►
Description
https://link.springer.com/article/10.1007/s10827-019-00729-1 from visiting scientist Florian Fiebig. He says:
Its a brief 6 page paper, and I think it can serve as a neat introduction to the kinds of spiking neural networks and model thinking about the cortical microcircuit I was working on for my PhD.
The main idea in short:
Many Hebbian Learning Rules violate Dale's principle (A neuron cannot be both excitatory and inhibitory, all its axons release the same neurotransmitter) in the course of dynamic synaptic weight learning, because it this may change the sign of an individual connection. On the example of a reduced cortical microcircuit originally built as an attractor model of working memory, we show how biological cortex might instead learn negative correlations through a di-synaptic circuit involving double bouquet cells (DBC). These cells are very particular in the way they are distributed regularly across the cortical surface and innervate the whole minicolumn below without affecting neighboring columns. "Indeed, disregarding some exceptions, there appears to be one DBC horsetail per minicolumn"
A
B
Currents
go
in
those
currents,
change
the
change,
the
potential
there's
something
akin
to
a
fine
edge
hold
it's
not
exactly
a
heart
threshold
because
it
actually
depends
on
the
sharpness
of
the
rice.
So
there's
some
time
here
this
exponential
you
see,
which
means
that
when
the
neuron
gets
close
to
the
spine
threshold
VT,
then
it'll
it'll
kick
off
right
and
there's
a
bunch
of
cars
that
go
into
it.
You
don't
need
to
go
into
much
detail
about
that
main
point
already
here
that
you
see
all
these
G's
all
the
times.
This.
C
B
D
B
B
B
Those
ringabel
which
drive
the
membrane
potential
towards
zero,
which
means
closer
to
the
firing
threshold
and
sort
of
if
you
have
a
lot
of
those
transmissions
and
inaudible
fire,
but
there's
also
other
other
driving
potentials
like
inhibitory
receptors.
So
GABAergic
transmission
that
drive
in
toronto
was
75
millivolts
or
something
so.
B
B
D
A
B
B
B
B
D
B
C
D
B
You
you
kick
off
the
system
and
it
just
it
exhausts
is
capability,
you
know
until
it
reaches
a
new
equilibrium,
you
know
someone
out
there
and
then
this
of
course
you
know
different
ones
up
here
that
then
kick
off
to
drive
it
back
again
yeah.
So
if
you
are
two
different
ions,
you
know
potassium,
you
know
like
change
in
an
outflow
and
thus
change
the
membrane
voltage,
but
I'm
not
here
to
talk
about
ion
channel
based
neurons.
Those
also
exist,
but
this
is
a
simplification
of
that,
but
it.
B
B
B
B
Yes,
here
it
is
right,
so
in
fact
what
is
happening
not
it's
not
that
at
every
spike,
there's
a
constant
particular
current
that
briefly
gets
injected
into
the
neuron,
but
instead
what
happens?
Is
that
there's
a
dynamic,
fast
change
in
the
conductance
towards
a
potential
that
depends
on
the
specific
transmitter?
B
B
Types
weird
like
NMDA
receptors,
which
are
also
excited
to
Tory
receptor
types,
which
also
drive
the
membrane
towards
zero,
but
are
a
little
bit
different
in
that
they
have
a
different
time
course.
So
that's
why
you
have
this
expression
here.
The
conductance
changes
over
time.
It
might
be
very
brief
change
and
in
that
case,
very
little
current
right
will
actually
flow
here,
because
it's
going
to
be
a
brief
increase
in
the
input
current
and
then
it's
going
to
go
down
again.
So
that's
on
the
order
of
5
milliseconds
one
powerful
example.
E
B
The
that
is
a
good
question,
so
that
goes
towards
that
biomechanics
of
these
ion
channels
that
get
open
right.
In
fact,
it
is
this,
there's
an
equilibrium
that
gets
reached
when
these
channels
get
opened
and
they
have
sort
of
a
limited
driving
ability
and
then,
depending
on
how
many
you
have,
of
course,
you
might
get
there
faster
or
slower,
but
indeed
that
equilibrium
depends
on
these
different
ions
and
and
those
don't
change
right.
