►
From YouTube: IETF101-T2TRG-20180322-1550
Description
T2TRG meeting session at IETF101
2018/03/22 1550
https://datatracker.ietf.org/meeting/101/proceedings/
A
B
B
D
B
Then
Matt
Acosta
will
report
from
the
hackathon
that
we
had
on
the
engine
from
the
Russia
approach.
That
was
underlining
that
then
John
will
talk
about
something
that
sooner
or
later
we
all
will
have
to
do
for
some
machine
learning
and
while
he
will
not
talk
so
much
about
how
that
is
going
to
influence
our
work,
I
think
we
all
can
start
imagining
that.
B
B
Be
centralized
and
violence
that
earlier
version
of
which
was
at
the
MDS
s
this
workshop
and
we
thought
was
interesting
enough
first
target
and
at
the
end,
don't
know
the
way.
Before
the
end,
we
will
have
a
quick
look
at
meeting
planning
in
this
case.
Mostly
look
at
the
meeting
will
have
to
with
w3c
and
ocf
okay,
so
just
to
remind
people
what
the
research
group
is
about.
Who
here
is
new
to
the
research
group
has
not
been
to
a
previous
meeting.
Okay,
so
let
me
quickly
say
why
we
are
here.
B
B
So
we
are
looking
at
issues
in
this
area,
in
particular
when
there
are
opportunities
for
IETF,
standardization
or
I
should
add
for
cooperation
with
other
SDSU's,
for
others
with
other
standards,
developing
organizations
to
make
them
happen.
So
we
don't
usually
do
radios
here.
We
sometimes
have
to
look
at
the
radios
if
they
behave
strangely.
Here
will
slide
on
that
in
a
minute,
but
we
not
only
started
the
IP
annotation
layer
and
end
in
the
hearts
and
minds
of
end-users.
So
this.
B
B
The
IOT
security
considerations
draft,
which
entered
iesg
IRS
G
last
call
we
had
a
workshop
and
mbss
on
decentralized
duty
security
at
standards.
I
quickly
reported
in
nature.
We
had
an
interesting
side
meeting
so
essentially
we
you
know
all
these
sounds
about
how
many
people
fit
into
a
phone
box.
Of
course,
these
don't
work
anywhere
because
they
are
for
marcus
any
law,
so
we
just
usually
use
the
meeting
room
for
the
same
thing
and
had
a
side
meeting
with
29
people
in
the
room
Waterloo,
which
is
about
the
size
of
the
school.
B
B
B
F
B
Out
early
security
and
the
other
that
IT
security
often
will
not
be
based
on
centralized
approaches,
but
will
require
decentralized
or
non-centralized
approaches.
So
we
had
12
papers
which
are
still
in
the
process
of
being
published,
so
it
will
take
until
about
mid
April
until
you
can
download
them.
B
D
B
About
so
that
would
be
later
in
this
meeting.
So
what
are
we
going
to
do
next
tomorrow?
We
will
be
in
Prague
with
OC,
f
and
w3c.
Where
are
things
so?
Oc
f
is
the
open
connectivity
foundation.
The
organisation
that
resulted
from
the
merger
of
oh
I
see
you
PNP
at
the
ancien
Alliance
and
they
are
doing
a
very
interesting
standardization
work
at
their
heavy
user
of
IDF
technologies.
So
it's
a
good
thing
to
talk
to
them
and
we
will
talk
you
particular
about
requirements.
They.
B
B
What
we
understand
interesting-
and
we
are-
they
are
also
interesting-
local
people
there
that
we
may
want
to
contact.
We
suddenly
will
do
more.
We
share
related
things
there.
There
has
been
some
discussions
about
distributed
this
country,
which
won't
follow
a
little
bit.
The
theme
that
we
had
this
workshop,
but
this.
G
Photo
from
Nokia
so
in
terms
of
these
joint
meetings,
it
may
be
interesting
also
to
have
a
joint
call
with
a
joint
sessions
with
we
may
we
have
a
July.
We
have
a
cos
fest
in
Co,
which
is
done
by
TTA.
It's
an
organization
standardized
a
test
standardization
organization
in
Korea.
It
may
be
interesting
to
showcase
some
of
this
and
also
participate
in
the
light,
a
table
to
amicus
fest
there
that
maybe
I.
B
B
G
B
But
yeah
definitely
is
one
of
the
organizations
that
we
have
been
cooperating
with
an
IETF.
So,
for
instance,
one
of
the
collab
flood
tests
was
hosted
by
IO
a
night
which
was
really
useful
because
in
the
process
we
learn
much
more
about
the
Iranian
trip
specification,
which
was
pretty
young
at
the
time.
So
we
certainly
make
sense
to
deepen
that.
B
B
So
the
other
document
that
is
nearing
completion
is
called
restful
design
for
IOT.
So
this
essentially
takes
the
when
your
architectural
style,
that
is
underlying
the
HTTP
protocol
and
much
of
the
web
work
and
explaining
how
to
use
this
by
Internet
of
Things
applications,
because
there
are
some
things
that
that
may
just
not
be
very
intuitive.
B
H
Okay,
so,
as
Carsten
said,
we
we
tried
to
fit
30
some
people
in
a
room
to
mentioned
for
about
10
or
12
other
than
that
the
meeting
went
very
well.
There
is
a
draft
about
some
of
these
issues,
but
the
main
idea
is
that
when
we
put
administrative,
Lea,
independent
IOT
applications,
so
something
I
put
my
apartment,
something
you
deploy
in
your
apartment,
no
administrative
relationship
between
them
with
these
packets
because
of
radio
interference,
and
that's
not
the
IETF
problem.
H
But
what
some
recent
research
results
suggest
is
that
the
protocols
are
inducing
timing,
behaviors
sensitivity
to
loss
of
certain
kinds
of
it's,
that
ours
asked
that
exacerbate
these
problems
and
we
can
get
kind
of
severe
performance
degradation
at
the
protocol
level
and
potentially.
That
is
something
that
that
we
need
to
think
about
it.
This
layer,
right
now
there
don't
seem
to
be
good
tools,
use
simulators,
test
beds
that
help
us
evaluate
these
sort
of
multi
Network
scenarios
or
if
independent
networks
sharing
the
channel.
H
Just
some
of
the
things
we
discussed,
miss
Ted,
just
potentially
a
number
of
IETF
protocols.
Anything
that's
touching
these
Mac
layers,
six,
five,
six
th,
some
other
things
that
came
at
the
discussion.
This
is
obviously
somewhere
on
the
IETF
I
Triple
E
border.
Clearly,
the
Mac
has
to
do
most
of
the
work
here,
but
it
administrative
ly,
independent
networks.
The
Mac
only
has
limited
abilities
to
isolate,
isolate
the
networks
in
the
link
sense.
H
So
I
think
there
are
some
rules
for
for
IETF
protocols
to
to
try
to
understand
the
performance
in
this
environment
and
potentially
try
to
make
some
improvements.
Some
folks
also
raise
various
possibilities
for
active
or
explicit
coordination
via
high-level
protocols.
So
there
is
a
draft
and
I
encourage
people
to
be
interested
in
this.
I
Go
get
Montenegro
a
yeah
I
was
wondering
if
you
could
elaborate
a
bit
more
on
this
research
because,
for
example,
in
Europe
the
ICM
bans
require
listen
before
talk,
not
that's
not
required
in
US
and
many
other
in
some
other
places,
but
most
that's
why
3gpp
with
court
and
coexistence
problems
in
in
bands
that
are
typically
used
by
Wi-Fi,
assumed,
LBT
and
built
on
top
of
that,
so
my
question
is:
is
this
research
that
you're
pointing
at?
Doesn't
it
US?
I
B
I
J
Juan
Carlos,
when
you
guys
see
cooks,
so
yeah
just
responding
to
to
Gabrielle,
because
I
guess
discussion,
but
you're
right
all
those
points
where
we're
definitely
part
of
the
discussion.
There's
multiple
regulatory
domains.
If
we
focus
let's
say
you're
nice
and
bans,
there's
different
rules
on
there
country
on
how
you
operate
on
these
IES
and
bands.
Not
all
these
protocols
follow
the
same
Mack
for
instance.
So
here
we
talk
about
the
Triple
E,
that's
just
part
of
it.
Some
some
protocols
may
not
be
following
the
same
as
long
as
they
comply
to
the
regulation.
J
They
are
legal
to
operate
in
this
country
that
they
may
not
be
legal
to
operate
in
this
other
country.
However,
I
think
at
least
a
take
that
I
that
I
got
from
from
the
discussion
is:
there's
still
some
potential
to
share
some
knowledge
at
higher
layers
and
and
that's
why,
with
there
there's
a
pointer
there
to
their
white
spaces
protocol,
that
was
the
past
protocol.
That
was
a
fine
before
where
yeah
I
mean
through
the
internet.
J
You
can
get
some
knowledge
that
could
be
useful
for
your
area
and
maybe
this
bans
the
the
sorry
this
operators
are
using
or
either
private
networks,
public
networks,
standardized
protocols
or
from
different
entities,
but
still
there
could
be
some
interesting
information
like
I,
don't
know
to
me.
