►
From YouTube: IETF102-NMRG-20180719-1810
Description
NMRG meeting session at IETF102
2018/07/19 1810
https://datatracker.ietf.org/meeting/102/proceedings/
A
B
C
So
welcome
everybody.
This
is
under
network
management
research
group,
but
this
is
not
exactly
about
network
management
or
not
only
about
network
management.
This
is
actually
a
special
session
on
artificial
intelligence
and
networking,
and
this
is
a
bit
of
a
trial.
Balloon
I
would
say:
I
did
a
heart
of
RFC
in
London
on
what
not
only
what
network
management
could
do
with
AI,
but
mostly
what
not
even
what
AI
could
do
to
networks,
but,
but
also
what
networks
needed
to
do
to
help
artificial
intelligence.
C
So
that's
why
we're
having
this
meeting
so
I
hope
you're
all
here.
For
that
again,
it's
a
bit
of
a
trial
balloon.
So
we
don't
really
know
where
that
is
going
to
go,
but
we
know
that
it's
a
topic
that
has
had
a
lot
of
attention
in
the
past
a
year,
I
would
say-
and
we
felt
it
was
important
to
at
least
have
an
open
discussion
about
it
so
again.
So
why
are?
Why
are
we
here
and
a
cute
little
typo?
Why
are
we
here
again
to
discuss
how
this
could
evolve?
C
We
have
a
press,
two
presentations.
What's
missing
there
we
have
a
net
for
AI
and
then
I,
don't
know
what
the
other
title
is,
but
there's
two
presentation,
so
it's
gonna
take
about
I,
would
say
3540
minutes
and
then
we're
going
to
come
back
and
we
have
a
summary
of
some
of
the
options
that
could
move
us
forward
again
by
looking
at
it
from
different
aspects,
easier
things,
the
low-hanging
fruits
for
this
and
the
things
I
get
to
be
more
complicated
and
again,
while
you're
here.
C
D
D
D
So,
during
the
a
distribution
phase,
the
model
parameters
are
distributed
from
the
parameter
server
to
all
the
workers
and
during
the
aggregation
phase,
the
workers
calculate
the
gradient
and
then
send
them
to
the
parameter
server.
So
the
traffic
model
here
of
the
two
phases
one
to
many
and
many
to
one
respectively.
D
So
we
have
done
some
kind
of
experiment
that
the
shows
that
the
data
transmission
between
the
parameter
server
and
the
workers
takes
the
majority
of
the
time.
So
we
can
see
from
dates
table
they,
for
example,
the
via
obesity
model.
The
calculation
time
takes
about
218
mini
second
yr,
one
single
iteration.
It
takes
around
seven
seven
thousand
millisecond.
So
so
it's
a
majority
of
the
time.
The
data
transmission
accounts
accounts
too
much
so
for
the,
on
the
other
hand,
Scott
Schenker's
team
from
the
UC
Berkeley.
D
They
have
to
also
some
experiment
about
this-
to
add
the
English
world
computing
and
multicast
to
reduce
the
end-to-end
training
time
of
the
obesity
idea
on
Jenny
model,
for
example,
to
129.
So,
theoretically
speaking,
we
when
we
introduced
the
in
network
computing
and
multicast,
we
can
still
use
the
cheap
switch
rather
than
they
much
more
expensive.
D
D
Traditionally,
the
switch
just
work
as
traffic
over
the
traffic,
the
workers
have
note
the
traffic
from
the
workers
to
the
switch
and
then
the
switch
just
deliver
the
traffic
to
the
parameter
server.
So
what
if
we
introduced
the
computing
and
caching
capability
into
network
after
the
image
work
aggregation
and
we
can
transfer
the
results
to
a
parameter
server,
so
we
can
greatly
reduce
the
traffic
from
switch
to
the
parameter
silver
so
which
is
almost
a
seminar
during
the
distribution
phase.
If
we
add
caching
capability,
we
store
model
parameters
and
then
it
distributed
to
the
workers.
D
So
yeah
most
recently,
the
there
are
many
researchers
about
the
in
network
computing,
which
has
missed
some
of
them.