I.
D
D
B
B
D
D
Going
back
and
forth
between
those
and
models.
Okay,
all
these
very
complicated
parameters.
You
can
build
models
only
way.
You
can
model
all
the
ion
channels.
You
can
learn
these
dendritic
properties
do
very
sophisticated
tools.
Others
deployments
just
describe
what
his
neurons
are
doing.
Is
it
or
not?
You
know,
on
the
run
so
now
my
channels
he's
just
beating
his
simulation
is
not
modeling.
That
level
of
detail
he's
modeling
other
levels
of
detail,
but
he
also
said
he's
not
meddling
and
modeling
dendritic
spikes
or
modeling
comportment.
All
these
pretty
model.
A
D
This
finally
good
selling,
all
these
little
circuit,
embedded
in
them
finally,
layer
which
are
these
channels
and
they're
just
and
there's
very
specific,
something
like
in
this
I
on
some.
Let
out
that
I
on
and
of
course,
they
can
probably
do
that
in
there's
a
gradient
right.
So
that's
part
of
the
question
of
the
week,
so
diem
channel.
D
B
You
will
not
capture
all
the
biological
detail,
but
because
it's
a
simplifying
assumption
that
you
can
replace
this
whole
zoo
of
you
know
important
ion
channels
with
with
sort
of
like
a
simple
equation
equation
equation
like
that,
but
it
captures
enough
that
it's
really
interesting
more
meaning
that
you
can
show
lots
of
things
that
a
simpler
model
would
not
be
able
to
show
good.
So
one
thing
that
this
original
term
up
here,
which
describes
the
evolution
of
the
membrane
voltage,
does
appear
right.
B
You
see
the
differential
equation
for
the
membrane
voltage
is
that
by
default,
there's
some
leak
channel
right,
so
some
fixed
conductance
that
will
slowly
leak.
The
membrane
voltage
towards
what
is
called
a
leak
potential.
So
if
we
take
a
look
at
parameter,
values,
I'm
sure,
there's
a
column
here
I
here
we
actually
go
here,
somebody
from
right,
so
you
see
that
the
leak
reversal
potential
is
at
minus
76
millivolts,
so
by
default,
if
you
don't
have
any
inputs
on
to
the
neuron
will
be
towards
this.
B
D
D
B
So
you
get
lots
of
physical
parameters
here.
These
parameters
were
fitted
for
a
specific
cell
type.
These
double
key
cells,
so
they
can
be
reasonably
well
described
in
their
behavior,
like
what
we
see
in
terms
of
firing
rates
and
how
excitable
they
are,
and
and
and
all
these
things
and
typically
observe
membrane
voltages
addressed
by
these.
These
parameters,
obviously
they're
tuning
involved
to
try
to
get
the
model
to
fit
what
has
been
observed.
D
B
B
A
B
With
like
a
big
premier,
pyramidal
cell
here,
right
with
the
base,
all
you
know
integration
zone
Africa,
and
you
also
see
that
capacitors
cell
here,
which
are
probably
the
most
powerful
inhibitory
cells
and
cortex,
which
are
capable
of
shutting
down
activity
in
their
surroundings
when
they
get
triggered,
and
they
can
fire
very
fast.
That
would
be
cases
are
not
like
that.
They're
a
little
bit
different.
There.
B
Neurons,
so
they
release
GABA,
which
is
this
inhibitory
receptor
that
drives
the
cells
that
it
targets
when
it
gets
released
towards
zero
millivolts.
So
it's
also
an
inhibitory
neuron,
but
it
has
a
rather
particular
shape
there.
It's
called
taboo
bouquet
cells,
because
there's
two
bouquets
that
you
see
here,
you
know
one
going
and
going
up
and
one
going
down
of
its
inputs
where
it's
sensing
and
then
you
have
this
long
horse
tail
right
of
connections
that
it's
making
through
the
entire
column.
That's
the
axon.
D
B
You
could
imagine
that,
where
these
holes
tails
right
these,
these
long
axons
are
getting
close
to
the
dendrites
of
this
parameter.