It
sounds
like
level
of
noise
in
this
region,
something
that
doesn't
touch
on
privacy
necessarily,
but
some
some
useful
information
that
could
help
me
operate
better.
H
The
document
really
kind
of
forces
on
focuses
that
kind
of
laying
out
the
challenges
systematically
it
points
to
a
couple
of
research
results,
both
mine
and
other
people's.
The
main
point
in
the
documents
is
that
I
recommend
and
that
the
TTT
RG
or
some
other
mechanism
we
try
to
understand
what
we
would
like
to
have
for
some
practices
for
evaluating
how
the
ietf
protocols
you
know
six
low,
60s
role
perform
in
this
environment.
We
need
to
have
a
better
understanding.
That's
really
what
the
document
says.
Apologize.
C
For
following
up
even
before
we
make
recommendations
about
protocols
right,
the
I
think
the
first
step
is,
we
might
make
sense.
I
agree
with
you
only
understanding,
as
part
of
getting
the
understanding
of
avista
useful,
to
have
some
sort
of
data
collection,
sauce
analysis
methodology,
so
that
we
can
better
characterize
and.
H
K
H
H
B
L
E
Milk,
media
interoperability
and
we've
actually
had
eight
calls
since
the
ATF
101,
because
we've
had
a
lot
of
activity
in
planning
this.
Our
participation
in
the
hackathon
and
we've
really
covered
a
diverse
set
of
topics
around
the
semantics
and
I
per
media.
I'm,
probably
not
going
to
try
to
read
through
everything
here,
but
essentially
we
looked
at
sort
of
top
to
bottom.
So
what
is
what
is
the
nature
of
abstract
semantic
time?
We
want
to
describe
something
you
couldn't
wait
list
protocol
neutral.
How
do
we
do
semantic.
E
Do
we
connect
with
RDF
ontology
x'
and
other
data
datasets
that
are
better
around
like
these
third-party
vocabularies
yeah?
So
you
know
the
qu
DT
as
though
I
say
SSM
that
are
fairly
hard
to
use
and
how
do
we
make
them
easier
to
use?
How
do
we?
How
do
we
integrate
symmetric
metadata
into
the
system?
Where
does
it
go
and
you
know
who
uses
it?
How
do
they
use
it?
This
idea
of
a
layer
semantic
stack.
There
may
be
different
semantics
at
different
layers
and
the
stack
at
different
protocol
levels.
E
E
The
goals
for
this
this
one
and
we're
probably
going
to
have
a
series
of
these,
but
the
goals
for
this
is
the
second
hackathon,
but
this
is
the
first
hands-on
where
we
bought
hardware
and
tried
to
make
it
do
things,
bringing
they
diverse
things
together
to
interoperate,
and
we
wanted
to
start
focusing
on.
You
know
where
you
would
start
in
an
application,
workflow
of
doing
discovery
and
synaptic
based
discovery
and
software
adaptation
to
different
data
models
and
different
protocols,
and
that's
that's
generally
the
theme
for
this
one.
E
We
started
with
really
simple
application
scenarios
like
can
a
signal
input
like
a
motion
sensor
turn
the
light
on
and
off
just
for
very,
very
basic
orchestration
and
the
implementations
that
people
brought
were
w3c
web
of
things
which
we'll
talk
more
about.
There
was
a
yang
coma.
There
was
a
coli
based
implementation
tried
to
interoperate.
There
was
a
om,
a
lightweight,
m2m,
client
and
server,
and
there
was
some
various
ad
hoc
device.
Api's
protocols,
just
basic
rest.
E
Api,
like
you,
might
see
from
various
connected
devices
that
don't
follow
any
particular
standard,
but
they
still
use
HTTP
and
they
still
have
a
a
rest,
ish
sort
of
API.
So
we
looked
at
connected
home
and
automotive
domains
and
the
idealist
you
know
how
do
we
make
them
work
together?
What's
the
technology
that
we
use
and
we
started
using
some
of
the
stuff
from
the
w3c
web
of
Things
group,
because
a
lot
of
the
goals
and
objectives
are
very
similar,
so
we
brought
that
that
technology
and
it
will
talk
more
about
that.
E
We
we,
we
started
integrating
komai
with
the
descriptions
so
to
show
that
you
don't
really
have
to
have
any
particular
kind
of
protocol
of
what
what's
the
range
of
protocols
that
can
be
connected
to.
We,
we
brought
a
thin
directory
and
we
stored
thing
descriptions
in
the
thing
directory.
I'll
show
more
about
what
that
that
whole
thing
looked
like,
and
then
we
had
a
at
the
end.
We
had
the
web
things
implementation,
communicating
with
these
different
device,
implementations
and
different
protocols.
E
So
the
thing
description
was
was
what
we
used
as
the
metadata
common
metadata
format.
Here,
it's
basically
a
media
type
with
RDF.
It
describes
the
abstract
and
interactions
with
things,
so
you
know,
read
the
temperature
or
lock
the
door,
this
sort
of
thing,
and
it
also
binds
those
abstract
interactions
to
concrete
instances
of
things
that
implement
those
interactions.
So
you
know
things
like
the
data
shape,
payload
structure,
data
types
and
transfer
layer.
E
So
here's
a
little
picture
that
illustrates
this
description
has
both
semantic
annotation,
high-level
information
models
and
capabilities
like
what
do
you
want
to
do
in
an
abstract
sense
and
protocol
bindings
sort
of
from
the
device
specific
problem?
How
do
you
do
it?
You
know
OCF
lightweight.if,
so
dot.
These
are
all
based
on
coop,
but
they
used
collab
quite
a
bit
differently.
There's
not
an
easy
way
to
to
use
them
all
in
a
single
application
without
doing
some
adaptation,
and
so
that's
what
we
do
with
thing
description
is
that's
up
for
me.
E
A
format
for
adaptation
between
these
different
protocols,
so
thing
directory
is
what
we
use
in
the
and
hackathon
as
a
central
registry
of
these
thing
descriptions.
So
basically,
each
each
device
or
connected
thing
registers
its
thing
description
in
the
thing
directory
and
then
applications
discover
them
and
find
them
using
semantic
queries
and
thing
directory
uses
the
same
protocol
as
a
core
resource
directory.
E
E
Small
applications
discover
them
after
that.
The
application
might
interact
directly
with
the
connected
thing
or
it
might
be
interacting
through
some
intermediary,
like
a
proxy
or
a
pub
sub
broker,
and
if
it
was
a
pub
sub
broker.
There
there's
some
coordination
here
to
register
in
thing
directory
the
address
of
the
pub
sub
broker
and
how
to
communicate
with
it,
as
opposed
to
the
thing
itself,
and
some
of
these,
like
my
entries,
for
example,
and
some
others
registered
both.
E
E
So
here's
some
examples,
sort
of
showing
that
the
same
format,
how
the
how
the
gangue
implementation
worked.
It
that
it's
stuff
to
a
thing
directory
and
then
the
Serbian,
which
is
where
the
application
runs,
can
then
access
the
device
same
thing
with
an
HTTP
device
that
might
have
some
ad
hoc
protocol
buttressed
fish
registers,
the
TD
and
then
the
application
discover
from
the
TV
and
interact
with
the
device.
M
N
But
with
which
one
so
I'm
trying
to
make
sense
out
of
this
time,
I
understand
that
that
the
thing
directory
is
the
directory
that
you
previously
explained,
which
one
corresponds
now
to
the
application
sort
of
like
and
to
the
connected
thing
and
the
intermediaries
on
the
next
slide
so
like
here.
So
so,
I
see
the
directory
finger
rectory,
which
one
is
now
the
client
that
wants
to
access
some
data
from
the
IOT
device
is
the
comma
I
device.
Is
that
the
IOT
device
and
and
the
application
is
the
w3c
wrap
of
things
Serbian?
E
This
is
acting
as
a
proxy
for
the
device.
The
Serbian
is
essentially
an
application
proxy
for
the
device,
so
the
Serbian
interprets
the
the
transfer
layer
instructions
from
the
thing
description
that
it
got
from
the
thing
directory
and
it
uses
those
to
access
the
device.
So
in
the
w3c
Serbian
software
we
already
have
the
adaptation
or
the
protocol
for
the
for
the
transfer
layer.
So
we
use
that
as
as
the
application
proxy
I'm.
N
Sorry
I
think
you
have
to
update
the
slides
to
get
them
in
line
with
the
architecture
picture,
because
there's
one
entity
than
apparently
missing
and
some
errors
may
be
missing
as
well,
because
in
a
previous
life
the
IOT
device
registered
itself
at
the
directory
directly.
In
this
case
it
doesn't
and
there's
no
application.
So.
N
E
Here's
a
better
picture
to
show
a
lot
of
these
things.
So,
for
example,
here
the
lightweight
m2m
device
going
through
the
lightweight
m2m
server
and
this
discovery
adapter
will
discover
things
on
the
lightweight
m2m
server
and
register
them
with
the
thing
directory.
So
that
shows
the
whole
loop
that
wasn't
shown
on
the
other
side.
So
unfortunately,
so
like
I,
take
it
a
good
point.
