►
From YouTube: Discussion of Active Dendrites
B
B
B
But
it
didn't
assume
that
was
delay.
So
my
respect
too
much
for
something.
That's
a
question:
where
are
these
ubiquitous
predictions
recurrent
in
the
cortex?
And
one
of
the
reasons
you
have
a
separate
integrations
arms?
Is
that
the
fatal
dendrites
you
don't
want
to
how
to
sell
fire-eaters?
What
I
have
is
all
the
colors?
It's
right
a
shoe.
So
if
even
if
you
could
put
all
the
synapses
on
one
day
trading,
some
you
don't
serve
the
purpose
and.
B
A
B
B
B
D
B
E
B
B
E
To
you
know,
just
if
you
just
extend
you
know,
I
want
to
know
I
want.
I
would
like
to
understand
if
the
the
number
of
units
that
you
need
in
a
single
representation
is
somehow
linear
to
the
number
of
patterns
that
you
want
to
recognize.
I
understand,
they
will
be
biological.
You
know,
constraints
is
but
just
mathematically
it
just
that
you
want
to
grow.
C
C
A
D
C
C
A
C
E
B
B
C
B
B
B
B
D
B
D
In
fact,
many
of
these
staining
techniques-
you
know
they
work
well
because
they
don't
actually
stay
in
every
neuron,
a
stain
every
one,
thousands
neuron
and
then,
when
you
do,
that
actually
looks
really
nice,
these
nice
beautiful
hours,
but
of
course
in
reality,
it's
a
pact
to
ship,
there's
no
free
space
right
there,
there's
no
empty
holds
between
the
neurons,
it's
all
packed
with
invites
and
axons
and
keel
cells-
and
you
know
fibers
are
passed
through
or
so
that's
important
to
keep
in
mind.
It's
really
a
struggle.
B
D
Do
this
thing
where
they
like
slice,
the
quarter
sheet
with,
like
you
know,
mind,
meter
thin
slices
and
then
they
use
computer
vision
technology
to
apply
to
track.
You
know
if
there's
an
axon
going
through
the
sheet
by
a
so
there.
You
have
all
these
slices
and
you
try
to
track.
You
know
how
it
goes
and
where
branches
enter
time
to
go,
to
build
these
creations.
C
So
one
question:
the
first
question
was
you
know,
yeah
and
I'm,
not
gonna
draw
arrows
coming
into
it.
It's
you
know.
The
first
question
is
well.
You
know
why
not
just
have
only
one
dead
right,
one,
okay,
so
the
answer
one
answer
there
that
Jeff
said
is
that
you
mean
at
least
a
second
dendrite,
because
these
ones
will
do
prediction,
and
these
ones
will
do
you
know
activations.
C
And
this
one
has
an
impact
on
time,
T
plus
1,
on
this
one
impact
right
on
time,
all
right.
So
this
the
function
of
this
on
the
cell
is
quite
different
than
the
function
of
this.
So
just
for
that,
you
need
at
least
two
and
then
you
can
ask
okay.
So
this
is
detecting
some
patterns
that
form
the
prediction.
You
know
why
don't
you
just
put
all
the
patterns
here
that
cause
a
prediction
and
treat
that
as
a
union,
so
their.
A
C
B
A
C
Really
a
union
limitation
issue:
you
could
quantify
this
exactly
so
because
of
that,
because
you
can't
put
too
many
patterns,
you
know
we
end
up
having
lots
of
dendrites
and
because
each
one
can
recognize
a
couple
of
having
one
or
more
patterns
and
they
can
be
completely
segregated.
You
run
run
it
to
the
Union
limitation,
so
you
can
do
hundreds
of
patterns,
contextual
product.
So
that's
the
second
benefit,
this
capacity
thing
and
I.
Think
there's
a
third
one,
which
is
a
learning
benefit.
C
So
if
you
recognize
a
pattern
and
it's
either
correct
or
incorrect,
you
want
to
update
just
the
synapses
for
that
pattern.
