
another
data
point
didn't
know
about
that.
One.
B
F
F
F
Yeah
yeah,
I
thought
that
was
an
important
piece
of
context.
It
is
another
you
know.
Big
picture
thing
to
understand
is
just
the
processing
of
olfaction
in
the
brain,
so
in
the
nasal
epithelium
there
are
these
cells
that
have
receptors
that
bind
to
specific
volatile
molecules,
and
then
these
cells
project
to
the
glomerulus
and
this
coupling
seems
to
be
very
precise,
so
cells
that
express
a
particular
odor
receptor
seem
to
couple
specifically
a
particular
glomeruli.
D
F
H
E
F
So
the
other
thing
that
they
point
out
is
that,
according
to
other
researchers,
the
activation
patterns
that
you
see
in
the
glomerulus
also
appear
to
be
stable
over
time.
So
this
paper
studied
the
piriform
cortex.
So
the
question
is
basically
what's
going
on
in
the
piriform
cortex,
and
you
know.
Why
is
that
different?
F
F
No,
so
I
guess
in
other
experiments
it's
kind
of
just
floating
here,
but
here
they
cement
it
into
the
skull
to
try
and
keep
it
from
moving
they're
targeting
layer.
Two
three
pyramidal
neurons
in
the
piriform
cortex.
They
make
a
note
that
they
might
be
hitting
inhibitory
neurons,
sometimes
but
they're,
mainly
targeting
these
layer.
2
3,
neurons.
F
F
One
possible
compound
for
spike
sorting
is
that
individual
neuron
waveforms
can
shift
over
time
and
start
to
look
different.
They
make
a
note
that
from
what
they
can
tell
a
single
neuron's
waveform
doesn't
really
change
over
time
in
this
experiment,
and
that
gives
them
a
little
bit
more
confidence
in
the
results.
So
the
waveform
is
a
bit
like
the
neurons
footprint,
and
so
it
stays
kind
of
the
same
over
time.
F
F
C
This
would
seem
to
be
the
the
the
largest
concern
for
this
result.
Right
is
that
they
say
it's
relatively
stable.
These
representations
for
over
a
day,
but
it
drifts
over
time
and
of
course,
the
neural
tissue
is
also
moving
around
and
drifting
and
things
are
growing
and,
and
so-
and
you
know,
and
so
one
could
easily
argue
that
that
the
change
is
not
because
they're
not
recording
from
the
same
cells
anymore
right.
That
would
be
the
most.
C
So
I
guess
they
had
to
go
through
a
lot
of
work
to
try
to
prove
that
that
wasn't
the
case,
but
that
would
be
the
most
seems
to
be
the
most
likely.
You
know
objection
one
might
have
to
these
results.
I
guess
would
be
to
say:
hey
you're,
not
really
recording
from
the
same
neurons.
You
know
so
I
I
don't
know
how
to
judge
the
accuracy
of
the
techniques
here
right.
I
guess
they
they're
making
the
argument.
They
did
everything
necessary
to
guarantee
they're.
Looking
at
the
same
units
over
32
days.
E
C
That,
like
you
know,
because
the
cells
have
these
membranes
and
they
this
this
research
showed
that
if
you
have
a
sharp
point,
the
cell
membranes
could
wrap
around
the
tip
of
the
probe
and
they
would
literally,
they
really
react
to
the
probe,
as
it
was
some
sort
of
physical
thing
that
they
were.
They
were
drawing
around
and
moving
relative
to.
So
it's
very
hard
to
imagine
you
stick
this
probe
in
there.
These
cells
are
just
not
moving
and
static
over
32
days.
F
F
Okay,
so
they
they
do
this
odor
test
every
eight
days.
So
it's
kind
of
confusing
there's
a
there
are
eight
odors
that
they
administer
to
the
mouse
and
they
administer
every
eight
days
and
on
a
day
where
they
administer
these
orders.
They
do
it
seven
times
throughout
the
day
and
in
figure
g
here.