The
most
important
publications,
for
example,
DeMarco
Collini,
proposed
in
that
work,
computers
at
an
idea
whose
time
has
come.
The
experiment
shows
that
a
large
number
of
notch
data
reduction
ratio
and
a
similar
decrease
in
workers
computation
time
and
also
scar
Shankar
seem
they
propose
to
the
change
Network
even
ocean
for
TNS,
so
the
experiments
showing
that
optimizing,
the
network
fabric
can
improve
the
training
time
and
also
someone
from
the
granham
its
structure.
D
They
propose
to
the
scalable
hierarchical
aggregation
protocol,
which
is
also
almost
to
the
same
the
same
effect
effective
way
Chile
in
improved
so
and
also
this
year
in
August,
there
will
be
a
in
network
computing
related
workshop
in
Sikkim,
so
you
might
be
interested
in
that
and
besides
computing
in
network
caching,
and
also
the
consensus
are
also
being
researched
recently
with
some
extra
extraordinary
achievements.
For
example
the
immigrant
caching,
it
can
obtain
the
capability
which
is
similar
to
the
mouth
cast.
D
And
meanwhile,
these
capabilities
are
generic
enough
and
decoupled
from
specific
applications.
Then
we
trial
it
first
today
in
the
constrained
areas,
for
example,
the
data
center
related
data
center
network
and
also
we
we
need
to
keep
in
mind
that
the
existing
program
of
all
network
devices
are
designed
for
forwarding.
So
it's
limited
for
the
X
functions,
for
example,
that
some
of
the
limitations
are
the
size
of
the
cache
table,
the
number
of
bytes
that
the
devices
can
duplicate,
manipulate
on
each
packet
and
also
the
floating
points
etcetera.
E
C
C
Ok,
so
I'm
saying
that
from
your
presentation,
we
understand
that
what
you
are
suggesting
is
that
to
enable
artificial
intelligence
in
the
network,
you
also
need
distributed
computing
and
the
network,
because
you
started
your
presentation
where
artificial
intelligence
and
finished
with
a
a
call,
a
call
for
action
on
distributed
computing.
Is
that
what
you
think
is
going
to
enable
artificial
intelligence.
C
F
F
That
can
be
right
as
long
as
I
mean
the
violation
can
happen
if
someone
some
entity
in
the
middle
decides
without
the
other.
The
two
ends
of
the
of
the
communication
to
make
a
deployment
of
anyway
computational
facility,
but
if
it
is
either
of
the
sides
and
I
would
like
to
see
these
more
in
terms
of
research
on
the
side
of
the
users,
if
I
as
an
user,
instead
of
using
only
this
and
connecting
to
a
cloud
that
is
who
knows
where
I
decide
to
push
my
function
somewhere
in
the
in
the
network.
C
Agree
with
you,
I,
don't
know
if
Rachel
agrees
with
you,
but
I
do
we
can
also
discuss
after
the
second
brick.
We
can
also
discuss
after
the
second
presentation.
Thank
you
very
much,
Rachel
and
I
think
she
needs
to
be
very
thanked
because
she
prepared
this
on
my
request
like
in
two
days.
So
thank
you
very
much.
Thank
you.
G
H
Reusing
so
if
you
happy,
if
you
have
been
in
the
in
the
previous
session
I,
maybe
so,
but
I
will
mention
everything
from
from
different
perspective.
That
means
concentrating
in
what
artificial
intelligence
is
and
means
for
for
the
network,
especially
for
the
network
management,
because
we
have
two
sides.
That
means
that
artificial
intelligence
can
be
used
both
in
the
data
plane
and
in
the
control
and
management
plane.
So
but
I
will
concentrate
in
general
and
how
AI
is
coming
exploited
in
the
whole
network.
H
Not
just
management
or
but
I
will
mention
may
not
mean,
but
it's
not
just
targeted.
Not
so,
but
I
will
first
start
clarifying
science
and
some
concepts
in
general,
many
people
say
Oh
a
I
and
an
L
Honda
Honda
they
are,
they
are,
but
no,
they
are
different.