Well,
we
should
fix
those
up
to
make
maybe
put
something
like
this
a
little
more
upfront
so
in
the
same
way
for
the
HT.
E
So
here
it
shows
the
device
isn't
registering
this.
The
device
is
doing
lightweight
m2m
to
the
lightweight
MTM
server.
It
only
needs
to
have
these
tiny
short,
URIs
and
numeric
identifiers,
and
all
of
that
then
this
this
piece
goes
in
and
expands
them
and
knows
how
to
create
all
the
semantic
annotation
based
on
its
knowledge
of
you,
know:
lightweight
m2m
number
codes
and,
if
so,
smart
objects
and
what
have
you
and
then
and
then
the
Serbian
can
can
say.
Oh
I'm,
looking
for
a
temperature
sensor
and
can
find
this
and
then
from
the
transfer
layer.
E
Instructions
knows
how
they
interact
with
the
device
and
the
same
thing
for
the
HTTP
and
comai
devices,
but
they
didn't
show
the
whole
loop
of
so
the
HTTP
device
was
a
Raspberry
Pi.
It
could
just
it
can.
It
can
basically
go
directly
and
register
it's
it's
things
with
the
thing
directory.
Also,
we
had
some
cloud
services
that
had
we're
running
on
a
digital
ocean
instance
that
had
enough
you
know,
CPU
and
enough
resources
to
be
able
to
go
and
both
expose
the
devices
in
and
create
the
semantic
descriptors
and
register
them.
E
So
next
steps
we're
here
we
go,
want
to
go,
look
at
semantic
annotation
and
discovery
like
the
thing
directory
case,
but
using
core
RT
and
link
format.
So
how
can
we,
semantically
annotate
web
links
with
the
same
information
to
enable
the
same
sort
of
discovery
on
a
general
web
linking
format?
We
want
to
look
at
different
end
device
protocols
and
data
models,
so
you
know
please
like
back
that
even
are
part
of
this
right,
more
automation
of
the
semantic
queries
because
we're
mapping.
This
are
:
point
to
you
our
eyes.
N
Mentally,
when
you
did
the
mapping
from
let's
say,
like
a
team
term,
two-finger
scription
from
the
w3c
or
other
like
kamae
to
that,
how
was
was
that
possible
without
losing
some
of
the
semantics
or
like,
because
that
has
always
been
a
challenge
like
we
have
all
these
different
activities
and
we
are
piling
on
more
and
more
standards
activities
on
different
descriptions
in
then
it
becomes
very
difficult
to
map
them
to
each
other
without
losing
some
of
the
semantics
of
it.
So
how
was
that?
What
was
the
experience?
E
When
we're
starting
we're
reusing
these
devices
that
are
biomes
and
temperature,
sensors
and
stuff,
there's
not
much
semantics,
there's
not
any
synaptic
box.
In
fact,
the
the
semantic
annotations
pretty
much
have
there
have
a
way
of
describing
all
of
the
what
you
want
to
do:
sort
of
sort
of
things
and
that's
all
extensible.
N
N
One
of
the
challenge,
or
one
of
the
observation
that
we
learned
from
the
IOC
workshop
was
you
can't
just
look
at
the
data
model,
but
you
also
have
to
look
at
the
interaction
model
because
they
are
closely
covered
together
somehow,
and
so
it
would
be
interesting
to
go
from
the
basic
example
to
something
a
little
bit
more
sophisticated
actuators
may
be
other
more
complicated
things
and
see
how
well
that
works.
I
would
be
curious
in
what
the
result
is.
P
Matthias
Kovich
siemens
talked
about
Michael
a
bit
here,
so
they're
different
tracks
going
on
so
in
in
the
debrief.
You
see
level
things.
We
look
exactly
at
these
interaction
models
that
have
partly
describing
what
protocols
are
used.
What
are
basically
the
abstract
yeah
operations
that
you
perform
there.
We
found
kind
of
this
narrow
waste
of
properties,
actions
and
events
that
cover
from
a
programming
model
perspective.
Basically
what
you
can
do,
and
so
we
can
read
out
some
states.
We
can
write
to
some
state
if
this
is
allowed
in
this
particular
ecosystem.
P
We
can
call
DC's
actions
that
often
correspond
to
something
that
goes
on
for
certain
time.
Some
others
might
be
RPC
based.
So
there's
also
this
coverage
and
we
can
cover
whenever
a
device
can,
as
synchronously
send
something
to
the
client.
So
there
we
have
this
disco
model
in
the
finger
description.
So
there's
a
formal
model
behind
the
description
document
itself,
and
this
is
then
extended
a
bit
by
the
binding
templates
that
Michael
was
mentioning.
P
So
in
this
example
here
the
common
protocol
was
co-opted,
so
we
are
able
to
describe
the
methods
that
can
be
used
by
coop.
We
can
describe
if
there
some
specific
options
that
I
use
for
the
protocol,
and
this
is
basically
a
noted
down
uniformly
in
the
finger
scription.
So
we
can
describe
the
differences
there.
What.
N
I'm
most
interested
in
is
reading
up
on
some
of
that
experience
on
when
you
did
that
I
am
like.
There
are
always
some
new
standards
efforts,
I,
don't
I,
don't
necessarily
care
about
them.
I
want
to
learn
like
did
those
things
that
you
try
to
accomplish
actually
work
out
like.
Where
are
the
problems,
and
maybe
maybe
you've
traded
those
descriptions
already?
Yes,.
P
P
Okay,
this
part
explicitly
about
that,
would
be
probably
a
report
from
this
hackathon,
which
was
just
the
last
weekend
and
I.
Don't
know
how
much
time
you
had,
but
definitely
this
is
something
we
also
working
on.
So
there
is
another
clock
fest
happening
at
the
debrief
you
see
face
to
face
at
the
upcoming
weekend.
We
also
again
people
from
different
organizations
come.
There
will
be
some
reporting
on
that
and
the
last
thing
I
just
wanted
to
add,
because
you
asked
about
a
semantic
level.
P
So
this
is
something
that
goes
on
top,
and
that
is
something
that
we're
currently
collecting.
So
it's
looking
into
ok,
but
in
kind
of
semantic
interactions,
do
we
have
in
the
different
ecosystems
and
actually
listing
this
down
and
creating
vocabulary,
but
there
we
need
basically
yeah
I
could
overview
of
what
is
out
there.
Maybe
activity
to
point
to
is
the
IOT
schema.org,
we're
kind
of
the
the
ones
that
are
easy
to
identify
because
they
occur
across
all
these
ecosystems
are
being
collected
again
in
some
kind
of
uniform
format
that
can
apply
over
these
different
ecosystems.
O
Are
kind
of
the
floor
so
quickly
comment
on
there,
like
with
enzyme
experience,
it
was
very
simple
thing
that
they
did
so
far,
so
for
sure
we're
gonna
be
learning
more
on
that.
How
well
does
it
my
pics
map,
so
right
now,
with
the
prototype
did,
is
that
it
curious
from
the
management
server
what
kind
devices
are
registered
there?
One
of
the
you
are
eyes
that
are
exposed
and
it
maps
them
to
think
thing
that
thing
description
well,
the
next
step
is
to
do
semantic
mapping
and
then
figure
out
like
how
much
tools
how's
vocabulary.
C
So
these
are
like
really
good
next
steps.
I
really
appreciate.
You
know
where
you're
going,
but
before
you
get
to
the
more
complex
devices,
I
think
it
might
be
useful,
also
to
expand
out
your
with
a
more
complex
model
of
the
existing
devices,
that
is
to
say,
Hennis,
has
worked
with
ace
and
other
things.
C
So
one
of
the
challenges
that
when
I
talked
to
my
partner's,
you
know
this
by
the
way
I'm
Elliott
I
work
at
Cisco
the
moment
we
add
in
an
authorization
model
it
gets
even
with
the
simplest
devices
it
gets
pretty
complex
and
I
would
I
think
that's
a
great
area
for
discovery.
I
realize
I'm,
saying
this
without
saying:
I'm
gonna.
Do
it
okay,
but
it's
just
we're
going
through
this
now
for
very
simple
cases
and
they're
turning
very
complex.
You
know
you've
been
trying
to
get
something
to
market,
it's
very
hard,
so.
E
That's
that's
a
really
good
point
and
what
we're
working
on
a
w3c
and
probably
need
to
look
at
here,
also
we're
looking
at
how
do
we
create
security
models
for
individual
endpoints,
so
you
can
authorize
and
learn
how
to
securely
communicate
with
them
and
we
sort
of
seeing
that
as
a
next
step,
but
not
far
behind
so
yeah
you're
using
access
control
entries.
How
do
you
generate
the
correct
token
subject,
role
for
our
back?
All
of
those
things
we're
trying
to
work
that
into
the
next
next
phase
of
design.
For
you
for
that,
my.
M
E
M
E
M
E
Okay,
so
basically
that's
these
all
really
been
covered.
The
one
thing
I
wanted
to
say
that
we
find
we're
looking
at
more
diverse
models,
we're
looking
at
an
automotive
model
and
the
automotive
model
is
very
dependent
on
being
able
to
understand
the
feature
of
interest
that
the
thing
is
bound
to.