Neuron,
for
example,
you
might
form
inhibitory
synapse,
so
this
w
key
cell
might
be
capable
of
changing
the
membrane
potential
of
his
pyramidal
cells
towards
zero
millivolt
whenever
it
finds.
B
B
Of
those
interesting
thing
about
WK
cells
that
they
actually
innovate,
the
entire
column
and
the
radius
of
these
horse
tails
is
something
like
32:16
microns
or
something
micrometers.
So
so,
as
a
result,
you
get
this.
So
if
you
look
at
it
from
the
side,
do
this
as
an
aside
here
right,
you
get
these
very
nice
long
streaks
that
go
through
the
entire
cortical
sheets,
so
this
is
immuno
st..
D
B
B
Spacing
of
these
double
kisses,
so
they're
kind
of
distributed
around
new
cortex
in
most
of
new
cortex
there's
some
areas
that
don't
have
them,
or
at
least
they
might
have
a
different
gene
expression.
So
we
don't
find
them
when
we
label
them
with
these
dyes
that
are
that
sort
of
get
triggered
by
certain
chemicals.
On
the
surface
of
these
double
kisum's
but
sort
of
in
in
the
majority
of
cortex,
you
find
this
very.
C
B
D
B
So
it
was
the
most
authoritative
we
could
find,
and
they
said
it's
pretty
much
one
per
mini
column.
There
might
be
sometimes
a
mini
column
or
two
that
don't
have
one
now,
particularly
you
know
just
neighboring
one,
because
they
tend
to
spread
out
so
nicely
in
cortex
that
when
you
have
a
double
kiss,
all
you're
very
unlikely
to
find
it
double.
Okay,
so
right
next
to
it,
do.
C
D
Asked
questions
like
well:
we
assume
that,
as
it's
absurd,
you
see,
these
cells
are
bramasole.
Excitatory
cells
in
the
mini-com
all
share
a
basic
receptor
field
problem
and
the
Carson's.
How
how
do
they?
How
did
that
come
about
right
through
something
has
a
for
staff
to
curve
and
other,
and
we
assume
that
these
cells,
probably
differentially
under
different
conditions
together.
D
Forcing
all
these
guys
to
do
the
same
thing
at
the
same
time-
and
we
documented
this
in
a
paper
or
my
speculation-
was
that
the
double
bouquet
cells
would
actually
inhibit
the
basket
cells
at
some
at
some
period
of
time,
because
otherwise,
the
basket
cells
don't
work
all
the
cell's
understanding
the
best
shut.
The
old
out
of
these
markings.
Ever'one
and
there's
give
me
some
perspective.
Oh
yes,.
D
B
It's
it's
very
fascinating.
It
was
to
us.
We
were
also
I
mean
we
were
looking
for
a
mechanism
to
get
targeted
inhibition
to
a
mini
column,
and
it
was
surprising
to
us
that
the
experimentalist
was
so
adamant
about.
There
might
be
exactly
one
double
bouquet
cell
per
minute
yeah,
and
it
turns
out
that
when
you
find
one
this
other
one.
D
B
B
The
double
bouquet
cell
is
capable
of
controlling,
inhibit
through
inhibition
right,
the
entire
population,
below
it's
virtually
impossible
for
a
pyramidal
cell
in
the
in
that
mini
column,
to
not
have
some
DBC
input
if
there
is
a
DBC
cell
in
that
Minnie
column,
so
you're
basically
guaranteed
that
you
have
control
right,
inhibitory
control
over
over
one
specific
column
and
these
horses,
that
don't
branch
out
to
other
mini
camps,
they're
very
constrained.
There
is
like
a
local
column
in
the
column
that
runs
through.
D
D
B
D
B
C
B
C
B
D
D
B
D
A
D
D
B
So
if
you
pick
two
pyramidal
cells
in
the
mini
column,
that's
like
a
20
percent
chance
that
they're
connected
as
that's
what
we
call
a
20
percent.
We
can't
connect
to
be
where
these
cells
locally
connect
with
one
another,
and
so
you
have
to
make
an
argument.