Ml
is
one
technique
inside
AI
a
has
a
broader
spectrum,
and
many
topics
like
perception,
reasoning
and
planning
are
not
widely
exploited
currently
in
both,
even
in
neither
in
the
network,
in
the
data
plane
or
in
the
management
plane.
H
So
on
the
other
hand,
we
have
to
differentiate
that
intelligence.
It's
not
equal
to
intelligent.
That
means
that
we
can
manage
intelligence
information
in
order
to
to
take
decisions
for
the
data
plane
or
management
plane,
but
that
doesn't
mean
that
that
that
decisions
are
being
taken
intelligently,
that
they
can
be
systematic
based
on
intelligence
information
taken
from
from
from
the
from
the
network.
In
this
sense
intelligence
emphasizes
in
the
gathering
of
that
information,
we
can
use
that
the
intelligence
methods
to
gather
information
and
then
use
systematic
or
AI.
That
is
the
same
method.
H
In
order
to
to
take
decisions
from
that
information.
On
the
other
hand,
intelligent
methods
would
be
we
can
use.
Telemetry
information
is
more
more
systematic
in
order
to
take
decisions.
Okay,
so
why
I
weigh
AI
must
be
in
the
network
in
general
and
in
the
management
plane,
specifically
first
one.
The
key
part
is
that
decisions
in
general
or
in
management,
but
also
in
the
data
plane,
are
more
and
more
complex,
because
this
the
system,
as
are
bigger,
they
are
composed
of
a
Terrigen
use
elements
that
you
have
built
on
machines.
H
We
have
physical
machines,
we
have
elements
that
can
be
moved,
etc
and,
and
all
of
that
must
be
joined
together
to
make
the
system
metaphor
system
work,
so
AI
is
required
for
that.
Also,
the
operational
environments
are
dynamic.
That
means
that
were
just
mentioned
is
not
a
static,
it's
very
complex
and
it
changed
on
time.
H
Moreover,
network
devices
becoming
autonomic.
We
know
we
have
here
the
the
the
animal
group
that
is
working
on
that,
so
they
have
to
take
complex
decisions
that
on
their
place,
so
they
need
some
kind
of
deliberation,
know
and
actually
zero
touch.
Networks
that
we
had
been
posted
recently
are
based
on
that
on
that
sense,
and
probability
is
shaky
on
that
approach.
And,
finally,
as
where
a
means
is
why
not
AI?
H
H
H
Exploiting
AI
into
the
net,
we
can
push
the
concept
of
what
we
can
say.
The
8id
net
is
intelligence
driving
networking
because
it
allows
to
easily
process
intelligence
information
and
take
decisions
to
drive,
know
the
network.
In
this
case,
aim.
Aai
methods
will
have
a
big
impact
in
the
in
the
management
and
control
of
the
network,
and
the
knowledge
needed
from
such
data
can
be
used
to
decide
the
strategic
to
take
staticy
decisions
for
the
evolution
of
that
network.
H
H
Adan
would
be
like
methods
that
used
that
used
to
retrieve
the
information
must
be
assured.
The
quality
of
those
methods
must
be
assured
in
some
sense,
so
we
can
be
sure
that
the
information
we
are
retrieving
for
day
what
I
mean
is
there
for
the
intelligence
for
less
aid
in
the
network
intelligence
plane?
That
information
is
from
good
quality.
Let's
say
the
matrix
and
data
has
been
gathered
following
some
procedure
that
ensures
that
types
and
quality
of
that
information
is
consistent.
A
H
That
information
must
be
formatted
using
some
ontology
that
common
using
common
concepts,
so
different
stakeholders,
meaning
different
elements
of
the
system
and
different
systems,
can
understand
that
information
and,
finally,
the
protocols
used
to
communicate
that
information
must
be
also
adapted,
or
at
least
well-defined,
no
to
be
sure
that
the
information
is
also
to
ensure
the
quality,
but
also
that
the
information
is
transmitted,
know
and
communicated
properly,
and
that's
it.
For
my
from
my
side
question
please
thank.
H
C
I
One
number
one
yeah.
C
H
It's
not
it's
not
yes,
okay,
that
the
formation
is
that
network
built
initiation
enables
Network
automation.