What
you're
looking
at
are
raster
grams
for
a
single
neuron,
so
the
x-axis
is
time.
A
dot
means
that
the
neuron
fired
a
spike
and
you're
looking
at
each
color.
F
Here
is
one
of
the
seven
trials
so
or
sorry.
You've
got
like
seven
trials
here
within
a
single
color
and
then
different
colors
are
different
days
and
in
the
two
rasters
that
I
put
boxes
around,
you
can
see
that
the
kind
of
firing
rate
activity
of
a
neuron
in
response
to
an
odor
is
changing
over
time.
F
F
B
Too,
is
these
odors
might
not
be
naturalistic
odors
that
they
sniff
regularly?
So
maybe
it's
just
a
learning
thing
that
over
the
course
of
several
days
they
they
learn
to
represent
these
orders.
It's
not
that
their
representation
is
changing.
It's
just
that
they've
never
smelt
these
odors
before
so
now.
They're
learning
to
form
representations
of
that.
F
Yeah
I
mean
they,
so
they
they
talk
a
little
bit
about
this
later
in
the
paper.
I
think
the
habituation
thing
possibly
makes
sense,
because
in
a
later
experiment
they
look
at
basically
how
frequency
of
exposure
to
the
odor
affects
drift
and
yeah
drift.
Is
it's
supposed
to
be
slower
if
you're
just
frequently
exposed
to
it?
So
if
you
just
expose
the
animal
every
single
day,
you
know
it
cuts
the
drift
rate
in
about
half.
B
G
F
F
F
F
I'm
going
to
walk
through
each
of
these
panels,
I'll
try
and
kind
of
move
quickly
to
cover
ground
so
figure.
A
we've
got
300
some
neuron
order
pairs,
so
you've
got
a
single
odor
and
you're.
Looking
at
the
firing
rates
of
neurons
now
different
neurons
can
have
just
different
firing
rate
statistics,
so
some
just
fire
more
than
others,
so
they
z-score
them
to
put
them
all
on
the
same
kind
of
scale
and
they
sort
them
based
on
rate
on
day
zero.
F
So
up
here
are
the
neurons
that
fire
fastest
in
response
to
this
odor
down
here
are
the
ones
that
actually
fire
less
than
normal.
In
response
to
this
odor-
and
if
you
apply
this
ordering
on
each
of
the
subsequent
days
where
the
rats,
or
rather
the
mice
smell
these
odors,
you
can
see
that
the
ones
that
used
to
respond
a
lot
to
this
odor
are
kind
of
changing
and
not
really
as
responsive
later
on,
and
meanwhile,
the
ones
that
used
to
you
know
be
anti-correlated
with
this
odor.
F
F
Okay,
they
look
at
the
raster
grams
use
that,
as
features
to
train
a
support
vector
machine
with
a
linear
kernel,
so
they
look
the
accuracy
of
a
linear
classifier
to
and
it's
this
is
eight
ways
classification.
So
there
are
eight
orders
that
you're
trying
to
determine
based
on
the
raster
grams
and
if
you
train
it
on
data
from
day
zero
and
test
it
on
data
from
day
zero,
you
get
decent
accuracy
and
then
that
accuracy
decays
monotonically
with
time.
F
F
So
the
correlation
here
is,
like
you
know,
it's
really
high
as
we
go
to
the
right,
we're
just
doing
the
same
thing,
but
now
we're
comparing
data
on
day
0
with
day
8
16
24
32.
The
blue
line
is
the
regression
line
and
you
can
see
at
a
glance
that
the
blue
line
is
getting
flatter.
As
you
go
to
the
right.
F
B
And
I
think
this
could
this
could
really
be
explained
as
the
animals
just
learning
to
represent
these
odors.
It's
that
first
takeaway
of
responses
to
orders
are
similar
within
a
day
but
change
over
time.
That's
true
for
almost
anything
that
involves
ltp
or
long-term
plasticity
and
there's
all
sorts
of
consolidation
and
stuff
that
happens
overnight
and
and
anything
that
involves
long-term
learning.