How
lot
of
population
of
that
size
would
be
capable
of
self
activating?
And
when
you
look
at
the
typical
size,
often
excited
to.
D
D
C
B
There
is
now
a
single
spike
right,
so
let's
say
this
time
point
here.
There
is
an
example
spike
that
comes
in
from
from
an
accept
renewal.
That
target
is
neuron.
It's
going
to
drive
to
know
towards
your
millivolt
for
some
short
amount
of
time
right,
so
that
might
be
behind
this
rise
here.
But
the
problem
is
that
this,
this
kind
of
a
rise
is
very
small,
like
the
the
epsp
size
is
about
about
a
millivolt
for
rather
strong
synapses,
meaning
you
will
never
get
from
from
the
resting
potential
head.
B
B
A
B
Of
course,
we
still
have
competition,
but
that
competition
tends
to
be
through
a
shared
pool
of
basket
cells
in
the
hypercar.
So
if
you,
this
is
a
schematic,
of
course
right,
not
anatomically
and
correct,
but
you
sort
of
see
this
macro
column
here,
right
in
gray
and
for
simplicity
sake,
you
only
show
now
two
of
these
mini
parts
and
inside
those
mini
columns.
The
main
workhorse
of
all
of
these
are
these
thirty
pyramidal
cells.
B
D
D
A
D
E
B
Reason
why
the
wk
say,
or
a
cell
came
into
this-
is
that
we
have
problems
with
the
the
problem
that
the
learning
rule
that
we
were
using
learns
negative
weights
and
the
only
way
to
get
around
that,
while
still
preserving
the
ability
to
learn
inhibition
amongst
the
currently
connected
pyramidal
cells,
is
to
have
a
Dyson
optic
circuit.
So
we
integrated
W
key
cells
because
we
had
a
very
functional,
a
specific
role
for
it
in
mind
to
to
fix
the
violation
of
Dale's
principle.
That
is
principle.
A
E
B
D
B
B
C
B
We
have
is
a
lot
of
dynamic
plasticity
with
regards
to
the
potentiation
of
deep
potentiation
of
connections,
which
is
a
big
difference.
There
are
to
some
extent
equivalent
this
people
who've
done
sort
of
the
math
on
that
like
for
klosty,
it's
been
interesting
papers,
comparing
structural
plasticity,
who
was
the
synaptic
plasticity.
D
B
D
B
C
D
B
B
B
B
B
D
B
It
can
still
be
quite
fast,
particularly
if
the
path
links
are
short
right,
if
you
are
in
the
same
column,
for
example,
so
that
I
septic
inhibition
and
when
you
want
to
learn,
there's
two
ways
to
learn
it:
either
you
learn
the
strength
of
excitation
onto
an
inhibitory
cell
that
is
specific
for
the
population
P.
So
that's
why
the
TBC
cells
are
so
amazing,
because
the
TBC
cells
sit
on
top
and
like
the
the
lower
parts
of
layer,
two
layers
we
a
and
then
they
have
these
false
tales
that
run
all
the
way
down.
So.
D
A
B
So
you
have
two
ways
to
learn:
now:
either
you
learn
the
strength
of
the
excitatory
connection
on
to
the
DBC
cell,
so
you
make
that
strong
when
the
anticorrelated
strong
positive
right,
because
it
can
only
be
positive
because
it's
an
excitatory
protection,
all
you
learn
all
this
is
fixed
right
and
you
learn
the
strength
of
the
inhibition
because
there's
two
different
ways:
we
look
at
the
literature
and
it
turns
out
that
these
synapse
is
inhibitory.
Sinuses
of
the
DBC
cells.
Don't
have
much
in
terms
of
plasticity.
C
D
D
D
C
B
Solution
to
the
circuit
is,
if
you
want
to
be
able
to
learn
anti
correlations,
then
you
need
to
be
able
to
train
this
connection
here.
Wait
that
is,
that
is
the
weight
you
want
to
learn,
and
so
previously,
in
this
patient
confidence
propagating
neural
networks,
the
learning
rule
was
such
that
it
can
learn
negative
weights.