There
is
not
a
requirement
is
that
you
can
automate
things
okay
in
the
in
automate
network
management,
but
you
cannot
automate
the
network
without
something
bit
work,
because
you
cannot
create
a
physical
router,
but
you
can
create
a
bit
order.
Is
why
what
I
mean
in
that
sentence?
Is
that
you
cannot
omit
the
net
or
itself
not
not
just
the
management.
H
You
can
automate
the
management
of
physical
reuters,
physical
elements,
okay,
but
you
cannot
manage
the
network
itself
if
it
is.
If
you
are
not
using
built
well,
some
chai,
you
can
use
some
slicing
okay,
but
this
is
some
technique
of
utilization,
so
you
need
some
kind
of
bitter
engine
technique
in
order
to
be
able
to
automate.
Actually
the
network
know
to
meet
operation
of
the
network,
but
automate
the
network
itself.
C
G
K
G
G
H
H
L
A
G
G
Missing
yours
or
the
intelligence
data
which
piece
you
you
said:
intelligence
data,
so
I,
I'm
wondering
what's
the
meaning
of
intelligence
data
because
thought
era
is,
is
not
a
core.
It
can
be
our
method
to
achieve,
of
course.
So
if
we
study
or
teach
or
make
up
algorithm
model,
then
there
are
is
all
sorts
of
the
piece
of
the
corporation.
So
would
you
want
you
some
intelligence
into
the
data?
No,
no.
What
I
mean
here
is.
H
Intelligence
data
is
sorry
here.
What
I
mean
is
that
is
here?
Is
data
gathered
it
for
that
is
gathered
for
the
specific
purpose
of
knowing
what
to
do
in
the
network?
It's
not
something
like
the
data
is
intelligent
by
itself,
no,
its
intelligence
later
sorry
for
the
then
ones,
but
it
is
that
means
that
this
this
data,
you
are
getting
that
for
the
specific
purpose
of
taking
decisions.
No,
you.
A
H
G
H
H
Do
you
need
you
need
some?
Do
you
need
some?
In
the
end,
you
need
some
ontology
and
some
common
format
to
let's
say
in
code
that
they
tell
that
you
have
gathered.
Of
course
you
need
not,
but
here
I
push
that
the
key.
Is
that
how
you
gather
that
the
data
and
how
do
you
exploit
you
exploit
that
data?
Of
course
we
need
some
format
and
I
push
also
the
format.
I
I
H
H
I
F
I
I
Is
that
why
it
is
obvious
to
me
at
least
or
maybe
not,
the
management
area
ecosystem
changed
substantially
last
period
and
is
moving
towards
managing
the
network
service
is
not
the
connectivity,
so
personally,
I
will
forget
about
the
connectivity
part
of
it
here
and
add
value,
if
possible,
with
some
techniques
in
managing
better
autonomically,
programmatically
and
2n0
touch.
If
possible,
this
new
service
network
interfere
either
per
slice
or
any
other
form,
and
look
for
something
simple
to
be
used
so
for
the
name
of
the
game,
changes
towards
service
management
in
practice.
Yes,.
I
L
N
Stew
card,
so
I
do
networks
by
day
and
I
do
machine,
learning
on
nights
and
weekends
and
I've
been
trying
to
bring
them
together
for
many
years.
So
I
agree
with
much
of
what
you
said.
Ai
is
not
the
same
thing
as
machine
learning.
Automation
is
not
the
same
thing
as
autonomy.
We've
had
all
of
these
pieces
separately
in
our
networks
for
many
years
frequently
with
a
human
in
the
loop
I
think
what's
new
is
where
we're
looking
at
deploying
autonomous
systems
that
continue
to
learn
online
and
so
I'd
like
to
introduce
into
the
discussion.
N
The
word
surprise
my
work
in
machine
learning
mostly
uses
genetic
programming,
I
evolve
code,
and
it
surprises
me
and
frequently.
It
surprises
me
in
ways
that
are
not
conducive
to
my
achieving
the
engineering
solution
that
I
saw.
But
if
it
didn't
surprise
me,
I
would
argue
that
it's
not
terribly
interesting.
Now
in
in
operating
networks,
we
generally
hate
to
be
surprised
by
our
networks,
so
I
guess
what
I
would
like
to
to
throw
out.