You
would
get
that
statement
that
that
statement
would
be
applicable
to
it.
B
F
F
F
Time
this
slide
is
more
or
less
just
doing
some.
Well,
there
are
a
couple
things.
First,
they
do.
They
do
some
controls,
so
they
look
at
how
individual
neuron
firing
rates
overall
change
with
time,
and
there
doesn't
seem
to
be
very
much
change.
They
look
at
the
population
sparseness,
so
what
fraction
of
neurons
are
on
at
any
given
time?
F
It
also
doesn't
seem
to
change
very
much
over
the
course
of
these
32
days,
yeah,
I
kind
of
forget
which
all
things
they
looked
at,
but
the
point
is
that
they
did
a
series
of
controls
and
there
wasn't
very
much
change
these
principal
component.
Plots
are
just
every
day
that
they
administer
odors,
they
calculate
principal
components
and
then
they
plot
the
odor
on
the
first
two
pcs.
H
F
F
Then
they
do
a
behavioral
assay
to
verify
that
when
the
animal
smells
this,
they
try
and
run
away.
So
the
animal
is
indeed
learned
to
be
afraid
of
this
smell.
The
cs
plus
smell,
I
think
ingress
is
like
I
don't
know.
Maybe
if
someone
knows
experimental
neuroscience
better,
they
can
say,
but
I
think
it's
like
the
animal
moving
backwards
or
something
trying
to
get
away
anyway.
F
F
And
now
you
look
at
the
rate
of
drift,
so
I
kind
of
glossed
over
it,
but
they
they
compute.
They
fit
an
exponential
decay
model
to
the
angle
between
population
vectors
and
you
basically
have
similar
drift
rates,
even
if
you're
shocking
the
animal
so
I'll
pause
briefly
to
reconsider
subutai.
Your
idea.
That
learning
could
just
explain
this,
because
here
the
animal
is
learning
to
associate
the
smell
with
you
know,
a
negative
reinforcement
and
the
drift
rate
is
the
same.
H
F
F
B
F
Yeah,
that
is
what
I
would
expect
yeah
so
now
the
rate
of
drift
is
the
same
still,
though,
if
you
spend
16
days
familiarizing
with
the
odor
and
then
you
look
at
drift
over
the
following
16
days,
where
you
only
have
the
animal
smell
every
every
eight
days.
So
that's
I'm
I'm
looking
at
this
figure
here.
F
F
F
F
E
B
B
B
Or
habituation,
it's
not
really
that
it's
constantly
turning
it
changing.
It's
just
you're,
either
learning
to
code
it
and
then
it's
fine
or
habituation
is
just
a
normal
thing
that
we're
kind
of
used
to
you
know
the
representation
representation
go
away
and
then
come
back
again.
If
it's
infrequent,
I
mean.
G
F
B
F
B
C
You
know
the
thing
about
this
to
me
is
like
the
piriform
cortex
is
such
an
unusual
thing.
I
mean
it's,
not
it's
just
different
than
other
neural
tissue,
and
so
if
this
phenomena
is
really
just
isolated
to
the
piriform
cortex,
then
it's
it's
still
interesting,
and
but
it's
not
like
a
more
fundamental
ground-shaking
observation
to
me.
C
F
Talked
about
what
other
people
did,
so
they
didn't
look
in
this
paper.
They
cite
a
bunch
of
papers,
saying
that
sensory
representations
are
more
stable
in
other
modalities
and
okay
in
the
discussion
they
note
a
handful
of
references,
maybe
only
five
references
in
total
that
there's
been
representational
drift
observed
in
other
areas.
F
G
G
No,
that
that
was
that
was
that
was
the
extent
of
it
because
we're
we're
saying
you
know
it
could
be
a
learning
thing.
The
question
is,
what
is
actually
being
learned?
Is
it
identification
or
is
it?
This
is
different
than
what
I've
you
know
currently
noted,
and
I
should
pay
attention
to
it
for
xyz
reason
or
not.