So,
if
I
have
a
door
switch,
the
automotive
model
really
requires
us
to
say
which
door
that
is
and
that
it's
a
door
and
and
that's
something
that
we
don't
have
in
the
IMT
schema
on
top
definition
model.
E
D
Turn
your
screen
with
us.
We
can
see
that
hello,
yeah,
so
I
need
something
completely
different,
going
to
talk
about
a
research
project
that
is
being
developed
in
my
team
right
now.
That
allows
us
to
do
deep
learning
on
microcontrollers,
which
I
think
is
one
of
the
most
exciting
products
that
we
can
be
doing
right
now.
My.
D
So
actually
we
went
to
do
fields
sitacles
the
conference
or
six
days.
So
during
that
time,
with
its
IT
training
like
how
do
you
build
devices?
How
do
you?
Where
do
you
store
the
data?
What
do
you
do?
It
is,
and
after
we
actually
went
into
the
fields,
so
we
went
to
dairy
farms
or
with
the
chicken
farms.
Actually
capturing
data
that
was
published
in
open
data
sets
that
a
variety
of
universities
then
used.
D
So
a
university
in
Kenya
was
actually
using
the
samples
that
we
generated
during
that
during
that
workshop
and
returning
in
May
very
happy
about
it
full
day
of
science.
Africa
we're
gonna
do
eight
days
a
hundred
students,
three
field
project,
three
fieldwork
projects
and
again
just
actually
building
something
and
getting
this
technology
in
hand
with
students.
Yes,
very
cool.
D
But
if
we're
taking
one
step
back
to
actually
a
lot
of
the
value
of
machine
learning
is
added
if
we
can
run
all
these
algorithms,
not
in
a
big
data
center,
but
rather
on
very
small
edge
notes.
So
one
of
the
cool
projects
that
we've
seen
in
the
last
year
is
a
sensor
fusion,
where
you
combine
a
number
of
relatively
cheap
sensors
and
get
data
points
from
all
the
sensors
and
see
if
these
these
points
are
not
worth
that
much.
D
You
can't
infer
much
from
it,
but
if
you
combine
a
lot
of
cheap
sensors
together,
you
create
like
this
was
the
guys
at
gee
rats
called
a
super
sensor,
so
there's
a
there's,
a
really
good
video
on
that
link.
The
slides
were
published
after
is
where
they
all
kinds
of
stuff
that
happens
in
the
house
like
stuff
like
a
faucet,
turns
on.
They
can
detect
from
this
very
very
small
group
of
sensors
and
that's
all
just
because
of
training
on
a
variety
of
different
sensors
at
the
same
time
in
the
school.
D
Another
thing
is:
what
happens,
what
what
Google
is
doing
with
their
keyboards
on
androids
a
technique
called
federated
learning?
Now,
there's
a
there's,
a
local
model
running
on
the
Android
phones
that
gets
locally
trained
as
well,
but
what
you
don't
want
to
do
is
send
raw
training.
Data
is
preferably
if
I'm,
if
I'm
a
keyboard
manufacturer
I,
want
to
get
data
from
all
my
users
and
get
that
back
and
refine.
D
My
global
model
then
push
out
to
all
the
Android
phones,
but
don't
want
to
push
any
raw
data
into
a
cloud
because
there
might
be
all
kinds
of
privacy
aspects
in
there
and
Plus.
You
know
from
living
on
how
often
I
type
it
costs
quite
a
bit
of
down
plates
or
just
stream
it
upwards.
So
people
is
doing.
Is
they
have
a
local
model?
They
keep
refining
that
and
they're
sending
in
the
changes
in
the
weights
in
the
model
back
to
cloud,
and
they
use
it
from
all
the
people
to
refine
their
global
model.
D
So
it's
a
hybrid
model
way
to
do
some
training
locally
and
I.
Take
the
learnings
that
you
do
in
that
model
and
push
it
upwards
and
uptake
your
global
model
that
runs
somewhere
on
the
cluster
of
GPUs.
Another
thing
is,
for
a
bunch
of
file
formats
were
really
good
compression.
If
I
draw
a
number
I
draw
a
letter,
then
I
don't
need
to
send
the
whole
image.
I
can
just
say:
okay.
D
Well,
that's
an
eight
and
I
just
send
one
byte
over
line,
but
if
we're
going
to
look
at
larger
data
sets
or
interesting
stuff
that
happens
all
around
us,
then
I
don't
have
a
model
that
can
compress
the
data
straight
away.
So
just
think
we'll
order
encoding,
where
you
basically
have
a
deep
learning
model.
D
That's
going
to
keep
producing
features
up
until
the
moment
that
it
has
its
own
compression
for
Matheny
had
the
model
on
the
outer
sides,
or
it
can
decompress
that
and
basically
give
you
a
learned
representation
of
what
you
try
to
send
it.
This
is
a
model
where
the
deep
learning
model
itself
is
going
to
figure
out
an
encoding
scheme
for
this
training.
Very
very
interesting
compression
schemes
that
we
didn't
device
before
for
spare
very
localized
sets
of
data.
D
It's
like
a
flying
self-contained
systems.
What,
if
you
want
to
do
something
in
an
area
where
there
is
no
internet,
if
I'm
going
to
deploy
a
machine
learning
model
on
a
dairy
farm
so
where
it
comes
in
Nia,
where
the
average
income
is
under
$100
a
month
to
me,
there's
no
internet
I'm
not
going
to
put
down
a
cluster
of
GPUs
extra.
Do
some
deep
learning
some
deep
learning
models
there,
but
I
am
very
interested.
What
machine
learning
can
bring
to
me?
D
So
one
of
the
things
that
we
wanted
to
detect
is:
can
we
see
if
a
cow
is
in
heat
at
any
points
based
on
data
that
we
get
straight
from
the
Scout
skin
and
it's
important
for
the
farmers,
because
a
semen
sample
is
about
two
to
four
dollars
and
to
the
four
dollars
is
a
lot
of
money?
So
you
only
want
to
put
that
in.
If
you
have
the
highest
chance,
the
cow
is
actually
getting
pregnant
and
there's
a
little
blog
post
about
that.
D
Metric
well
suitable
non
once
he
was
doing
her
PhD
internship
at
arm
and
she
was
looking
at
some
inferencing
numbers,
and
this
was
for
a
very
simple
environmental
monitoring
system
or
shakes.
He
found
that
the
energy
per
inference
for
doing
it
on
the
edge,
if
no
thix
relatively
longer
and
more
cycles,
was
over
ten
times
more
efficient
for
other
stuff,
like
image
or
image,
recognition
or
video
or
object.
Recognition
in
videos
might
be
the
other
way
around.
There
are
use
cases
where
do
you
than
the
edge
just
more
efficient.
D
D
Well,
I
say
so:
just
the
small
they're,
cheap
and
they're
efficient,
I
mean
if
they
don't
do
anything
they
just
basically
don't
consume
any
memory
either,
but
the
downside
slow,
a
very
limited
memory
so
that
prompted
us
winning
an
arm
were
a
couple
of
guys
with
in
arm
in
our
research,
not
a
complete
research
team
in
or
the
practical
end
to
come
up
with
a
library
to
do
deep
learning
on
these
type
of
devices.
Deep
learning
on
microcontrollers
and
that's
the
the
micro
transfer
projects,
so
micro
tensor
is
an
open-source
library.
D
It's
a
budget
to
licensed
specifically
made
to
run
tension
flow
models.
So
what
does
it?
You
normally
train
on
your
GPUs
that
you
can
run
directly
on
my
controller,
so
it
only
does
it
is
points
classification,
so
no
training.
So
what
you
do
is
you
push
the
trains
model
that
you
train
in
tensorflow
down
to
the
mic
controller?
And
after
that
you
can
do
classification
with
a
couple
of
really
smart
tricks.
D
We
are
able
to
do
classification
of
a
lot
of
machine
learning
models
and
under
two
minutes,
it's
6km,
so
that
includes
stuff
like
Android
and
digit.
Recognition
will
just
show
in
a
video
but
even
object,
classification
in
and
it's
a
really
good
step
forward.
Think
it's
really
cool
it.
Dis
enables
all
kinds
of
incredibly
interesting
use
cases
without
having
to
either
upstream
all
your
raw
data
to
clouds
or
do
your
inference.
You
know
a
really
expensive
machine,
so
it's
developed,
it's
not
an
arm
project
per
se.
D
It's
originated
from
arm
and
these
two
four
core
contributors,
as
you
can
see,
I'm
not
there
Neil
Neil
was
actually
in
my
team.
He's
the
developer,
evangelist
infinite
things
for
a
pack.
I
helped
it
seem
a
little
bit
so
I
wrote
the
simulator
for
this.
Who
makes
a
little
bit
easier
to
develop
it
I'm,
not
part
the
core
team,
also
not
a
machine
learning.
It's
machine
data
scientist,
anything,
that's
very
fun
it
after
it,
so
he
like,
hey
I,
want
to
develop
on
that
feel
free.
D
You
know
for
people
in
the
core
team
of
which
none
of
them
is
their
full-time
jobs
actually
develop
this.