H
In
both
in
both
scenarios,
I
would
either
they
don't
surprise
us.
So
everything
is
doing
what
we
want
to
do.
We
will
be
happy,
but
if
they
should
price
us,
what
I
mean
is
that
we
are
in
control
of
that
means.
Intelligence
doesn't
mean
that
on
uncontrolled
things,
no,
you
we
will.
We
will
not
just
put
a
seat
there
unless
hope
that
the
network
or
network
system
or
service
is
working
properly.
No,
no.
We
tablet
some
policies,
rules
etc.
So
all
we're
ecosystem
of
intelligent
methods
is
working
inside
our
parameters.
So
in.
G
G
Okay,
so
I'm
laying
a
message
from
the
mythical
from
Jefferson,
since
you
are
using
terms
from
the
other
area
like
AI
ml,
it
would
be
nice
if
you
could
bring
some
references,
and
the
second
point
is
more
in
the
form
of
a
question.
Would
it
be
useful
to
have
a
document
on
AI
ml
terminology
in
network
management?
Okay,.
H
Good
comment:
maybe
it's
good
to
push
forward
this
technology
and
also
maybe
a
broader
ontology.
In
that
sense,
that
means
not
just
definition
of
concepts,
but
also
the
relation
of
the
those
concepts
that
they
have.
That's
again,
taking
the
words
of
this
commentary.
I,
don't
remember
the
name
it
was
to
define
the
a
I
make
reference
of
where
to
get
proper
information,
how
to
get
the
proper
definition
of
a
IML,
etc.
Okay,.
C
G
O
O
H
G
O
H
But
we
need
some
kind
of
ontology.
We
need
to
define
the
concepts
that
were
I
mean
the
relations
and
how
and
also
the
format
and
how
to
encode
that
we
can't
rely
on
some
intelligent
methods
in
order
that
can't
this,
for
example,
that
in
mind
determine
the
attributes
of
the
information
we
are
gathering,
but
the
resulting
piece
of
data
in
order
to
be
useful
need
to
follow
some
kind
of
ontology,
so
otherwise
it
would
be
difficult
or-
and
you
must
adapt
all
algorithms
to
that
specific
data.
H
O
O
O
H
O
H
G
H
P
P
Yeah
I
think
that'd
be
interesting.
This
you
know
sort
of
how
do
you
it's
the
what
happens
to
kind
of
roll
back
of.
If
you
change
something
at
the
moment
have
quite
a
good
idea.
Oh
we've
made
a
little
change.
It
didn't
work,
we
can
roll
it
back
and
I
kind
of
feel
that
in
this
it's
going
to
be
much
harder
to
do
that
sort
of
thing.
P
So
people
have
their
will
the
will
the
human
to
have
able
to
come
in
and
control
in
some
sense
if
something
goes
wrong
or
you
just
have
to
leave
it
to
the
AI
to
work
out
how
to
get
out
this
problem,
that's
happened
and
those
sort
of
high-level
questions,
I
think,
would
be
very
interesting.
Research
topics
as
I'm,
not
aware
of
of
you,
know
kind
of
work
to
address
that.
Okay,.
G
G
We
expect
what
and
if
something
goes
out
of
this,
it's
usually
dropped
or
discard,
and
if
we
move
towards
more
machine
learning,
especially
in
AI,
used
to
be
different
machine
learning
approaches,
we
have
to
understand
a
bit
more
of
randomness
or
stuck
at
stochastic
aspect
of
the
behavior,
which
is
quite
different
from
the
determinism
of
current
protocol.
So
when
those
things
will
collide-
and
it
will
be
interesting
to
see.
C
So
this
is
something
that
we
discussed
together.
So
there's
there's
a
number
of
options
and
I
agreed
with
the
previous
presentation
that
machine
learning
and
AI
there
are
tools,
they're
not
ends
in
themselves,
so
the
current
low-hanging
fruit.
If
you
look
at
what
is
being
published
right
now
in
a
lot
of
conferences
on
AI
networking,
is
that
you
take
an
AI
algorithm,
you
apply
it
to
a
network
problem
and
you
see
what
happens,
and
you
know
this
again.
