►
From YouTube: Edge-Cloud Synergy, the Last-Mile Delivery for AI
Description
A
B
Before
I
kick
off,
I'd
like
to
first
have
a
question
for
you.
How
many
of
you
have
been
interested
in
edge
intelligence?
A
lot
of
you,
you're,
really
ahead
of
curve.
Indeed,
edge
intelligence
is
a
new
topic
in
academia
and
vertically,
but
actually
it's
somehow
still
new
in
the
vertical
industries.
I'm
o
chen
guang.
B
I'm
from
hai
team
from
huawei
cloud
edge
cloud
innovation
lab
where
is
seeing
on
edge
cloud
and
the
key
technologies
and
the
challenges
in
cloud
edge
as
well
as
low
hanging
fruits.
Today,
I
will
share
with
you
something
on
a
vertical
line
and
its
innovation
in
the
cloud
edge.
First,
I've
been
working
on
the
cloud
computing
and
notch
and
cloud
computing
and
am
now
responsible
for
sender,
development,
operation.
B
B
B
It
centers
some
cloud
only
when
you
have
a
hub
cloud,
you
can
do
computing,
otherwise
you
can't
do
that
or
you
might
refer
to
other
scenarios
as
embedded
scenarios.
That's
why
to
articulate
hai.
I
have
to
do
some
comparison
with
two
things.
The
first
one
is
about
the
cloud
intelligence
which
centers
on
cloud.
Second,
one
is
about
embedded
ai.
So
later
I
will
touch
upon
that
to
give
you
a
sense
about
what
do
I
mean
by
that.
B
Like
there's
a
detection
of
a
falling
elderlies,
as
well
as
the
crying
of
babies,
that's
all
relevant
to
intelligence,
and
the
second
is
about
the
smart
health
care.
It's
about
smart
diagnosis
in
smart
cities.
We
see,
for
example,
those
municipal
management
for
those
unlicensed
running
of
business
along
the
roads.
The
fourth
one
is
ar
vrm,
because
there
are
many
data
to
be
transmitted.
B
There's
also
a
kind
of
scenario
of
hai,
and
next
one
is
intelligent
transport,
industrial
internet
talking
about
intelligent
transformer
means
the
smoke
detection
internals,
as
well
as
reverse
driving
all
the
wrong
driving
detection
and
industrial
internet,
for
example,
detection
of
electric
circuit
boards.
The
next
one
is
about
smart
agriculture.
I'm
not
sure
whether
you
know
that,
for
example,
we
use
drones
to
try
to
provide
those.
B
B
We
just
talked
about
the
concepts
like
hai
and
embed
ai,
but
what
are
differences
between
cloud
intelligence
and
ed
ai?
I'm
talking
about
cloud
intelligence,
ai!
So
because
you
put
everything
on
the
cloud,
it's
easy
to
have
latency
and
cost.
That's
quite
obvious
challenge,
but
what
is
embedded
intelligence
like
at
the
entrance
of
the
campers
it's
necessary
to
recognize
license
plates,
of
course,
but
the
challenge
is
that
you
they
will
never
update
the
software
running
in
the
equipment
of
the
detector
for
one
or
two
years.
B
So
that's
why
the
challenges
about
how
can
you
really
run
the
system
with
the
minimum
cost?
So
this
is
something
that
will
be
researched
by
combining
hai
and
cloud
by
collaborating
those
two
factors
and
we
believe
it's
possible
to
cope
with
the
above
mentioned
the
challenges
when
talking
about
hai
what
are
specific
or
what
challenges
are
specific
to
at
ai,
the
first
heterogeneous
software
southern
hardware.
B
B
We
also
see
ai
firmware,
like
my
sport,
tensorflow
pedal
pedal
and
I
don't
list
some
of
them
here,
but
those
are
mainstream
ones
and
when
we're
in
the
system
of
hydrogenated
software
and
hardware,
how
can
we
ensure
consistent
experience
of
customers
by
only
calling
apl
one
time?
Is
it
possible
to
just
call
or
manage
a
large
number
of
applications?
I
think
that's
represent
all
that
is
the
biggest
challenge
ahead
of
us.
Second,
about
limited
edge
resources.
B
B
B
When
cameras
take
picture
of
background
is
different,
mostly
graying
in
spring,
the
black
white
or
gray
in
winter.
Mostly,
so,
you
see
the
varieties
of
colors
in
different
seasons.
What
do
I
mean
by
spatial
differences
and
it
means
the
gate
of
all.
The
entrance
of
the
eastern
campus
is
different
from
that
of
the
one
in
the
west,
because
there
are
different
dimensions
and
data
laid
out
is
different.
B
B
B
B
So
I
would
like
to
use
some
typical
hai
scenarios
to
elaborate
on
that.
So
this
one
is
safety,
helmet
detection,
for
example.
How
does
it
look
like
when
you
detect
those
helmets,
for
example,
in
foggy
days
rainy
days,
a
strong
light
at
construction
site
when
there
is
any
light
at
the
construction
line
during
night,
and
you
might
have
a
disruption
of
the
sharp
rays
of
light
or
installation
and
angles
differ
from
each
other
because
of
the
distance
between
the
camera
and
construction
size?
A
A
A
From
the
beginning,
from
the
end
of
last
year,
until
last
month,.