If
everything
was
all
Union
together,
even
if
it
was
within
the
capacity,
it
would
be
really
hard
to
update
just
that
pattern,
because
there's
no
real
knowledge
of
it
anymore.
What
is
that
pattern?
But
if
it's
segregated
like
this,
you
could
do
brain
specific
learning
on
these
things,
so
that
you
update
the
synapses
or
ways
only
for
that
pattern
and
I.
Think
that
makes
learning
much
more
efficient.
B
C
I'm,
just
sort
of
answering
from
an
abstract
machine
learning
point
of
you
ignoring
biological
considerations
for
a
second,
but
this
is
obviously
all
these
inspirations
come
from
biology.
This
is
completely
and
then
from
a
continuous
learning
standpoint.
We
know
that
if
you
have
very
sparse
representations
that
the
chance
of
a
single
pattern
activating
falsely
some
of
these
other
one
who's
very
small.
So
now
you
can
learn
new
things
on
new
dendritic
segments
without
any
interference.
C
E
C
That's
totally
true,
so
you
know
you
could
have
in
our
temporal
memory.
We
have
multiple
cells
per
column,
each
one
has,
let's
say
three
different
active
dendrites
or
you
could
have
three
times
as
many
cells
with
one
active
done
right
each
and
that
would
be
fine.
It's
just
know
it
gets
a
question
of
resources
and.
D
B
D
See
that's
what
I'm
wandering
out
there's
actually
back
my
question
so
whether
for
the
model,
it
would
be
helpful
and
understanding
that
it
is
actually
specially
divergent
that
it's
a
tree,
that's
branching
out
and
I
don't
mean
to
introduce
multiple
levels
on
the
dendrite.
That's
not
what
I
mean,
but
what
I
mean
is
that
and
we
draw
them
here
in
parallel,
which
kind
of
you
would
suggest
that
you
could
have
like
a
you
know,
a
descending
fiber
that
would
activate
all
of
them
in
practice
right.
C
B
Looks
big
compared
to
the
cell
body,
but
it's
not
very
big
overall
I'm
trying
to
think
about.
You
know
it
can
be
maybe
I'm
tempted
to
put
a
number
I'm,
not
sure,
maybe
half
an
elevator
or
something
like
that
and
that
change
everything
is
going
quickly.
We
believe,
all
over
the
place
since
I'm
not
sure.
C
B
D
B
Haven't
had
much
problem
of
interference
or
it
really
it's
just
not.
It's
never
really
been
a
big
issue,
only
guests
when
you,
when
we
train,
like
the
sequence
memory
and
we
trained
it,
is
until
it
started
to
fail
and
you
can
apps
without
there.
You
know
it
wasn't
like
even
close
to
being
a
critical
parameter.
Yeah
I
think
it's
worth
point
out
a
couple
of
distinctions
here.
This
is
a
biological
distinction
and
may
not
be
important
from
just
just
to
reiterate.
There
are
no
excitatory
synapses
on
the
soba.
B
No
right
there
are
inhibitory
synapses
in
somebody,
so
I'm
like
a
point,
neuron
and
they're
all
networked
or
so
they're
all
being
something
somebody
that's
not
going
to
happen
here.
There's
a
there
are
there's
basically
within
some
distance
of
the
cell
body,
the
proximal
synapse
is
they
are
close
enough
that
they
act
as
if
they're
somehing
at
the
cell
body,
and
so
there
can
be
depending
how
many
basic
branches
coming
off
the
soma.
You
know
there's
some
number
of
synapses
here,
a
nice.
B
B
Then
of
course,
then
all
the
others
are
further
away
and
I'm
actually
not
sure
how
long
how
distance
it
is
to
get
your
first
branch
point
I,
don't
know
if
there's
something
here
that
don't
qualify
as
proximal,
but
clearly
after
the
first
branch
point
the
synapses
are
all
too
far
away.
These
are
all
these
sort
of
prediction.
Synapses.
The
other
thing
I
want
to
point
out
is
that
these
branches
on
a
real
neuron
this
the
length
along
these
turnarounds
are
here,
can
be
200
microns
and
that
that's
a
lot
longer
than
the
integrations
on
earth.