I
will
briefly
show
it.
D
B
So
that
the
basic
idea
of
the
of
the
pcpn
and
learning
rule
is
that
you
learn
the
level
of
activity
pre
and
postsynaptic
Li,
and
then
you
use
those
statistics
of
pre
and
postsynaptic
activation
and
Co
activation
to
compute.
These
two
properties
here
wait
between
two
notes
and
a
bias
term,
so
you
can
think
of
this
as
the
probability
of
those
two
neurons
firing
together
and
these
as
the
probability
of
the
first
neuron
fire,
and
this
is
the
probability
of
the
second
neuron
to
fire.
B
So,
let's
make
a
very
simple
example
like
just
to
trivialize
it
a
little
bit.
Let's
say
the
probability
of
the
connection
from
cell
I
to
cell
J.
Right
say
that
is
a
connection
right
and
let's
say
that
the
probability
of
the
pre-cell
right,
the
eye
cell,
is
50
percent
right,
so
it
fires
half
of
the
time.
Let
obviously
you
have
to
define
a
time
frame.
B
B
Non
je
pre
and
postsynaptic
neuron
and
let's
say
this:
one
fires
with
50%
probability
and
this
one
fires
with
50%
probability
and
the
likelihood
that
the
fire
together,
if
that
is
actually
the
product
of
the
two,
then
this
term
will
go
to
1,
which
means
that
the
weight
goes
to
goes
to
0
right.
It
means
there
is
no
connection
between
those
two
neurons:
they
don't
have
an
influence
on
each
other,
they
are
actually
independent
and
you
derive
that
from
the
statistics
of
activation.
B
If,
on
the
other
hand,
the
neurons
have
a
strong
connection,
then
this
term
will
go
up
right,
so
you're
much
more
likely
to
see
a
spike
in
the
jano
one
when
there
was
also
a
spike
in
the
idler
on
this
might
be
a
higher
term.
I,
don't
know
0.9
from
connection,
and
now
you
still
have
to
normalize
this
term
by
the
individual
firing
rates
like
the
P
I
and
the
PJ.
B
But
you
will
now
get
a
strong
excitatory
wait.
The
reason
for
this
logarithm
here
is
so
that
it's
mathematically
useful
in
the
sense
that
you
avoid
weight
explosions
because
of
the
logarithm,
even
if
the
correlation
is
really
strong.
Let's
say
it's:
a
combination
of
1
by
2
guaranteed
a
postsynaptic
spike
when
there
is
a
priest
that
they
always
fire
together
right.
That
weight
is
still
not
going
to
explode
in
your
learning,
so
it
kind
of
keeps
it
on
the
way
to
little
bits.
B
B
Argument
that
a
colleague
of
mine
Holly
made
in
his
thesis
I'm
not
sure
it's
worth
presenting
here,
but
the
broad
point
is
that
there's
a
mathematical
reason
for
that
model,
rhythm
Anna
turns
out
to
be
quite
useful.
The
problem
with
it
is,
though,
that
it
means
this
network
is
capable
of
learning
negative
weights,
and
that
is
not
physiologically
possible
between
excitatory
neurons.
D
B
There
were
only
static
negative
weights,
then
you
could
always
make
a
model
argument.
Look
if
their
weight
is
negative.
We
just
presume
this
and
inhibitory
neuron
in
between
that
does
something
like
this,
but
the
problem
is
that
the
learning
rule
is
actually
dynamic,
so
weight
might
be
positive
at
first,
because
neural,
a
and
B
tend
to
be
very,
you
know,
collective
and
they
not
just
by
chance.
C
B
The
way
from
positive
to
negative
that
works
and
model
just
fine,
it's
a
dynamic
model,
it
can
run
and
adjust.
It
might
change
its
behavior,
but
it
does
not
work
in
topology,
because
an
excitatory
neuron
cannot
become
an
inhibitory
neuron
and
the
other
way
around.
So
you
need
to
if
you,
if
we
still
want
to
be
able
to
defend
this
kind
of
a
learning
model
as
biologically
defensible.