We
we
definitely
work
on
contributions,
so
just
listen
get
up.
So
here's
a
little
demonstration
and
what
we're
going
to
do
here
is
draw
something
on
touchscreen
banana
for
scale
and
after
that
it
will
do.
It
will
basically
tell
you
what
you're
just
drawn
on
it.
So
I
hope
it's
physical
from
the
back,
because
I
was
expecting
a
little
bit
larger
screen.
D
D
And
there's
a
nice,
so
this
is
a
basic
handwritten
digit
recognition
running
on
a
microcontroller
in
less
than
two
minutes,
56k
of
ram
we're
seeing
school.
So
what
are
we
seeing
here
is
essentially
the
the
hello
world
of
machine
learning,
and
this
is
the
Emma
Stata
set,
which
consists
of
60,000
images
of
handwritten
digits.
D
So
this
is
trained
in
tensorflow
just
on
a
normal
computer
and
the
script
is
listed
there.
So
you
can
run
it
yourself
and
when
we
do
the
classification,
because
that
is
what
we
doing
in
the
microcontroller,
not
the
training
we
take
whatever
we
just
put
in
on
the
touchscreen,
we
rotate
it.
So
it's
upwards.
D
In
the
end,
we
get
an
output.
So
it's
a
single
neuron,
so
we
need
to
work
backwards.
We
need
to
scale
784
neurons
down
to
one.
Just
before
we
have
the
output.
We
have
an
output
layer
which
is
ten
potential
outputs,
because
there
are
10
numbers
0
to
9,
and
then
we
run
softmax
on
it.
So
we
only
picked
one
that
has
the
highest
probability
of
actually
matching
in
there.
Then
we
run
so
what's
kind
of
reported.
Is
that
what
we
do?
We
jump
from
the
input
layer
to
the
hidden
layer?
D
So
all
of
these
neurons
are
connected
on
both
the
input
layer
and
the
hidden
layer,
and
just
before
we
do
a
metrics
multiplication
with
nones
ratings.
We
don't
set
theta
to
be
known
for
the
training
set
with
a
little
bit
bias,
which
also
no
because
we
trained
this
model
and
they
were
in
an
activation
function.
We
did
it
twice.
We
can
talk
about
like
the
actual
details
of
this
later
now.
Why
is
this
important?
Because
how
are
we
going
to
run
this
on
a
mic
controller?
This
looks
hard.
D
This
game
looks
cool
and
so
what's
important.
This
is
matrix
multiplication
table,
because
that
is
what
does
kind
of
drive
the
memory
usage
of
this
model.
So
we
have
on
the
left
side,
784
input
neurons
and
you
have
an
hidden
layer
behind
the
28
and
an
hidden
layer
of
64
and
an
output,
well,
actually
a
layer
of
10
and
an
out
layer
of
one
holding
at
all
in
memory.
D
That's
what
we
want
to
reduce,
because
for
a
lot
of
machine
learning
models
during
classification,
RAM
usage
is
the
thing
that
just
explodes
so
the
biggest
trick
that
we're
doing
is
that
we
do
quantization
and
quantize
it.
Typically,
what
you
hold
to
weights
of
the
neurons
as
health
in
a
32
bits
floats
and
that's
completely
necessary
during
training,
because
it's
the
only
way
can
probably
hold
everything
have
like
proper
accuracy.
We
figure
out
and
the
guys
intense
flow
figured
it
out
as
well,
based
on
research
by
Pete
worden.
D
Is
that
if,
instead
of
32-bit
flows
use
8
bits
integers,
you
have
a
little
bit
of
loss
in
accuracy
so
test
against
the
cypher
10
data
set,
which
is
a
object.
Classification
in
video
feeds
about
0.4
percent
accuracy,
was,
but
that
drastically
reduces
the
amount
of
memory
that
we
need
with
leave
four
times
so
tense
flow
requires
deep
floating-point,
D
quantization.
D
So
if
you
quantify,
if
you
quantize,
you
holds
you
map
a
floating-point
to
tune
into
eight,
so
you
hold
the
minimum
value
and
the
maximum
value
that
you
might
have,
and
then
you
met
with
in
between
tensorflow
requires
to
put
it
back
to
a
float
whenever
you're
done
with
an
arrow.
We
worked
around
that
because
this
is
very
slow
on
my
controllers
that
doesn't
have
a
floating-point
units
of
them.
D
This
way
around
this,
and
now
we
see
the
current
memory
usage,
we're
very
aggressive
with
birching
layers,
so
only
the
if
we're
going
from
the
input
layer
to
hidden
layer,
one
we
don't
let
anything
else
in
memory.
So
and
then
you
see
that
the
the
metric
multiplication
table
he
needs
in
the
first
hidden
layer
is
dominating
the
RAM
usage.
D
So
for
this
model
we
see
a
big
memory
usage
of
98
kilobytes,
and
that
is
a
lot
less
than
256
K
and
we
can
push
this
even
down
further
with
a
couple,
a
lot
of
tricks
that
we're
doing.
First
of
all,
we
have
Beijing
of
memory
for
larger
models.
So
if
we
see
that
we
can't
fit
everything
in
memory,
because
we
have
too
many
neurons
at
some
points,
we
have
the
message
multiplication
table
doesn't
fit.
D
So
it's
great
actually
and
something
we're
looking
into
right
now
is
that
we
want
to
take
a
function
whose
sparsity
of
data,
so
it's
going
to
cost
a
little
bit
of
accuracy,
but
we
might
win
a
little
in
memory
with
that,
not
all
operated
for
there.
What's
the
most
important
ones
that
are
missing
right
now
are
convolution
and
pooling,
which
are
necessary
for
image,
recognition
or
object,
recognition
and
video
files,
and
these
operate
somewhat
work.
D
These
are
working
process,
and
now
the
operators
are
important
because
these
disable
predators
in
tensorflow,
so
the
whole
idea
of
micro
tensors
that
we
want
to
keep
parity
with
what
tensorflow
is
doing.
So
we
don't
want
to
have
a
new,
more
new
way
of
training
your
data
set,
because
all
that
happens
currently
intensive
loan.
That's
that's
really
cool!
There's
a
lot
of
innovation
there,
like
there's
startups
now
that
doing
JavaScript's,
machine
learning
and
I
pack,
everything
by
tensorflow
models.
D
D
That
is
a
real
set
flash
tensors,
where
the
commutes
from
flash,
of
course
at
the
sacrifice
of
speeds
and
the
sparse
tensors
and
there's
also
Network
tensors,
which
I
figured
why
the
interesting
here,
people
in
IETF
and
so
one
of
the
things
we
can
do
there
is
say
well
we're
going
to
split
part
of
her
data
set
between
devices
and
those
devices
might
be
in
a
mesh
and
they
can
all
work
together
to
actually
solve
this.
Do
this
classification?
You
think
it's
cool?
It's
something
else.
It's
floating
ideas!
D
Nothing
else
will
implement
this,
but
it
might
be
cool
that,
like
a
hundred
very
small
devices,
all
work
together
to
do
a
classification,
of
course,
only
if
your
computation
is
costs
more
than
actually
overhead,
the
against
assembled
data
runs,
and
so
the
workflow
my
credenza
consists
of
two
tools:
the
CLI
and
the
library.
So
you
start
with
a
great
set
of
data.
You
throw
it
to
tensorflow,
it
gives
you
a
protocol
file,
they
run
the
you
tensor
CLI
tools
that
does
two
things
one.
It
creates
the
metric
multiplication
tables
from
the
trend
model.
D
They
contain
the
tensor
graph,
you
combine
with
the
microtones
lip
and
then
you
have
something
running
on
a
micro,
solar
and
they
you
can
put
data
and
get
plus
quite
a
tells
if
you
look
at
and
if
you
guys
ever
if
anyone
at
U
stands
fly
before,
but
you
have
tens
report
which
can
basically
show
you
your
current
graph
visually
represented
there,
and
we
just
put
it
in
codes
here
on
the
right
and
that's
that's
how
we
put
the
graph
into
wrong
looking
at
time.
Developing
can
be
a
lot
easier.
D
We
have
something
called
the
simulators
developed
in
the
same
team
in
also
my
team,
so
we
have
the
same
mes
data
set
here
and
we
don't
want
to
run
it
on
the
device
because
it's
pretty
slow
and
if
you
want
to
refine
the
way
to
keep
your
classification,
it's
bit
annoying
as
well.
So
we
have
an
all.
My
simulator
just
runs
in
the
browser
and
two
three.
D
And
what
it
does
is
actually
cross
compiles
the
complete
model
into
JavaScript
and
then
runs
the
complete
in
inferencing
in
JavaScript.
It's
really
what
very
nice
way
of
doing
your
development,
something
that's
new
that
has
just
come
out
is
a
bunch
of
kernel,
extensions
for
cortex-m
schools,
the
CMS
snn
developer
arm,
and
the
whole
idea
is
that
it
lets
the
DSP
and
Cynde
functions
in
your
silicon.
D
If
you're
using
cortex,
m4,
cortex,
m7
or
cortex
I'm
33
and
used
it
to
speed
up
a
bunch
of
operations
that
you
normally
do
like,
for
example,
convolution
or
raloo.