B
40
microns.
So
with
the
die
that
I
get
the
integration
zones,
you
have
to
have
some
number
of
within
40
microns
of
distance.
You
have
to
have
some
number
of
synapses
15
to
20
85
or
some
like
that.
That
are
activated
some
very
short
period
of
time
that
they
integrate
together
to
produce
a
dendritic
spine.
But
you
stretch
those
out
over
200
microns.
It
wouldn't
work.
So
already
what
that
tells
you
is
it's
not
as
simple
as
we
draw
it
here.
We
just
keep
adding
segments.
A
B
You
know
20
synopsis
for
each,
so
we
don't
really
again.
None
of
that
seems
to
matter
at
the
moment,
and
even
the
thing
that
subitize
is
pointing
out
is
if
I'm
going
to
mix
and
match
them
on
here
on
these
200
microns
they're
not
going
to
be
mixed
and
matched
over
the
whole
distance,
because
that's
more
than
40
microns,
so
a
natural
100
micron.
B
You
know
you
can
have
to
have
some
pattern
right,
guys
on
the
floor
to
hear
another
panorama
can
either
maybe
got
40
here,
I
kind
of
thing,
so
you
know
the
pattern
over
here.
The
time
over
here
not
really
combined
together,
they're
not
really
competing
with
one
another,
and
somehow
the
learning
signal
comes
back.
We
don't
know
if
it
affects
all
these
synapses
or,
if
there's
still
a
local
effect
from
this
little
area
that
was
depolarized,
it
generated
is
possible
to
dis
generated
appendix
pi.
B
Can
that
affect
their
metabolic
effect
or
something
that
that
this
gets
you
trained
on.
We
don't
really
know
that,
but
I
just
want
to
point
out
that
these
dendrites,
the
eternal
dendrites,
tend
to
be
long
and
the
other
going
back
to
what
I
mentioned
ago.
It's
related
to
it.
The
tips
of
these
dendrites
are
constantly
growing
and
retracting.
B
It's
as
if,
as
if
they're
just
trying
to
find
something,
you
know
see
if
I
make
the
connection,
but
if
they
do
make
a
connection,
if
they
find
something
else,
is
there
useful?
Then
they
stop
that
and
I.
Don't
know
why
you
know
I
guess,
presumably
they
don't
keep
going
forever.
The
climbing
things
connected
also
maybe
they're
just
the
signaling
efficacy
Coast
now
ready
yeah,
there's
always
the
step-back
action
potential
and
the
forward
with
the
ice
there.
There
might
be
a
physical
limit
to
the
distance
here,
but
all
the
time.
B
B
C
B
B
D
D
Though
the
way
that
they
use
it,
so
in
that
sense,
it's
kind
of
fascinating
to
me
that,
even
though
it's
like
more
machine
learning
or
it
is
actually
capturing
more
biological
detail
in
some
aspects-
well,
obviously
abstracting
it
very
far
with
you
know,
sort
of
like
ideal
life.
You
know
like
segments
and
bothering
with
all
the
biophysical.
D
D
Complex
machine
as
well
to
sort
of
evolution
like
how
many
different
proteins
do
you
find
at
a
synapse,
and
if
you
look
at
sort
of
the
earliest
forms
of
mines
that
have
some
kind
of
nervous
system,
you
might
find
some
2030
different
proteins
and
the
human
sign
outside
is
at
its
peak.
There's
like,
like
estimates,
range
a
little
bit,
but
thousands
of
different
proteins
in
in
the
small
confines
of
you
know.
D
B
E
B
In
one
of
these
train
is
that
you're,
essentially
you're
always
coming
all
the
inputs
that
are
on
and
therefore
to
get
the
know
and
respond
differentially
under
different
conditions.
You
have
to
very
fine
to
these
ways,
based
on
the
the
grave
specific
activations
of
the
individual
input,
yellow
veins,
you're.
B
C
C
E
C
This
huge
matrices,
so
that's
one
one
answer
one
possible
thing:
that's
important!