B
D
B
D
E
E
B
B
B
Its
value,
how
is
that
changing
so
the
to
get
to
the
to
the
actual
showing
of
the
paper
here?
What
you
see
here
in
this
little
little
sort
of
next
to
the
circuit
diagram,
I
hope
you
roughly
understand
now.
I
mean
I,
there's
a
lot
of
lines
here,
but
the
point
is:
there
are
two
different
patterns
in
the
system.
One
gets
activated
by
still
zero
and
the
other
one
is
activate
by
still
one.
B
B
D
D
B
D
D
D
A
B
D
C
B
B
B
B
B
B
So
what
you
see
is
that
when
there's
no
stimulus
present
right,
these
cells
tend
to
a
sparsely
fire
like
this
occasionally.
So
this
is
like
a
full
second
down
here.
You
see
the
time
in
milliseconds
below
so
from
zero
to
1000
is
one
second,
and
you
see
one
two,
three,
four
spikes
right.
So
it's
like.
B
D
B
B
D
B
D
D
B
B
C
B
A
B
B
To
do
this
between
excitatory
cells
right
instead
we're
doing
all
this
where
we
learn
these
excitatory
ways
and
we
learn
these
excitatory
ways,
and
so
we
showed
that,
in
fact,
this
is
exactly
what
happens
so
you
used
to
have
negative
lights,
say,
say
of
you
know:
what
is
this
minus
1.8
between
competing
mini
columns
and
now,
instead
in
the
new
model
that
shifts
right?
So
the
previous
model,
that
is,
the
learned,
weight,
distribution,
lwd
and
now
the
learn?
B
C
B
D
D
D
A
D
They
form
these
little
sets
of
synapses
along
the
beginning.
You
can
see
them
right
here.
They
form
a
set
to
synapses
along
these.
Are
these
I
believe
are
synapses
from
the
where
they
not
for
the
most
general,
because
they
have
no
Campbell's
vertical
Campbell
and
Beyond
they're,
like
all
over
Baltimore.
Many
comes
like
yeah.
D
E
Which
are
pure
inhibition,
suggested
here,
I
here,
Division
one
set,
would
say
well
what
these
things
to
be
fully
and
correlated
suppose
you
want
to
say
they're
only
partially
partially,
because
basically
there's
these
overlaps
as
you
between
the
down.
So
if
you
want
to
learn
if
it's
useful
in
some
kind
of
math
sentence,.
D
E
E
When
you
have
something
it's
an
area
here
lies
like
this
yeah
and
then
the
next
one
over
is
whatever
it
is:
15
degrees
yeah,
wherever
that
there's
in
the
in-between
state,
between
those
orientations,
both
of
them
are
going
to
say,
there's
something
close
to
I
and
trained
to
look
recognized
right.
So,
if
you're
looking
at
this
thing
and
say,
okay,
well,
just
raw,
you
learning,
you
know
this
is
firing
it.
This
is
firing
with.
B
Proposed
so
an
interesting
tip
but
I
mean
I,
don't
think
it
solves
it.
But
it's
an
interesting
observation
but
of
course
it's
important
to
keep
in
mind
that
these
columns
they
have
an
anatomical
equivalent,
often
but
technically
they
are
not
anatomica
columns.
They
have
functional
commerce,
meaning
yeah
I'm.
Talking
about
the
minicamp
well,.
B
So
so
that
suggests
that,
like
a
more
anatomical
view
of
this
is
proper,
but
the
question
is,
of
course,
is
what
makes
a
mini
column
a
mini
coma
and
one
of
the
most
interesting
papers
that
I
read
couple
years
ago
must
have
been
three
four
years
ago.
There
was
a
paper
that
came
out
that
showed
that
there
that
their.
D
A
C
D
B
B
B
D
B
That
was
just
a
little
bit
fuzzy.
That
would
mean
that
even
if
they're
tuning
from
the
input
C
is
very
strong,
the
neighboring
column
will
automatically
move
to
something
that
is
reasonably
close
and
that's
the
same
mechanism
as
the
self-organizing
map.
It
will
yield
the
code
that
gradually
changes.