So
we
see
you
speed
up
about
four
to
five
times,
which
is
pretty
awesome
because
stuff
that
you
might
think
is
kind
of
impossible.
Like
object,
classification
in
an
image
becomes
viable
this
way
and
because
it's
hardware,
acceleration
Micro
tensor,
doesn't
use
it
right
now,
but
it
will
do
that's
very
very
soon.
D
So
here's
a
video
of
the
cortex
m7
device
running
at
I
think
300
megahertz
more
or
less
doing
an
object
recognition.
So
here
in
the
left,
corner
you'll
see
the
what
is
currently
classifies
and
in
the
right
top
corner,
really
see
the
data
feeds
from
a
webcam
and
then
it
gets
down
sample,
I,
think
32
by
32
pixels
and
you
color
image
so
sees
a
frog.
Now.
D
Okay,
I
think,
is
pretty
cool,
so
there's
actually
running
in
the
cortex
m7
on
a
mic
controller.
Doing
live
up
to
classification
in
an
inner
video
feeds.
That's
it's
pretty
amazing.
So
the
speed
of
this
it
is
measured
on
the
cortex
m7
running
in
260
megahertz,
with
the
hundred
32
kilobytes
of
RAM
used
still
running
on
the
236
K,
and
at
this
point,
because
we're
at
this
point
we
have
enough
RAM
to
give
this
in,
so
that
the
speed
of
the
mic
controller
is
actually
the
limiting
parts
having
a
cortex
m4
core
exam.
D
Seven
is
definitely
going
to
help
you
with
use
cases
that
require
lots
of
computation.
So
here
we
have
three
convolution
layers,
so
before
CMS
SMN
this
took
about
half
seconds
and
now
with
shims
design
and
introduced.
We
can
do
this
in
100
milliseconds
so,
which
is
10
frames,
2
seconds
analyzing,
a
video
feeds
on
a
microcontroller
that
is
2
and
off
dollars.
D
This
is
absolutely
crazy,
so
recap
I'm
holding
a
little
bit
short
because
we're
also
already
over
time
and
how
to
get
started
well
on
why
development
boards
those
tokens
to
do
any
promotional
activity,
because
people
in
ITF
don't
like
that,
but
we've
punched
development
boards
that
are
compatible
here,
they're
link,
20,
you
tensor
profit
in
the
end,
I
think
this
is
really
cool
like
for
me
when
I
ran
into
this.
D
My
other
interests
are
LP
wins
you
know,
and
sending
large
sets
of
data
over
the
line
is
impossible,
and
this
allows
me
to
do
actual
actual
intelligence
on
the
edge
or
actually
intelligence
in
an
automation
system
where
device
to
talk
to
each
other,
but
where
I
don't
actually
have
to
push
everything
up
to
the
cloud.
I
think
this
is
going
to
there's
going
to
be
shipped
in
a
wide
variety
of
products,
and
it's
definitely
has
a
chance
to
push
the
internet
of
things
into
the
hands
of
a
lot
more
people.
So
with
that,
forgive
me.
B
Q
Yeah,
well
maybe
it's
more
Edgar
Ramos
from
Ericsson.
It's
more
like
comment
that
a
question
the
problem
I
see
with
this
kind
of
architecture
is
the
same
problem
that
we
can
see,
for
example,
with
mobile
phones
and
having
this
kind
of
many,
let's
say:
micro,
controllers
or
or
equipment
that
have
their
own
things
and
then
how
you
can
actually
address
all
of
them.
Q
D
O
F
Good
afternoon,
everyone
first
of
all
my
upload
is
because
I
couldn't
be
there
for
the
presentation.
Server
have
to
do
it
remotely.
My
name
is
Samir
Khan,
the
data
I'm,
a
research
engineer
in
eurocom,
located
in
Sochi
Antipolis
in
France,
and
in
this
presentation,
I'm
going
to
talk
about
testing
semantic
interoperability.
It's
a
small
industrial
extension
of
European
in
horizon
2020
project
called
F
and
drop,
and
for
those
who
don't
know
about
your
account
just
in
one
slide,
you
already
see
it.
F
So
we
are
a
graduate
school
and
Research
Centre
in
digital
science,
and
we
have
many
academic
industrial
as
well
as
institutional
partners
whose
of
whose
logos
you
see
in
the
screen
already
next,
please,
okay,
so
my
presentation
would
be
very
brief
because
we
have
just
started
our
work
and
basically
reporting
what
we
have
identified
as
gap
in
semantic
interoperability
testing.
What
do
we
propose
the
two
ways
and
I'm
conducting
a
survey
and
finally
conclude
the
presentation,
X.
F
Okay,
so,
as
you
know,
the
Internet
of
Things
not
only
requires
a
smart
infrastructure
but
also
deployment
of
new
services
capable
of
supporting
multiple
skill
level
and
cross-domain
applications.
It
is
often
identified
and
said
that
interoperability
is
the
key
to
achieve
the
full
potential
of
the
IOT
market
and
due
to
the
highly
dynamic
nature
of
the
IOT.
This
is
a
strong
need
of
interoperability
at
the
data
level,
and
in
that
case
it
would
be
easier
for
aggregation,
processing,
managing
and
storing
data
coming
from
all
different
IOT
devices.
F
F
So
we
need
basically
a
way
to
test
a
way
to
test
that
will
validate
the
semantic
compliance
and
interoperability
among
different
IOT
systems,
and
this
in
order
would
boost
acceptance
and
adoption
of
semantic
technologies
by
IOT
market,
and
this
is
particularly
important
because
when
I
talk
to
industry
a
lot
of
people,
they
tell
me
that
they
understand
the
benefits.
Semantics
is
bring
to
IOT,
but
they
don't
have
a
way
to
test
and
quantify
those
things.
F
So
that's
why,
in
this
industrial
extension
projects,
we
identify
distinct
semantic
interoperability
as
a
quite
a
gap
in
the
current
IOT
research
and
industry
initiatives,
and
as
mentioned
so
within
the
scope
of
F,
interrupt
projects.
This
sentence
incentives.
We
are
actually
looking
into
how
we
can
provide
some
tools,
requirements
and
guidelines
for
testing
semantic
interoperability.
Next.
F
Okay,
so
in
sentence
we
propose
two
types
of
testing:
the
first
one
being
confirmed
semantic
conformance
testing.
Here
we
want
to
test
if
a
piece
of
semantic
data
confirm
I
have
to
reference
unloading,
any
reference
ontology,
but
it
has
to
be
a
reference
ontology
and
in
the
second
type,
it's
about
interoperability
testing.
So
it
is
to
check
if
two
parties
or
two
system
under
tests
understand
correctly
the
exchange
semantic
data.
Next.
F
Initially,
when
we
started
working
on
it,
we
started
gathering
requirements
for
both
the
conformance
test,
as
well
as
the
intra
public
test.
So
for
the
conformance
test,
we
basically
clustered
everything
around
three
main
checks,
so
we
should
first
perform
lexical
check,
then
go
into
a
lot
of
syntactical
checks
and
finally
doing
a
semantic
checks
next.
F
So
this
is
a
very
basic
test
scenario
that
you
see.
So
we
have
one
system
on
our
test
and
a
tester,
so
the
sut
basically
sends
piece
of
semantic
data
to
the
tester
and
it
would,
in
in
sequence,
would
perform
all
the
three
checks
generate
a
validation
report
and
send
the
validation
report
back
into
SUT.
F
Next,
then,
we
came
into
semantic
interoperability
testing.
So
here
we
feel
that
again,
three
levels
of
validation
is
necessary.
The
first
one
being
communication
level
check.
So
here
we
basically
want
to
see
that
the
message
sent
from
the
first
system
should
be
received
completely
and
correctly
by
a
second
system.
Then
we
have
a
lexical
or
format
level
check
here
at
the
message
issued
from
the
first
system
should
be
incorrect
and
understandable
format
for
the
second
system
for
death.
F
Further
data
processing
and
the
third
level
is
the
data
processing
level
check
next
said:
I
tried
to
represent
a
very
basic
scenario
using
two
system
under
tests
here
and
in
in
this
particular
test.
Our
objective
is
to
test
the
semantics
crossing
results
in
two
I
mean
from
the
two
systems
so
consider
a
suti
one.
It
sends
a
piece
of
somatic
data,
let's
call
it
D
1
2
SOT
and
the
sot
2
runs
a
query:
q1
aspartyl
query,
Q
1
on
top
of
D
1,
and
then
s
UT
2
sends
back
the
result.
F
R
1,
as
well
as
the
query,
Q
1,
a
31
and
the
system
under
test
one.
It
would
execute
the
same
query,
Q
1
on
the
same
data
I
can
send
before
and
let's
say
the
result.
This
time
is
our
own
complement,
so
su
t1
would
actually
verify
if
both
r1
and
r1
complement
are
the
same.
If
they
are
same
or
equivalent,
we
would
say
we
have
100
percent
probability,
otherwise
would
say
it.
These
two
system
under
tests
are
not
semantically
interoperable.
Next.