The
other
thing
is
I,
don't
know,
there's
a
benefit
in
in
theory.
You
could
create
more
efficient
representations
here.
Then
you
could
here
because
if
multiple
things
essentially
mean
the
same
thing,
you
could
group
them
here,
whereas
it's
difficult
to
do
that
here.
You
know
you
have
to
learn,
and
we
talked
about
this
a
little
bit
before.
E
C
D
So
essentially,
what
just
kicked
in
my
head
and
what
is
interesting
about
these
dendrites
is
exact.
Don't
make
it
necessary
for
the
learning
to
go
down
all
the
way
it
doesn't
need
to
suffice.
The
whole
network
criteria.
It
doesn't
need
to
be
retrained
on
all
days
if
you
wanted
to
marry
that
to
deep
learning
network
architectures
recognize
their
first
apostles
positive.
D
So
you
still
can
back
propagate
the
error,
but
only
if
the
dendritic
segment
actually
took
pilot.
If
it
didn't
apart,
you
stopped
propagating
the
Arabic
word
that
makes
training
cheaper
and
it
allows
exactly
for
the
this
wheel
property
that
not
every
neuron
takes
part
in
in
you
know:
sufficing
the
constraints
of
all
the
components
of
you,
train,
I,
think
to
some
extent,
I
would
actually
just
magically
happen.
If
you
set
up
this
network
when.
A
You
set
it
like
all
the
logic
with
maxing
and
knitting
and
stuff,
and
those
gates
already
have
all
that
magic
built
in
to
where.
If
something
gets,
maxed
out,
something
is
clamped.
My
name
stops
right
there.
So
I
think
that
would
be.
What
would
what
someone
would
accidentally
could,
as
you
just
provided
a
justification
part.
E
Right
yeah,
the
problem
is
that
you
know
they're,
not
sparse,
so
you
don't
tell
me
he's
probably
in
for
so
much
it's
just
if
you
consider
normal
evasions,
it's
60%
of
these
needles
of
acting.
So
it's
a
large
population.
It's
exactly
of
the
same
time
he's
you
know.
The
part
of
this
globalization
problem
is
all
the
ways
are
important
so
itself
he
still
to.
A
E
Mean
it's
all
about
enemies,
entangle
parameters
and
representations
as
well.
So,
of
course,
if
you
have
a
very
sparse,
you
know
sees
them
in
terms
of
parameters
and
in
terms
of
activations,
you
can
work
much
better,
so
I.
This
was
also
another
requester.
So
if
for
you
for
example,
what
do
you
think
that
sparsely
is
is
enough
for
coordinating
it?
We
don't
take
advantage
of
the
multiple
active
dendrites
and
the
thing
that
is
just
about
you
know
having
more
neurons
that
are
sparse
and
produce.
D
C
B
B
E
B
B
B
C
E
D
And
are
thinking
where
there
would
be
more
helpful
for
talking
to
the
machine
learning
community
to
think
of
this
act.
Look
we're
gonna
introduce
like
the
second
class
of
unit
here,
which
we
call
a
dendritic
unit,
and
then
mainly
they
would
understand
him,
because
then
you
know
how
to
talk
about
the
biology
of
it.
You
can
just
talk
about
the
fact
that
these
units
are
bit
special
on
the
sense
that
they
have
these
converging
inputs
onto
these
other
classes
of
notes.
I.
E
A
E
C
C
C
E
D
It's
mono
clergy
treats
has
like
the
zoo
of
classical
neural
networks
in
there,
and
so
they
show
by
different
colors
like
green
hidden.
You
know,
yellow
our
input
units
and
red
our
output
units,
and
it's
like
this
big
memory,
cell,
recurrent
cells
and
memory
cells,
different
memory
cells,
matched
input
output
cells,
like
you
know,
so,
there's
like
different
different
classes
of
notes
when
RNA
is
sort
of
like
talk
about
these
different
architectures
and
sort
of
these
terms,
and
they
just
put
them
together
and
like
these
nice
color
diagrams
and
one.