D
A
D
Basket
all
right,
and
so
with
you,
if
you
just
a
basket,
tells
this
thing
ends
up
being
very
different
than
this
gonna
be
or
can't
the
curl
you're
gonna
separate
as
far
as
apart
as
possible,
so
our
spaceship
Euler,
you
give
us
a
set
of
inputs.
The
adjacent
mini-cons
would
not
be
something
right,
but
so
so
now
you're
having
the
second
thing
name
is
what
you
were
saying.
You
had
any
second
thing:
it's
okay,
I've
got
this
one.
D
It's
strong
and
ambition
which
tries
to
separate
these
two
but
you're,
saying
maybe,
unfortunately,
to
have
some
crosstalk
here.
Unfortunately,
I'm,
not
gonna,
let
you
on
one
of
those
because
I'm,
it's
like
I
could
say
well
why
maybe
I
can
just
normally
I
would
say.
Well,
you
know
I'm
trying
to
separate
these
two
any
connections,
this
parameter
darling.
Why
would
it
be?
Remember
them
guys?
You
just
forget
it.
It's
not
very
useful.
It's
photo
staying,
there's
going
to
be
some
excitatory
connections
between
these
mini
columns.
That.
E
D
These
many
columns
be
clear
these
many,
the
general
believes
these
many
columns
are
not
they're
so
skinny
that
they
do
not
represent
different
areas
of
the
venom.
It's
sort
of
like
hyper
column,
represents
one
areas
around
and
within
that
outer
column,
there's
no
differentiation
of
that
patch
of
red
show,
but
no
so
just
be
careful
about
that.
So
it's
like
when
something
is
coming
across
the
red
knob,
it's
not
moving
across
them,
this
massive
column.
It's
it's!
That's
not
the
evidence.
It's
more
like.
E
E
B
A
D
D
D
E
E
D
D
D
D
It
just
some
forces
sparsity
everyone.
This
is
this
and
that's
the
most
common
individual.
They
vary
and
indiscriminate
connections.
All
cells
around
them
just
get
a
bottle,
require
something
or
just
make
sure
the
pulled.
It'll
be
quite
something
I,
don't
even
know
it's
possible
and
we
had
a
visitor
here
kind
of
exported.
D
All
I'm
wrong
requirements
that
if
one
of
these
cells
gets
a
spike
off
a
little
bit
sooner,
that's
it's
in
that
predicted
state.
It
causes
this
guy
to
spike.
This
kind
causes
all
the
cells
the
spot
to
be
inhibited
except
this
guy.
So
it's
like
mine,
I
wasn't
wanted
to
totally
basket
so
to
go
the
best
discounts
gun,
forgive
me
and
and
said
you
can
keep
going
because
your
problem
and
I
said
one
caucus.
You
never.
D
C
D
D
D
A
D
B
B
B
C
B
Where
loss
with
regard
to
close
sort
of
a
father
defining
special,
but
is
it
short,
the
time
consent
form
for
I
mean
there's
two
different
gaba
receptors?
Can
anyone
else
levels
past,
but
that
the
main
point
is
that
the
action
of
these
boxes
is
very
short,
meaning
the
cell
does
get
it
clamped
down
for
five.
Second,
it's
impossible
to
fire,
but
if
you
have
out-of-phase
excitation,
meaning
outside
of
the
five
millisecond
window,
there
will
be
excitatory
spikes
coming
in
by
other
cells
that
have
shared
that
activity.
Long
range,
so
different
type
of
column.
Here
then,.
B
D
B
B
D
B
D
C
D
D
A
B
C
C
A
B
B
An
interesting
thing
is,
of
course,
that
the
overlap
is,
is
big,
doing
this
models,
but
there's
some
big
differences.
Also
I,
just
like
with
regards
for
plasticity,
for
example,
I'm
not
saying
that
that
structure
post
isn't
it
doesn't
exist
in
the
brain
and
I'm,
pretty
sure
that
my
model,
the
models
of
not
son,
lab,
are
and
estimating
it
by
not
incorporating
it,
but
at
the
same
time,
I
also
think
that
the
HTM
model,
for
example,
is
actually
reduced
in
power
quite
a
bit.