A
lot
of
people
have
done
that.
There's
lot
of
results
and
I
think
some
of
it.
C
You
know
for
a
research
group.
It's
interesting.
Nice
makes
nice
paper,
but
there's
probably
a
more
interesting
way
to
look
at
it.
Actually,
a
few
months
ago,
I
listened
to
a
presentation
from
Andrea
Goldschmidt
who
actually
had
developed
an
AI
approach
actually
for
the
physical
layer
of
a
network
which
was
like
a
very
it
was.
It
was
actually
very
refreshing
and
everything
was
that
she
wanted
to
develop
a
network,
specific
AI
or
ml
algorithm
to
solve
a
specific
problem.
C
In
your
case,
it
was
diffusion
of
information
and
then
turn
once
you
have
them,
then
turn
them
around
and
say.
Well
now
we
knew
that
they
worked
in
that
concept.
Let's
turn
them
around
and
see
if
they
can
actually
themselves
now
start
produce
new
paradigms
for
and
and
that
that
could
be
helpful
in
the
ietf.
So
I
think
this
is
more
interesting
because
it
starts
from
the
knowledge
that
people
have
from
the
network.
C
It's
not
some
some
generic
thing
that
we
can
come
out
on
the
web,
I
have
and
I
think
this
is
what
a
little
bit
of
Rachel
was
doing.
I
have
another
approach
and
I
think
the
second
one
is
very
much
appropriate
for
the
people
in
network
management
and
people
who
actually
do
fault
detection
and
things
like
that,
the
other
one
which
is
a
little
bit
what
Rachel
was
was
doing,
and
it
is
some
thoughts
that
are
outside
of
this
room
and
relate
to
all
kinds
of
things
related
also
to
edge.
C
C
The
other
thing
is
this
is
the
thing
that
Laura
and
I
really
believe
in
is
the
same
thing
you
flip
them
around
and
now
by
flipping
them
around.
You
can
say
how
these
algorithms,
in
their
own
way,
can
influence
the
behavior
of
the
network
and
not
just
being
part
of
the
you
know
the
non
forwarding
functions.
C
Obviously,
in
the
case
of
a
distributed
intelligence
and
computing,
the
decentralized
approach
is
more,
maybe
more
appropriate,
but
they
all
share
into
a
few
things.
They
need
CPU,
they
need
energy,
and
then
you
get
into
sustainability
issues.
If
you
just
start
distributing,
how
do
you
make
sure
that
they're,
federated
and
synchronized
and
again
that's
what
what
is
needed
inside
the
networks?
You
support
that
of
course
performance.
C
If
we're
going
to
do
that,
and
it's
not
more
performing
than
what
we're
doing
right
now,
it
makes
nice
papers,
but
it's
not
the
tools
that
the
first
presenter
was
doing.
So
we
would
like
to
open
the
discussion
we
do
not
have
and
just
to
start,
we
do
not
have
a
specific
way
forward.
We're
very
much
interested
to
know
what
the
community
thinks
about
again.
If
we
take
a
I
as
a
tool,
it
may
not
have
and
again
decentralizing
it.
It
may
not
have
a
role
as
a
specific
research
group.
C
It
may
have
a
role
in
a
number
of
research
groups.
Obviously
there's
a
lot
of
applications
for
it.
There
could
be
also
future
work,
that's
more
into
a
single,
a
single
research
group.
So
we
want
to
open
the
discussion
to
people
who
are
here
and
to
see
what
people
are
thinking.
So
now
we
have
about
twelve
minutes.
B
G
L
Extreme,
so
just
as
one
comment
I
think
in
terms
of
what
I
would
find
interesting
is
when
one
hand
busy
there
are
the
AI
applications
and
the
network
management
are
genes
of
outsider
standard
I
think
all
those
applications
are
actually
fairly
straightforward.
Basically,
the
applications
about
there
are
not
so
many
surprises
there,
but
I
think
for
research
group
is
actually
more
interesting
is
to
apply
to
really
actually
networking
problems,
so
not
specific
in
that
network
management,
but
naturally
problem
so,
for
instance,
routing
decision
as
well
can
can.