D
This
in
machine
learning
from
context
this
exact
yourself
well,
there's
one
more
categorical,
know:
DM
right,
just
not
a
you
know,
which
we
call
it
like
a
dendritic
cell
and
there's.
Some
special
condition
has
to
be
layered
with,
like
these
other,
you
know
main
cells,
and
then
you
could
build
an
architecture
rather
than
machine
learning.
Person
would
understand
without
having
to
talk
about.
D
A
A
Like
we
have
this
kind
of
fairly
complicated
the
four
layer
network
that
we're
like
I
mean
it's
like
a
big
form
of
four
layer:
CN
n
there's
some
T
cells,
I'm,
leaving
out
here
and
we're
what
they
were
optimizing
for.
Is
they
want
to
be
able
to
to
classify
objects
from
this
top
layer
and
in
even
be
able
to
classify
novel
objects
without
ever
having
to
change?
A
That's
not
right,
and
the
interesting
observation
was
that
by
training
a
network
on
that
purpose,
they
got
something
that
where
the
top
two
layers
like
were
that
they're
representations
could
be
linear
map
onto
representations
between
these
things.
I'll
just
do
like
dotted
lines.
You
could
do
linear
regression
from
here
to
here
to
predict
units
firing
and
now
the
one
one
thing
that
I
wasn't
able
to
respond
well
to
intelligently.
To
was
that
was
that
that
this
is
self-evident,
that
this
would
happen.
A
The
idea
that,
if
you
can
classify
you
something
linearly
from
here
and
from
here
that
almost
implies
that
there
must
be
a
linear
about
thing
between
the
two
yeah
I
think
that
was
the
crux
of
the
issue.
Yes,
that
true
or
not
so
I
think
I
can
so
I
think
over
here
I
can
describe
that
statement
in
a
way
that
makes
it
obviously
false.
A
So,
first
of
all,
like
I,
think
part
of
the
confusion
came
from.
We
were
talking
about
both
linear,
classifiers
and
linear
regression
and
I.
Think
the
conversation
got
a
little
confused
from
that.
So
keep
in
mind
that
a
linear
classifier,
so
network
output.
This
is
this-
is
either
this
or
this
they
call
it
X
a
linear,
classifier
comes
from
applying
weights,
you
get
a
set
of
scores,
but
then
you
like
remove
most
the
information
from
this.
You
do
a
maximum
for
it.
A
B
D
A
B
A
Yeah
so
here
I'm
just
trying
to
show
picture
two
different
network
outputs,
one
of
them.
This
is
a
schematic
picture.
You
could
think
of
it
as
two
output
units
or
you
could
think
about
as
a
schematic
picture
for
many
units
and
here's
the
wave
matrix
combined
with
weights
and
the
bias
turn
causes
this
kind
of
decision
boundary
and
from
and
basically.
A
Here
here
is
like
I
call
it
a
decision
boundary,
but
really
it's
it's
a
scoring
function.
It
assigns
scores
like
you
could.
You
might
actually
have
a
triangle
over
here,
but
anyway
the
point
is
it's
a
scoring
function
and
now
I've
shown
particular
examples:
three
distances
of
class
to
go
three
instances
of
class
one
and
I've
shown
how
two
different
networks
might
put
them
in
different
places
and
I've
intentionally
drawn
them
in
different
orders
or
where
a
b
c
b,
a
c
e
DF
d
EF
to
show
just
like
it's
pretty
obvious.
A
You
couldn't
do
a
linear
mapping
from
this
to
this.
You
can
multiply
some
matrix
by
the
student
than
be
here,
because
you
can't
read
it.
You
can't
change
the
order
of
things
and
anyway,
I
think
that
this
was
worse.
There
was
some
confusion
here
today,
so
it's
at
least
from
the
perspective
of
linear,
algebra,
linear
mappings
of
matrix
multiplication,
whatever
it's
not
the
case
that
this
is
self-evident.
B
C
A
Now
I
think
maybe
some
of
your
intuition
came
from
something
that
might
be
true
and
in
the
paper
kind
of
points,
this
out
that
this
might
be
evidence
that
this
statistics
of
the
world
are
such
that
there
really
aren't
very
many
bases
where
this
is
possible.