D
D
D
B
That
is,
that
is
the
fair
enough
argument
from
the
functional
point
of
view
yeah.
But
if
you're
talking
about
the
biology
I,
think
it's
pretty
clear
that
there
are
rapid
forms
of
plasticity
that
that
do
exist
and
then
do
shape
connections.
In
addition
to
what
is
also
very
true,
the
rapid
spine
turnover-
and
there
is
right.
D
D
So
the
idea
is
that
in
some
places,
especially
like
the
hippocampus,
where
you
want
to
form
these
very
quick
memories
and
my
way
to
campus,
you
find
Travis,
also
three
thousand
synapses
on
and
a
white
portion.
Those
are
silent,
so
they're
like
potential
synapses
that
you
can
turn
on
very
quickly,
some
of
them
it's
it's
still,
not
a
graded
response,
it's
more
of
a
minor
response,
but
it's
like
it's
like
pre-wired
structural
possibilities,.
D
B
The
weird
thing
is
that
they
even
synthesize
receptors.
So
this
these
these
pictures
I've
seen
of
I'm,
not
an
expert
in
that
and
I
feel.
But
if
you
look
at
the
synapse
right
and
you
normally
have
sort
of
AMPA
receptors
in
the
membrane
so
where,
where
the
transmitter
right,
we'll
we'll
open
these
channels
right.
But
what
you
do
then
see
sometimes
is
that
there
are
inserted
into
the
membrane
receptors
but
they're
outside
of
the
synaptic
cleft.
So
so
they're
already.
B
E
B
D
D
B
B
B
And
they
spill
their
contents
like
that,
and
what
the
another
price
was
about.
Was
this
whole
mechanism
by
which
this
thing
gets
recycled
and
that
mechanism
by
which
they
opened
the
brain
close
up
again?
So
it
was
super
important
process
because
it
needs
to
be
super
super
fast,
and
it's
done
all
over
the
brain
all
the
time.
But
you
can
show
that
there
are
in
the
process
of
reinserting
the
use
of
transmitter
like
the
transmitter
reuptake,
so
that
can
we
use
to
put
them
into
new
new.
B
D
E
D
Have
actually
I,
would
you
know
my
guess
from
the
Sun
this
was
somebody
use
that
I
would
argue
the
best
nap
or
the
actions
happening.
You
can
think
about
these
vesicles.
This
is
sort
of
a
distribution
of
random
distribution
and
why'd.
You
release
them
here
then
move
down
to
the
place.
You
know
it's
a
diffusion
process
and-
and
these
guys
are
not
really
very
smart,
because
when
comes
into
this.
A
A
D
B
This
means
is
that
the
pre
synapse
has
some
way
of
measuring
activity
and
there
will
be
a
chemical
as
a
chemical
signal
Kaskade,
where
this
calcium
will
increase
the
probability
for
vesicle
release,
meaning
there
is
a
way
to
affect
the
transmission
at
a
synapse,
not
just
on
the
postsynaptic
side,
but
also
the
presynaptic
side.
But
Jeff
is
still
I.
B
Think
perfectly
right
in
saying
that
the
postsynaptic
side
is
a
bigger
game,
because
when
we
look
at
the
dynamic
change
right
really
most
of
what
happens
is
that
the
synaptic
cleft
gets
like
packed
with
lots
and
lots
of
AMPA
receptors.
And
then
a
couple
of
these
and
at
the
air
receptor
site
in
yeah.
D
Yeah,
so
we
don't
know
I
think
more
amuse
ready.
You
know
we
need
to
understand
what
the
capabilities
are
here
and
some
of
the
we
can
always
say.
Well,
some
of
these
don't
know
yeah.
What's
the
big
picture,
yeah,
that's
a
big
picture.
So
that's
why
we've
chosen
to
do
binary,
synapses
and
and
because
of
this
stochastic
nature
of
these
guys
and
populations,
but
you're
gonna
argue
that
there's
more
to
be
gained
than
we
think
yeah.