L
How
can
a
I
helped
with
that
all
the
work
I
see
on
the
on
the
on
the
controllers,
path,
computation?
That
means
a
etc,
etc
so
I
think
trying
to
apply
or
see.
Basically
what
role
I
can
play
to
make
networking
better
is
I,
think
actually,
the
more
interesting
question
or
the
question
I
would
find
more
interesting
as
a
research
group
than
looking
at
the
management
problems.
So
it's
just
from
virtual,
but
above
it.
F
F
Normally
we
tend
to
think
about
the
AI
as
a
helper,
the
tool,
but
also
can
be
always.
You
can
use
a
hammer
to
nail
down
to
network
to
put
a
nail
on
the
the
wall
and
you
can
use
a
tool
tool.
You
can
use
a
hammer
to
key
someone
and
say
I
can
be
an
extremely
complicated
tool
if
it's
used
by
wrong
hands.
F
A
F
F
Neuromancer-
and
this
is
something
that
we
should
I
mean
not
talking
about
this
exactly
but
about
an
environment
in
which
you
can
somehow
find
ways
of
reacting
to
potential
attacks
that
are
enabled
by
AI.
They
are
supported
by
a
race.
This
is
one
thing,
and
the
second
thing
is
some
mechanisms
for
doing
some
kind
of
attestation
or
security
assessments
over
the
a
behavior
is
something
that
we
should
consider
as
well.
Before
I
mean
let's
put
security
as
I
had
in
the
design
and
the
thinking
process.
P
Philibert
lien
I
mean
AI
is
driven
by
data,
isn't
it
so
if
my
network
is
being
run
by
the
way
way,
its
operating
is
being
driven
by
AI
decisions
as
an
attack,
all
I've
got
to
do
is
work
out
how
to
get
data
in
the
manipulates.
My
AI
to
get
my
network
in
some
statements,
a
disaster
so
I
think
I
mean
in
terms
of
this
I.
Think
the
first
one
is
fair
enough
and
will
continue
to
happen,
but
in
terms
of
the
others,
I
think
it'd
be
interesting.
C
We
want
to
have
access
to
every
instantiation
of
every
sensor,
and
that
creates
these
terabytes
of
information
and
data
that
have
to
be
moved
across
the
network,
and
that's
why
this
is
actually
where
the
bottom
ones
are
about.
A
lot
is
one:
what
is
needed
in
the
network
to
support
that
incredible
amount
of
data
which
is
essentially
coming
to
the
edge,
and
how
do
we
need
to
send
it
completely
back
to
the
through
the
cloud
for
all
the
processing,
or
could
there
be
some
intermediate
processing
that
would
reduce
the
load
on
the
hops?
C
This
is
actually
why
these
two
other
elements
are
there
is
that
when
we
think
about
intelligent
elements
and
data,
we
I
don't
think
we
actually
I
was
shocking
from
and
they
have
the
same
thing
with
surveillance
cameras
in
smart
cities.
So
I
think
there's
this
idea
that
what
what
also
can
these
algorithms
can
help
us
do
to
to
reduce
the
size
and
the
sheer
overwhelming
you
know,
quantity
of
information.
M
But
mom
woman,
so
it's
interesting
because
I
had
a
comment
bolt
regarding
the
surprises
and
how
much
data
we
have
to
process
there's
a
couple
of
things
that
come
to
my
mind
is
regarding
surprises:
is
how
much
control
do
we
have?
What
is
the
privacy,
for
example?
How
are
we
going
to
maybe
reduce
the
amount
of
data,
so
we
reduce
the
amount
of
surprises
because
we
buy
feeling
so
much
data
we
might
I.
Try
come
up
with
answers
that
we
didn't
even
want
to
ask
the
question.
O
Juwan
INRIA,
so,
first
of
all,
just
emphasize
what
I've
been
selling
security
and
privacy,
so
I
think
we
should
take
the
opportunity
of
you
speak
closer
to
the
mic.
You
should
take
the
opportunity
of
AI
in
order
to
actually
collect
less.
That
I
think
we
are
usually
talking
about
collecting
a
lot
of
data
with
the
AI,
which
is
not
I,
mean
always
necessary,
so
maybe
from
a
user
perspective.