There
are
that,
maybe
maybe
the
statistics
of
the
world
are
such
that.
A
Know
if
that's
true
or
not
again
and
I,
think
my
my
Gus
here
is
if
an
author
on
the
paper
were
standing
here
right
now,
they
would
have
the
perfect
answers
to
that
question.
They
would
say
we
intentionally
only
sampled,
a
small
number
of
units
making
sure
we
solved
for
that
case,
but
I'm
not
going
to
be
able
to
give
a
good
answer.
A
And
so
yeah,
so
ok,
the
closing
point,
and
that
part
of
this
is
is
it
is
there
are
two
possibilities,
but
two
possibilities.
One
is
that
there
really
aren't
very
many
of
these
bases
where
this
is
possible
and
it's
pretty
cool
that
the
deep
networks
are
able
to
find
it.
Our
brain
finds
it
too.
It
seems
you
could
think.
A
Maybe
one
way
to
map
this
I'm
not
talking,
maybe,
for
example,
the
equivalent
of
the
spatial
cooler
and
and
for
a
several
cars,
a
90
bat-spaceship,
that
spatial
cooler
has
a
basis
of
sorts,
and
maybe
it's
actually
quite
similar
to
this
basis
and
you'd
be
the
right
way
to
think
of
it.
That's
just
that's
me
trying
right
now
to
map
it
onto
our
world.
A
C
C
A
So
I
guess
when
I
say
the
same
basis,
I'm
really
thinking
about
her
being
of
Basie's
living
in
these
kind
of
families
or
like
this
family
of
bases,
aren't
lynnie
are
kind
of
related
to
each
other.
You
can
kind
of
linearly
not
between
them.
This
other
family
over
here
and
I'm,
saying
but
they're
all
and
they're
in
yeah
and.
C
A
The
other
possibility
I
was
going
to
bring
up
as
a
possibility
one.
There
aren't
very
many
solutions
to
this
where
you
can
linearly
separate
them.
Possibility
to.
There
are
a
lot
of
them,
but
it
just
so
happens
that
maybe
convolutional
neural
networks
learning
rules
are
sufficiently
similar
to
the
brains,
learning
girls
that
they
find
similar
ones.
A
C
A
C
C
A
There's
multiple
levels
of
training
here,
so
the
conversation
can
get
a
little
confused
for
one
when
they
were
doing
the
back
propagation
over
here
they
were
using
a
totally
different
set
of
images.
It's
so
totally
different
set
of
image
classes
yeah
later
when
they
were
doing
this
testing,
they
were
using
a
new
set
of
image,
classes
and
and
training.
Yes,
a
linear,
classifier.
Sorry
I
need
to
need
to
pause
for
a
second.
Yes,
they
were,
they
were
doing
linear
regression
from
units
here
to
units
here.
Was
there
a
trace?
C
C
B
A
You
can
kind
of
decompose
this
into
two
parts.
It's
like
you
want
the
basis
to
continue
keeping
keeping
everything
linearly
separable
for
the
old
images,
and
you
probably
have
to
update
your
classifiers
for
the
for
the
old
classes
and
I.
Don't
know
that
just
provides
a
point
of
view
that
it
really
matches
I
know
like
and
what
the
trends
have
presented
has
like
the
feature
layers
and
then
at
other
ones.
So
it's
like
it
goes
in
line
with
that
I
guess.
I'm.
A
This
is
kind
of
a
mental
model
of
of
what
deep
networks
are
doing
when
we
talk
about
our
compartmentalized
dendrites,
that's
like
adding
memory
to
the
system,
but
this
is
more
of
like
a
generalization
system.
This
is
more
like
a
memory
system
and-
and
it's
not
just
gonna-
imagine
I
it's
worth
experimenting
with,
but
it's
it's
our
words
not
cut
out
for
us.
It's
not
just
gonna
magically,
there's
some
cleverness
as
needed
to
sell
some.