F
Now,
since
you
come
and
our
partner
in
the
project
is
a
global
market,
so
we
both
work
as
well
as
follow
the
1
M
2
M
standards.
We
want
to
map
the
interactions
and
the
system
under
test
in
the
previous
slide
with
one
and
two
entities.
So
in
this
presentation
you
see
that
we
have
here
Lisa
re
I'm,
getting
some
echo
okay.
F
So
in
this
slide
you
see
that
we
have
basically
identified
the
application
entity
as
well
as
common
service
entity,
of
the
1m
tomb
architecture
and
wish
for
the
same
set
of
operations,
and
we
divide
them
with
respect
to
the
application
in
the
D
and
the
common
service
entity.
So
it's
basically
the
same
representation
of
the
slide,
then
with
respect
to
Valen
to
them
in
and
dislike
that
next
now,
I
think
starts
to
get
a
little
bit
more
interesting
when
we
introduce
additional
computational
components
between
the
system
under
tests.
F
So
let's
say
we
have
a
very
server
which
is
generating
the
query
q1
and
it
is
sending
the
same
query
both
to
in
51
and
two
and
then
the
system
under
test.
They
execute
the
query
1
on
their
semantic
data
and
return
the
results
to
the
query
server
and
the
query
server
is
going
to
compare
our
run
and
r2
and
check
if
they're
same
it's
the
same.
Then
we
have
achieved
interoperability.
Otherwise
we
haven't
next.
F
F
Check
if
they
share
the
same
vocabulary,
so
when
a
suti
one
sends
the
semantic
data,
do
you
want
to
the
tester,
so
the
tester
would
retrieve
the
vocabulary
v1
from
the
semantic
data
it
receives,
and
it
would
do
the
same
thing
for
D
2
coming
from
AC.
To
do
so,
it
would
retrieve
the
book
ability
to
and
then
compare
v1
and
v2,
and
the
comparison
result
would
be
sent
both
to
both
of
the
system
under
test.
F
F
Then
we
also
can
think
of
a
little
bit
more
complicated
scenario
for
data
level.
Interoperability,
testing
and
the
main
motivation
of
imagining
are
considering
this
type
of
scenario
is
that
we
have
to,
since
we
are
building
an
industrial
extension
of
the
Avenger
of
platform,
so
we
would
like
to
use
what
it
if
enter,
of
offer
offers
and
extend.
E
F
F
Doing
one
service
so
I
would
request
all
of
you
to
take
note,
5
minutes
and
complete.
This
survey.
Re
has
obvious
already
done
it,
and
so
I
request
all
of
you
to
do
it.
It's
gonna
help
this
project
a
lot
and
if
you
see
I,
have
missed
any
requirements
or
did
not
consider
any
requirement
feel
free
to
drop
a
note
there
as
well.
Next.
F
So
this
is
our
acknowledgments
in
synthesis
is
an
industrial
extension
of
the
F
internal
project
which
I
received
funding
from
European
Union
H
turn
20
project,
and
some
tests
is
a
joint
collaborative
project
between
easy
global
market
and
York.
So
we
both
are
located
very
close
to
HC.
If
you
have
visited
so
beautifully,
so
you
definitely
know
that's
all.
F
If
you
have
any
question
you
can
ask
me
or
read:
if
you
are
coming
to
the
the
opt-ins
meeting,
so
we
can
meet
and
discuss
personally
and
please
take
the
service
it.
It's
going
to
help
a
lot.
So
with
that
I
would
say.
Thank
you
very
much
for
the
opportunity
to
speak
here
and
thanks
for
your
listening.
F
B
R
R
Okay,
thank
you
and.
R
R
Input
and
then
provide
without
selecting,
enter
verify
that
the
result
is
correct
and
then
in
the
entrance
perform
the
payment
item
right
picture
so
right
now
this
is
done
in
the
cloud,
so
there's
Microsoft
Amazon
at
the
computations
pallets
and
resources
and
get
back
the
result.
However,
we
have
I
think
anyone
running
it
at
the
end
has
several
benefits
in
terms
of
like
cost
and
also
privacy.
R
So
in
our
efficient
we
want
to
have
a
mesh
at
nodes.
So
even
if
I
have
my
laptop
I
have
some
spicy
PU
I
can
just
accept
some
tasks,
run
locally
and
device
for
the
execution.
However,
it
changes
kind
of
the
the
trust
management,
because
Google
or
Amazon
will
just
keep
them
our
data
and,
in
fact
they're
not
here,
it's
like
untrusted
knowledge
with
the
security
and
privacy
must
be
deal
in
the
system
design.
So
we
cannot
just
trust
that
the
other
part
is
not
going
to
misbehave.
R
E
R
Wanted
to
reward
for
that,
so
it
should
be
also
the
main
driving
point
for
joining
the
system,
because
in
my
gt-r
system
that
was
the
problem,
there
was
no
incentive.
So
if
you
consider
torrents
everyone
else
to
get
files,
but
then
why
would
I
share
my
like
to
use
my
bandwidth
to
give
to
other
people?
However,
it
is
for
paying.
R
I
would
like
to
verify
the
result,
so
be
sure
that
I
didn't
just
random
link
of
bytes
and
there's
also
this
problem
when
if
the
payment
should
be
done
because
if
I
am
requesting
now
I
don't
want
to
pay
in
advance
first
and
execute
a
stroke
and
just
take
the
money
and
walk
away,
and
the
same
applies
from
the
execution
now.
So
you
don't.
R
Just
walk
away
without
pain,
so
the
result
cache
is
also
very
crucial
part
of
the
system.
So
we
can
have
different
types
of
tasks,
so
we
can
proof
of
work
types
of
tasks,
but
it's
very
easy
to
to
check
if
the
task
buzzer
was
correct
or
not
so
it
just
compared
to
get
hash,
however,
usually
means
we
have
to
recompute
the
whole
task
and
then
compare
it.
The
result
is
the
same.
R
We
can
generate
prospects,
your
knowledge
across
to
posted
correct.
However,
usually
it
applies
very
high
cost,
sometimes
it's
even
much
higher
than
the
computation
itself,
and
also
it's
not
available
for
every
single
computation.
Another
solution
would
be
to
send
the
task
to
many
people
that
we
get
back
the
result.
We
compare
them
they're
the
same,
there's
a
high
term
that
it
was
the
result
is
correct.
However,.
R
So
then
we
also
submit
some
input
and
we
get
some
result
and
we
want
to
make
it
private,
but
not
only
private
from
them
for
the
network,
but
also
from
the
execution
platform,
and
this
is
difficult
because
the
execution
platform
us
to
actually
compute
something
on
top
of
our
data.
So
in
the
cards
usually
now.
We
use
encryption,
however
again,
and
it
reduces
quite
a
lot
of
overhead
and
it's
not
always
possible.
R
However,
they
require
dedicated
Hardware,
so
now,
I
just
filled
in
our
system,
so
fast
is
Intel
SGX.
So
it's
an
example
of
this
trusted
execution
environment.
So,
basically,
what
it
does.
It
allows
us
to
create
protected
part
of
memory.
That
is
an
animal
running.
On
top
of
that.
So
then,
only
this
process
can
read,
read
and
write
to
this
part
of
memory.
It's
invisible
and
encrypted
from
the
operating
system,
world
hypervisor.
R
I
can
check
first
of
all
this
on
the
remote
host.
I
have
the
valid
internet,
yes,
but
for
then
I
can
verify
that
the
function
running
on
the
remote
host
is
the
function
I
requested
and
in
the
end,
I
call
so
secure.
Communication
channel,
so
I
can
communicate
directly
with
the
function
kind
of
bypassing
the
the
execution
node,
and
we
also
leverage
hariom
block
time.
So
we
built
on
smart
contracts
that
allowed
to
add
some
logic
on
top
of
a
blockchain.
In
our
case,
it's
Italian
with
solidity.
R
D
R
It
is
like
almost
no
cost
and
no
delay
in
processing,
so
there
are
secured
by
deposits
run
on
top
of
blockchain.
However,
all
the
communication
is
of
blockchain
and
the
last
component
are
Oracle's,
so
smart
contracts
by
themselves
they
have
no
la
légion
or
they
cannot
communicate
with
the
external
world.
So
we
have
like
virtual
machine
and
Oracle's
can
act
as
trusted
data
feeds.
So
if
I
want
to
query
a
trusted,
HTTP
HTTP
server
I
can
do
it
with
an
Oracle
and
then
they
die.
God
is
valid.
R
R
However,
both
you
know
they
trust
the
blockchain
and
also
we
assumed
that
execute
enough
has
complete
control
over
its
system,
operating
system,
hyper,
hyper,
visor
and
so
on.
There's
going
to
be
a
lot
of
so
here
requesting
now
some
input
on
the
top
on
the
bottom.
We
have
the
execution
that
and
on
the
right
web
block
change
running
with
smart
contracts.
So,
first
of
all
and
the
execution
part
some
brands
tank
light
enter
now.
R
The
the
requesting
node
can
establish
this
secure
communication
with
the
with
their
function,
enter
and
identity
to
generate
the
proof
enter
and
send
it
back
to
the
to
the
requesting.