A
A
A
We
have
edge
nodes,
managements,
application
management,
hardware,
resource
management
and
cross-edge
traffic
governance.
This
is
a
cube
edge
infrastructure
capability,
and
on
top
of
that,
we
built
ai
framework
cross
edge
data,
set
management,
collaborative
inference,
incremental
learning
and
lifelong
learning.
A
A
A
We're
now
working
with
partners
from
the
community
on
some
cases,
if
you're
interested
in
that
you're
welcome
to
participate
in
our
community
regular
meeting.
Usually
we
have
very
heated
discussion
each
meeting.
Usually
the
meeting
will
last
for
two
hours,
but
actually
the
meeting
will
go
beyond
two
hours
up
to
two
and
a
half
hours
or
even
three
hours,
so
I'll
talk
about
what
is
hai
and
what
are
the
challenges,
the
things
that
we've
done
in
the
community
and
how
people
are
working
together
to
address
those
challenges
concerning
setna.
A
A
A
A
A
A
A
A
A
A
B
B
B
B
B
Distribution
challenge
so
in
some
cases,
for
example,
smart
traffic-
there
are
a
lot
of
long
tail
protocols
algorithm,
for
instance,
smoke
detection
containers
is
similar
to
license
plate
recognition.
Facial
recognition,
they're
seldom
used
it's
hard
to
gather
data
for
those
rarely
happening
scenarios.
So
that's
why
we
then,
usually
the
accuracy
is
not
very
big
70
or
80
percent
are
really
good
enough.
Already,
in
that
case,
is
it
possible
once
we
push
model
to
to
the
edge
and
the
context
in
the
context
related
to
edge,
we
may
collect
relevant
data
and
we
may
collect
samples.
B
B
So
the
current
incremental
learning
is
quite
basic.
We
can
only
achieve
very
simple,
incremental
learning,
but
we
have
a
multiple
task,
incremental
learning
to
make
it
even
better
in
the
future
and
a
community
we're
still
discussing
so
how
about
our
policy
of
a
data
gathering
or
sample
gathering?
Is
it
possible
for
a
system
to
collect
the
most
efficient
samples
to
enhance
incremental
learning?
B
What
we
are
discussing
about
how
to
live
it
to
another
level,
but
we've
already
amassed
some
fundamental
technologies,
so
next
one
is
about
federated
learning.
As
you
know,
this
is
a
design
to
solve
one
thing
without
having
data
leaving
the
campus,
because
data
is
supposed
to
be
at
edge
in
ash
computer
as
computing.
That's
why
combination
between
edge
computing
and
federal
learning
is
quite
natural.
The
combination
can
help
us
to
solve
fifth
challenge
which
is
related
to
security.
B
B
It's
applicable
to
industrial
internet,
because
first
enterprises
owner
do
not
want
let
their
data
leave
the
campers.
The
second
data
amount
in
an
enterprise
or
product
production
lines
are
limited
if
it
is
possible
to
use
federal,
related
learning
and
then
it's
possible
to
train
a
good
model.
So
this
is
all
about
federated
learning.
B
Suppose
there
is
a
huge
amount
of
data
at
one
node,
but
not
much
another
narrow
to
run
federally
learning.
How
can
we
ensure
the
results
will
not
link
to
the
node
with
more
data
samples,
so
this
is
to
solve
buyers
and
besides
and
how
to
process
batch
status
because
after
training,
it's
really
hard
to
input
all
of
those
large
amount
of
data
into
the
learning
that
will
really
slow
down
the
whole
process.
So
those
are
the
discussions
we
are
having
today.
B
B
This
is
about
case
how
to
use
edge
cloud
collaborative
inference
and
and
incremental
learning.
We
see
40
percent
of
precision,
improvement,
left
hand
side,
you
see,
someone
is
wearing
a
casual
hat
who
hasn't
been,
who
was
not
detected,
but
he
was
detected
after
the
collaboration
of
the
adjacent
cloud
on
the
right
hand,
side
and
the
casual
head
was
misrecognized
as
safety
helmet.
B
I'd
like
to
share
with
you
some
roadmap
in
march,
we
released
a
version
and
we
built
the
basic
models
and
frameworks.
On
top
of
that,
we
add
a
federal,
related
learning,
collaborative
thing
for
instrumental
learning
and
then
in
q2
will
add
lifelong
learning,
transfer
learning.
What
is
lifelong
learning.
B
B
If
I
get
some
results-
and
I
will
also
add
this
new
knowledge
and
accelerate
the
learning,
it
means
the
previous
samples,
as
well
as
the
parameters
of
models,
can
be
reused
to
accelerate
the
learning
target.
So
this
is
what
do
I
mean
by
lifelong
learning
and
we'll
see
a
release
in
june
for
this
lifelong
learning,
and
then
we
also
later
have,
for
example,
collaborative
inference
between
ads
and
at
cloud
collaborative
tracking,
and
it
means
through
multiple
edge.
B
B
We're
now
designing
this
solution,
and
the
fourth
milestone
is
that
we
will
have
optimization
for
automated
edge
model.
You
know
architectures
and
the
computing
operators,
as
well
as
memories,
are
all
different
across
astronauts.
Is
there
any
way
for
me
to
really
automatically
optimize
and
price
the
most
suitable
edge
model
that
fits
a
node?
That's
a
very
interesting
topic,
and
last
we
also
integrate
with
huawei
cloud
ief.