C
C
You
know
the
hippocampus,
of
course,
is
also
known
as
a
fast
learning
system
and
it's
it's
one
of
the
first
places
people
observed
new
neurons
being
created
all
the
time,
so
there's
a
general.
You
know
belief
that
the
neural
population
is
turning
over
there
more
rapidly
than
other
parts
of
the
brain.
C
C
C
C
C
H
Fine,
I'm
not
completely
familiar
with
everything,
but
I
know
that
some
part
of
the
olfactory
bulb
project
preferentially
in
some
subsport
of
the
prefrontal
cortex.
So,
for
example,
other
balance,
if
you
like
an
order,
if
you
don't
like
it,
have
some
preference
in
where
it's
projecting
onto
the
peripheral
cortex,
which
seems
to
have
some
link
with
reward
system,
for
example,
but
I'm
not
a
specialist
of
the
peripheral
cortex
itself.
So.
E
C
F
That's
what
I
had
heard
and
I
think
there
was
like
kind
of
I
don't
know
how
much
evidence
there
is
for
this.
But
there
was
like
the
idea.
I've
heard
before
that
smells
are
supposed
to
be
more
evocative
of
memories
than
other
senses,
and
the
explanation
had
to
do
with
it
being
close
to
the
hippocampus.
C
So
yeah,
I
guess
I
was
much
more
bothered
by
this
paper
when
I
thought
they
were
going
to
make
a
broader
argument
that
representations
about
central
modalities
are
going
to
be
changing
rapidly
and
if
it's
really
just
the
olfactory
system,
well,
then
the
hell
that's
screw
system
to
begin
with
so
well
it
doesn't
our
theories,
don't
involve
it
too
much.
You
know
you
don't
have
good
representation
of
the
olfactory
system,
the
cortex,
it's
less
of
a
sensory
motor
system.
You
know,
there's.
C
C
F
D
B
C
C
C
Well,
I
think
our
representations
of
the
world
are
pretty
stable
right.
I
mean
I
can
remember
things
from
long
terms
ago
and
and
we
have
to
be
able
to
call
them
and
interact
with
you
know
all
my
memories
about
the
world
have
to
be
able
to
interact
properly
with
one
another.
So
if
representations
change
dramatically
over
time,
I
don't
know
how
that
how
any
of
that
sort
of
coordination
would
work.
F
C
Well,
if
it
was
drifting,
then
you'd
have
to
have
everything
would
be
drifting
right
and
all
members
would
have
to
be
formed
over
and
over
and
over
again
it
would
be
like
everybody's
learning,
all
the
time.
It's
like
I
mean
it's
possible,
it
just
would
be.
It
would
just
be
a
pretty
big
surprise.
I
guess
to
me
at
least.
F
E
B
C
I
you
know,
that's
sort
of
the
predictive
coding
idea
right
right
and
that's
that's
never
how
I
thought
about
it.
I
mean
it's
not
about
minimizing,
surprise
like,
but
you
could
argue
that
making
an
accurate
prediction
in
some
sense
minimizing
the
price.
I
don't
know,
but
it's
it's.
You
know
in
the
predictive
coding
ideas
you're
trying
to
like
you
know
you
only
pass
on
surprises
like
you've
said
so
you
you
things
that
that
are
unexpected.
You
pass
them
on
to
the
next
regenerator
whatever
and.
C
F
C
F
C
E
C
C
Yeah,
if
the
thing
you
know
the
way
this
might
work,
let's
say
you
and
I
are
two
columns.
Looking
at
the
water
bottle
and
I
said
okay,
I
know
what
this
thing
is
and
you
say
you
know
what
this
thing
is
and
we
can
okay.
At
the
same
time,
we
agree
that
something
that
we're
seeing
is
the
same.
Then
we
would
make
an
associative
connections
between
our
representations
of
that.