C
A
E
You
know
somehow
you
can
maybe
get
rid
of
the
words
that
and
then
eating
them.
You
can
perceive
a
lot
of
the
connections
and
their
relationship
receivers
is
in
under
representation.
That's
right
and-
and
you
can
somehow
have
these
nice
property-
or
you
know
long
time
to
back
propagate
everything.
Repair
stops
very
different
points.
You
can
preserve
a
lot
these
weights
yeah,
it's
not
even
that's
cool
but-
and
we
don't
apply
here.
C
E
Yeah
without
doing
anything,
much
more
than
sparse,
okay,
I
was
trying
to
think
about.
There
is
a
memory,
a
comic
efficiency
issue
yeah
and
then,
but
they
still,
we
need
a
university
to
preserve
the
weights.
At
the
same
time,
a
bit
more
like
two
consolidations
raising
or
something
like
that.
As
you
know,
sometimes
maybe
a
urine
can
fart
detecting
a
feature
that
is
useful
for
recognizing
an
old,
don't
cost
as
well
as
a
new
class,
and
then
what
that
means
is
that
for
something
up
to
layers.
E
E
C
E
Know
later
say
that
you
know
please
neurons
can
activate,
and
so,
for
example,
this
is
a
beautiful
to
predict
a
senior
class
as
well
into
this
previous
class
that
this
is
gonna
be,
and
this
one
it's
going
to
be
decreased.
If
we
don't
see
if
this
is
if
these
cuts
right
across
trees,
maybe
these
feature
is
very
important
to
that
particular
year.
C
C
Yes,
the
sparsity
will
reduce
the
chance
of
that.
They
all
still
happen
and
maybe
with
the
elastic
consolidation
or
something
like
that,
and
it's
not,
but
the
other
thing
that
we
do
is
you
know
our
our
negative
error
in
the
HTM
is
often
in
this
case
of
where
you
have
a
misprediction,
we
think
about
in
temporal
memory.
C
The
negative
error
is
much
much
smaller
than
the
positive
a
amounts
up
yet
so
the
decrease,
so
you
see
I
mean
partly
for
this
reason
you
have,
as
you
can
make
in
the
case
of
temporal
memory,
you
can
make
a
lot
of
predictions.
Only
one
of
them
will
actually
happen
doesn't
mean
the
others
were
wrong
all
right.
So
what
may
be
the
translation
here?
I've,
no
idea
this
will
work
is
in
the
case
where
it
correctly
predicts
something
the
increase
in
the
word
yeah.
C
A
But
that's
something
that
came
to
mind:
I
want
to
try
one
more
time
to
articulate
it
up
so
far,
I
guess
what
feels
strange
to
me
with
this
okay
here,
I'm
I'll,
try
to
say
back
what
word
what
we're
trying
to
do
with
these.
With
these
separate
dendrites,
we
are
making
it
so
embed.
Subsequent
training
data
doesn't
stomp
all
over
the
weights,
doesn't
stop
all
over
old
weights.
A
We
want
them
to
be
kind
of
guarded
from
from
later
training
data,
but
there's
this
conflicting
interest
that
we
want
the
network
to
always
be
searching
for
this
really
good
basis,
and
we
wanted
to
continue
using
a
future
data
to
change
its
basis
and
those
two
are
at
odds
with
each
other.
Those
two
things
see
my
top:
preserving
weights
versus
having
your
bases
get
better
and
better.
As
you
know,
as
you
observed
on
internet,
they
seem
there
there's
a
contradiction.
There,
there's
something
yeah,
there's
a
dissonance.
C
Maybe
one
day
I
can
say
the
the
temporal
memory
sighted
on
one
extreme,
we're
literally
memorizing
everything,
yeah
and
pure
against
backdrop.
Is
that
another
extreme,
where
you
don't
care
at
all
about
the
previous
table
every
morning?
Stuff
is
trashing
everything
and
maybe
there's
a
sweet
spot
there.
Where
you
know
you
still
have
some
of
that
problem
of
trashing
all
stuff
that
you
really
introduced
it,
but.