I
would
be
happy
if
busy
yeah.
We
can
at
the
end
phonetically
this
an
opportunity
to
take
Segond
regarding
security.
O
Yes,
I
think
what
which
is
our
topic.
Ii
and
I
of
the
adversity
are
learning
in
machine
learning.
Let's
say
address
our
learning,
so
if
you
put
AI
for
doing
some
operation,
you
increase
incredibly,
your
a
let's
say,
attack
surface
particular
I,
really
like
the
idea
of
putting
some
something
to
the
networking
network
computation.
K
K
Regarding
the
last
thing,
one
thing
is
the
the
objective
the
network
operating
line
may
have,
and
that's
that
the
network
runs
and
I
think
you
may
have
the
same,
and
then
you
may
have
conflicting
policies,
but
I
wouldn't
use
intently
there
in
that.
In
that
specific
I
would
use
simpler
words
like
objectives
and
policies.
E
Leon
from
China,
Mobile
and
I
have
a
general
comments
on
this
topic,
so
I
think
artificial
intelligence
is
a
great
topic
to
be
tackled
in
IETF,
especially
you
know
for
for
for
network
use
applications,
but
in
terms
of
skill
but
I
eat.
A
variety
of
here
I
think
it's
much
more
interesting
to
to
tackle
the
problem
of
network
for
artificial
intelligence
instead
of
artificial
intelligence
for
network.
You
know
and
well,
I
think
it's
more
important,
for
example,
as
an
operator.
E
E
You
know
it's,
it's
it's
just
useless,
it's
there,
so
it's
very
important
to
tackle
the
problem,
how
we
can
collect
the
data
from
from
the
network
more
efficiently
and
to
filter
out
the
data
to
be
the
input
of
your
AI
algorithm
and
to
get
the
output
you
want.
For
example,
if
we
have
a
long
massive
collection
of
bandwidth,
I
mean
live
bandwidth
data
from
the
real
network.
E
Would
that
give
you
a
prediction
of
how
you
configure
different
domain
of
network
to
to
have
to
give
you
a
out
spirit
of,
for
example,
for
low
latency
network
I
mean
we
have
some
research
project
with
vendors
to
tackle
this,
like,
like
converting
the
dimension
of
how
you,
using
some
type
of
algorithm,
to
converting
the
dimension
of
from
bandwidth
to
latency,
using
some
existing
data
yeah?
So
that's
what
I
mean.
C
Thank
you,
Alex,
we're
out
of
time.
Okay,
so
I,
don't
I,
think
just
one
an
hour
one
hour
for
this
was
obviously
not
enough
and
I,
don't
think
we
came
up
with
a
I
think
we
still
have
these
two
things:
AI
for
network
and
network
for
AI
and
I
think
it
follows
pretty
well
with
the
hi
and
I
had
discuss
that.
Maybe
there
could
be
two
different
types
of
approaches
to
this
so
AI
as
an
application
in
the
network
and
the
network
being
a
feeder
for
the
AI,
and
we
can
continue
the
discussion.
G
Yeah
and
to
build
on
that
I
would
say
if,
if
you're
interested,
if
you
think
we
need
to
keep
having
meetings,
discussion
or
try
to
organize
and
structure
this,
those
questions-
I,
don't
I
ate
here
for
IOT
F
scopes.
Okay,
we
are
still
here
I'm
here
also
tomorrow,
so
come
to
us,
I
mean,
if
not
only
to
us,
I
mean
anybody
can
discuss,
but
we
have
to
understand
you
can
use
the
meaning
list
or
if
we
need
to
prepare
something
for
the
next
meeting
in
between
the
meeting.
G
But
if,
if
you
want
to
go
beyond
is
just
one,
our
discussion
and
all
the
previous
attempts,
we
need
to
understand
a
bit
more
to
define
some
directions,
or
maybe
some
areas
where
we
say
we
want
to
give
priority
to
this,
because
there
is
five
ten
person
that
wants
to
invest
a
bit
of
time
on
that.
So
just
but
we
don't.
We
go
to
step
beyond
and.