C
So
you
know
there's
no
meaning
in
what
my
neurons
mean
to
you
just
saying
it's
the
same
thing
that
I'm
looking
at
now
so
I'll,
just
I'll
just
associate
your
pattern
with
my
pattern
and
now,
if
your
pattern
changes,
I
can't
associate
it
anymore.
I
would
know
like
it's
like
so
now
later
I
just
see
the
water
bottle
and
you
don't
but
my
but
but
and
if
I
didn't
have
drift,
I
can
invoke
the
water
bottle
in
your
column.
C
I
can
see
you
know,
I'm
seeing
a
water
bottle,
you're,
probably
seeing
one
too,
and
you
go
well
given
everything
I
know.
I'm
gonna
assume
there's
a
water
bottle
here,
but
if
my
representation
changes,
then
you
have
no
idea
what
I'm
looking
at
I'd
be
like.
Oh,
I
don't
know
if
this
is
a
strange
pattern,
so
the
voting
wouldn't
work
if
somebody's
representation
changed
or
if
everyone's
changed.
C
Or
continuously
yeah
I
mean
any
associated
memory
is
going
to
have
that
property
right
any
associated
memory.
Once
you
form
the
memory,
you
assume
that
the
two
populations
you're
associating
each
other
are
stable
and
if
not
then
you're
right,
then
you'd
have
to
continually
learn.
You'd
have
to
say
you
know
yesterday,
we
all
knew
this
was
the
water
bottle,
but
today
we're
going
to
learn,
it
again
know
so,
which
could
it
could
happen?
I
wouldn't
say
this
out
of
the
out
of
the
question.
E
B
C
You
know
the
the
one,
the
one
counter
example,
and
maybe
someone
else
knows
this
literature
better
than
I
do,
but
there's
there's
a
whole
bunch
of
papers
that
came
out
some
few
years
ago
about
hey
when
you
retrieve
a
memory
it
actually
erases
it
and
writes
it
again.
I
don't
remember
the
details
of
this
at
all,
but
it
was
like
the
idea
that
that
reading
is
a
destructive
process
in
some
part
of
the
brain.
I
don't
know
which
part
of
the
brain
was.
C
I
don't
remember
any
details,
but
it
was
like
hey
when
you're
recalling
this
particular
memory.
It
destroys
the
memory
and
it
has
to
be
right
right,
write
it
out
again,
and
so
people
were
using
this
as
sort
of
a
way
of
getting
doing
some
behavioral
therapy
because
they
could
say
you
know
you
could
try
and
get
people
rid
of
their
phobias
or
something
like
that.
You
could
yeah.
C
Yeah,
so
I
mean
so
if
what,
if
that
were
true
everywhere
in
the
brain,
I
have
no
idea,
I'm
just
totally
making
this
up.
So
what?
If
that?
Would
you
everywhere?
And
you
say,
every
time
you
read
anything
out
it
that
it
has
to
form
new
synapses,
it
seems
impossible.
But
what,
if
that
were
true,
then
I
guess
you
would
be
continually
learning
all
the
time.
So
I
don't
know
they
would
maybe
get
around
the
drift
issue.
H
H
H
You
can
modify
it's
how
false
memory
emerge.
Basically,
you
can
modify
your
memory
when
you
recall
it
and
that's
why
for
integration,
you
need
to
ask
new
to
not
prime
start
a
person
you
integrate
interrogating
with
some
false
memory
that
you
want
to
implement,
because
when
they
recall
something,
then
you
can
modify
that
memory.
As
you
recall,
yeah.
C
C
It's
a
it's
a
weird
tweak
on
the
memory
recall
process.
It
could
lead
to
false
beliefs
or
false
memories,
but
it
really
doesn't
really
change
anything
fundamentally.
Does
it
is
it?
Is
it
and
I
couldn't
think
of
how
it
did
in
some
sense?
I
said:
well
just
try
to
ignore
it
for
now,
because
it
may
not
be
that
important
for
the
general
idea
of
how
information
is
processed,
it's
obviously
important
for
biological
brains,
but
it
wasn't
clear
to
me
that,
like
but.