Now
and
now
we
know
that
it's
secure
to
send
our
input
to
the
Sudan
cliff
and
again.
This
input
that
we
sent
is
now
secured
from
the
from
the
executing
platform.
So
it's
only
visible
for
the
function,
and
so
now,
after
some
time,
the
language
will
will
compute
a
result
and
at
this
point
the
execution
path,
some
generates
the
secret.
R
E
R
Is
sent
back
to
the
food
executioner
so
now
we
have
the
result,
however,
it's
encrypted
enter.
We
also
send
a
hash
of
the
scheme
that
we
use
encrypt
the
result
enter
and
both
the
result
and
hash
have
at
the
station
from
the
Envoy,
so
we're
sure
that
they
are
correct,
and
so
now
the
execution
note
is
send
again
a
transaction
with
the
payment
photo
executing
node
in
platform.
However,
it
can
be
unlocked
only
if
we
submit
this
the
secret
key
and
can
verified
it
using
this
hash
enter.
R
So
now,
if
the
executing
node
submits
the
secrets
to
the
blockchain,
it
will
unlock
the
payment
enter
and,
at
the
same
time,
the
key
the
key
becomes
visible.
So
the
so
the
client
gets
the
secret
and
use
it
to
unlock
for
the
result,
and
so
here
it
works.
However,
we
have
two
problems,
so
first
of
all,
it
can
be
pretty
slow
and
expensive
to
do.
R
Blockchain
and
second
of
all,
I
can
always
send
some
computation
to
the
executing
node,
wait
until
it
executes
my
stuff
and
then
just
walk
away.
So
this
kind
of
denial
of
service
attack
answer
so
to
deal
with
this
problem.
We
here
introduce
payment
channels
so
now
sending
this
money
and
unlocking
is
she
cheap,
has
almost
no
cost,
and
it's
also
very
fast.
So
we
can
so
can
divide
our
ties
into
very,
very,
very
well
into
a
lot
of
small
tasks
into
everything
chunk
by
chunk
enter
so
here.
This
whole
process
is
the
same.
R
Do
it
very
quickly?
So
now,
even
if
I
submit
a
chunk
of
work,
I,
don't
know
10
seconds
of
CPU
and
just
walk
away.
The
execution
platform
uses
only
10
seconds
of
its
time
and
not
more
and
and
to
make
it
perfectly
secure.
The
problem
is
in
this
communication,
so
in
this
channel
of
data
between
the
repeating
note
and
the
Enclave,
because
it's
secured
and
no
one
knows,
what's
happened.
R
So
if
the
communication
fails,
the
smart
contract
or
no
one
else
is
not
able
to
say
if
it's
because
of
the
network,
failure
or
because
the
execution
doesn't
didn't
want
to
compute
anything
almighty,
just
the
requester
walked
away,
and
so
basically
we
can
either
sum
everything
to
the
blockchain,
and
then
everyone
will
know.
But
of
course
it
applies
big
cost
and
in
our
solution
we
decided
to
go
with
ipfs.
N
Is
the
time
for
short
question
I
was
wondering
whether
you
did
a
comparison
with
some
of
the
others
of
computing
solutions
that
are
out
there.
Did
you
did
you
do
a
sort
of
a
comparison
between
your
approach
and
di
some
other
edge
computing
platforms
out
there
like
this?
As
you
know,
this
work
on
open
fork
like
Microsoft
Amazon.
We
and
others
have
sort
of
an
edge
computing
solution,
and
so
I
was
wondering
whether
you
have
done
a
comparison.
No,
it
is.
N
R
N
Terms
of
in
general,
the
solution
which
you
have
worked
on
an
edge
computing
solution,
as
you
explained,
and
once
with
one
specific
set
of
characteristics
in,
but
there
are
other
edge
computing
solutions
out
there
that
people
worked
on
and
they
have
different
characteristics.
I
would
be
curious
on
how
they
compare
to
each
other
yeah.
R
N
R
However,
if
you
want
to
have
model
that
will
just
dispatch
tasks
automatically
between
nodes
and
I
think
it's
difficult,
because
then,
if
the
payment
is
involved,
how
do
we
define
kindness
in
this
case,
especially
if
we
have
like
different
priceless
loads
and
not
capacities,
also
how
to
estimate
the
cost
of
the
computations
in
terms
of
CPU
bandwidth
memory,
quality
of
service?
Knowing
again
that
we
are
very
heterogeneous
environment
and
how
to
provide
full
privacy,
because
in
this
case
we
will
protect
input
and
the
result.
However,
we
still
know
who
involved
which
function.
S
Dave
Eric,
so
mine
is
not
about
the
agenda
per
se
about
the
logistics
of
getting
there.
The
meeting
is
tomorrow
afternoon
in
Prague
and
I
know.
Several
people
are
on
the
7:20
a.m.
flights
out
of
he
throw
so.
My
question
is
as
far
as
carpooling,
since
this
is
before,
like
Heathrow
Express
runs
and
so
on.
I'm
organizing
a
car
can
I
get
a
count
of
how
many
people
would
want
to
travel
together.
We
would
be
leaving
the
hotel
lobby
at
about
4:20
a.m.
tomorrow.
S
Okay,
if
you'd
like
to
carpool
and
come
together
in
a
shared
car,
please
raise
your
hand,
so
I
can
count
one
two
three
anybody
else
here.
Ok,
you
should
need
to
give
the
the
the
concierge
account
for
how
big
of
a
car
to
get
so
so
far,
I
see
three
in
this
room.
If
you
know
of
anybody
else,
let
me
know
because
after
this
meeting,
I
will
go
and
get
the
car
booked.
S
B
D
B
That
we
actually
having
33cm
and
LCIF
in
one
room
together
with
a
research
group
and,
of
course,
we're
going
to
do
status
updates
between
each
other.
We
will
also
talk
about
Vichy
again,
but
there
are
a
few
issues
that
really
aren't
standardized
yet
and
so,
for
instance,
one
issue
that
comes
up
repeatedly
is
that
when
you
actually
request
some
action
from
an
IOT
device,
how
do
you
know
what
data
you
actually
need
to
send
with
that
action
so
that
the
energy
device
can
do
the
right
thing?
B
B
So
that's
one
thing
we're
going
to
talk
about
more
generally,
we
will
talk
about
the
model
interoperability,
some
things
we
tested
in
the
hackathon
and
yeah.
What's
the
next
step
to
make
sure
the
various
models
that
are
popular
right
now
in
to
operate
the
episode
or
that
way,
the
model
is
relatively
simple
and
straightforward:
pretty
easy
to
translate
three
descriptions:
very
powerful
ocf.
Has
this
Remlinger
based
modeling
schemes?
Maybe
you
put
these
things
together?
B
Another
thing
again,
the
protocol
level
is:
how
do
we
model
the
various
forms
of
pushing
data?
So
we
know
publish/subscribe
is
a
pretty
popular
way
to
model
the
edge,
but
also
his
problem
problems
and
would
be
interesting
to
see
whether
yeah
additional
experience
of
rest
beyond
what
we
already
having
card,
which
is
the
observed
function,
would
be
used
for
a
tool
to
better
handle
these
push
models.
So
there's
actually
a
research
group
draft
for
non-traditional
responses
and
we
want
to
discuss
how
use
what
that
would
be.
B
We
want
to
talk
about
the
area
of
data
that
you
will
actually
install
or
obtained
from
an
IOT
device
about
its
relationships
to
other
devices.
So
how
can
you
actually
find
the
light
switch
to
a
light?
How
do
you
describe
the
interfaces?
How
do
you
build
the
links
and
how
do
the
resource
representations
look?
Like
and
then,
of
course,
we
have
housekeeping
on
the
base
document
like
new
response
codes
talking
about
the
resource
directory,
there
are
different
ideas
on
how
to
use
that
and
so
on.
B
And
finally,
we
are
going
to
talk
about
ace
in
the
security
authorization
working
group
within
the
ITF
and
how
that
could
be
useful
for
the
ocf
which
have
their
own
security,
our
model
already
in
place,
and
they
also
have
interesting
authorization
formats
at
Salomon
but
talked
about.
But
one
interesting
question
is:
how
do
we
get
multi-point
security
in?
So
somebody
sent
me
a
link
to
the
fair
hair
white
paper.
B
The
fails
security
white
paper,
which
nicely
describes
how
they
are
going
to
do
all
the
security,
and
then
there
is
this
section:
seven
there
which
and
where
is
security,
I
think
that's
really
something
that
we
need
to
do
some
work.
So
this
is
the
sanitation
and
finally,
of
course,
we
all
benefit
from
the
availability
of
reference
implementations
and
from
test
cases.
So
we
want
to
discuss
how
we
get
to
a
better
situation
with
color--the
particular
property,
CPC
well
and
other
standards
that
we
are
using.
B
So
this
and,
however,
we
want
to
go
in
one
afternoon
that
will
be
here
as
background
program,
but
yeah
we
can
always
affect
our
things
to
telephone
part
later
on.
So
this
is
so
next
meeting,
I
already
said
a
little
bit
about
two
other
meetings
going
on
and
I'm
very
certainly
we're
going
to
have
a
summary
meeting
too
much
car.