C
H
C
H
C
But,
but
is
it
again
forget
the
intermediate
in
the
long-term
form
just
say:
I've
got
this
memory.
How
does
the
system
work
right?
How
is
it
processing
information?
There's
neurons
that
represent
location?
There's
no
represent
the
sensory
pattern
there
you
know.
Does
it
change
that
I
mean?
Does
you
can
separate
out
this
sort
of
biological
details
of
memory
formation
and
I'm
not
hearing
from
you
that
this
the
difference
in
me,
you
know
none
of
our
theories
really
rely
from
industry
at
all
between
short-term
mid-term,
long-term
memories.
C
G
Jeff,
when
you,
when
you're
talking
about
the
you,
don't
want
drift
because
the
the
voting
would
start
to
fail.
Is
there
a
mechanism
there
that
allows
you
to
have
a
general
to
have
you
know
one
concept
of
what
the
coffee
cup
is
and
generalize
it?
Is
it
a
constant,
rewriting
or
reinforcing
allows
it
to
be
less
specific
and
more
general?
Is
that
a
mechanism
that's
kind
of
built
into
your
formulation?
I.
C
C
So
so
I'm
just
clarifying
that's
a
broader
claim
than
voting
and
voting
is
just
to
talk.
You
know
just
one
instance,
so
could
it
lead
to
some
form
of
generalization?
I
don't
know
you
have
to
work.
You
have
to
work,
walk
it
through
and
explain
how
you
know.
Just
drifting
on
its
own
doesn't
seem
like
a
good
way
of
generalizing.
G
No,
no,
I
mean
if
it's
if
it,
if
it's
drifting
by
not
being
re-introduced
to
where
the
stimulus
was
then
that's
not
very
helpful,
but
I'm
I'm
just
trying
to
think
of
if,
if
you're
not
referencing
up
to
a
hierarchy.
For
a
surprise,
you
I'm
looking
at.
If
there's
this
intrinsic
mechanism,
you
have
to
explain
how
you're
also
able
to
generalize
as
well
and
that's
what
I'm
trying
to
grasp.
Well.
C
C
C
G
C
Well,
I
don't
know,
look
in
general
the
way
I
approach
these
things
is
like
there's
so
many
things
going
on
in
neuroscience.
You
have
to
you,
have
to
be
very
selective
and
very
the
challenge
is
to
figure
out
which
ones
you
can
put
aside
for
the
moment,
which
ones
you
can
and
that's
a
really
hard
task
and
that's
like
the
hardest
part
of
the
whole
thing.
It's
like
what
are
the
key
things
we
have
to
focus
on.
C
What
are
the
ones
we
don't
want
to
focus
on,
and
so
everything
we
do
is
is
going
through
that
process.
So
I
I
know
I
the
drift
thing
was
new
to
me,
but
the
the
rewriting
of
the
memory
idea.
That
was
one
that
a
few
years
ago
I
said
yeah,
it
doesn't
seem
to
be.
That
seems
like
one
we
can
put
on
the
back
burner
here
and
not
worry
about
it
right
now,
along
with
a
gazillion
other
things,
and
I
would
generally
put
anything
that's
unique
to
the
olfactory
system.
C
But,
but
that
you
could
argue
about
everything
right,
we
just
don't
know,
but
everything
could
be
argued
like
you
know,
like
you
know,
maybe
the
the
the
you
know
the
clusters
and
neurons
in
your
spinal
cord
are
really
going
to
be
essential.
For
you
know
I
don't
know,
understanding
physics
seems
unlikely.
C
C
This
is
the
general
principle
of
theorists
in
any
field.
You
have
to
decide
what
things
you
can
ignore
and
what
things
you
have
to
pay
attention
to
and
that's
where
the
skill
designs
a
lot
of
the
skill.
So
anyway,
I'm
just
giving
my
I'm
not
saying
I'm
right
about
this.
I'm
just
saying
this
is
how
I
thought
about
those
things.