►
Description
The combination of the potential behind machine learning (ML) and AI and its sudden accessibility in the form of open source libraries has led to an explosion of excitement about it. But what will this mean for you and your datacenter? Join Mirantis‘ Nick Chase as he gives you a fast-paced introduction to machine learning and how it actually works, as well as a look at what it might look like in action in your datacenter — especially in terms of projects that already exist.
Attendees will learn about:
- The basic types of machine learning and AI, and how they will affect the datacenter
- How machine learning can improve performance and functionality
- Using machine learning to predict and prevent issues
- How today‘s open source cloud tools can be used in an ML-enabled environment
- How machine learning and AI might shape the evolution of the datacenter
Join us for KubeCon + CloudNativeCon in Barcelona May 20 - 23, Shanghai June 24 - 26, and San Diego November 18 - 21! Learn more at https://kubecon.io. The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy and all of the other CNCF-hosted projects.
A
All
right
that
seems
to
be
pretty
stable,
so
we
need
started
now.
My
name
is
Mark
Coleman
I'm,
the
CN
CF
marketing
chairperson,
we're
joined
today
by
Nick
chase
former
Ensis
who's
gonna,
be
telling
us
about
machine
learning
and
AI
in
the
dentist
data
center,
not
dental
center,
and
how
it
will
affect
you.
Now.
We've
had
an
unprecedented
number
of
sign
ups,
and
that
means
there's
also
a
lot
more
people
in
the
call.
A
None
of
you
are
able
to
speak,
but
what
you
can
do,
if
you
have
questions,
is
put
them
in
either
the
QA
or
the
chat
and
I'll
try
to
find
opportune
moments
to
interrupt
Nick
and
ask
him
those
questions,
but
I
can
tell
you
ahead
of
time
that
we're
not
gonna
be
able
to
get
to
all
of
them.
So
I
apologize
for
that.
Nick
will
probably
give
you
social
media
details
or
something
drew
out
during
the
call,
and
you
could
perhaps
follow
up
him
later
I.
A
B
Well,
thank
you
very
much
Mike
good
morning.
Everybody
welcome
to
today's
webinar
I
kind
of
like
the
whole
dental
center
thing,
though,
but
we'll
move
on
before
we
do
go
anywhere.
I
want
to
tell
you
a
little
bit
about
me
if
you've
ever
been
at
any
of
my
webinars.
Do
you
know
that
I
tend
to
whip
right
on
by
this
slide,
because,
let's
face
it
you're
here
for
the
content?
B
B
Looking
into
it
and
I
found
something
out
first
off,
if
you
are
one
of
those
math
PhDs
who
does
all
this
fabulous
research
and
discovers
new
algorithms
and
creates
tool,
it's
like
tensorflow
and
scikit-learn
to
make
this
stuff
possible.
God
bless
you,
we
couldn't
do
it
without
you.
Please
don't
kill
me
for
what
I'm
about
to
say
next,
here's
what
I
found
out
actually
doing
machine
learning,
that
is
understanding
the
basics
and
getting
the
computer
to
figure.
Something
out
is
something
you
can
pretty
much
do
with
high
school
algebra.
B
If
that-
and
that
is
because
we
have
these
tool
kits-
that
let
us
mere
mortals
do
machine
learning,
which
is
about
to
open
up
a
whole
new
world
of
what
we
can
do
and
that's
what
we're
going
to
talk
about
today.
So
if
you
work
in
IT
or
your
developer
or
you're
just
interested
in
what
machine
learning
can
do
and
how
it
does
it
and
what
that's
gonna
mean
for
the
data
center
going
forward,
this
webinar
is
for
you,
if
you
don't
know
anything
about
machine
learning.
We're
gonna,
give
you
the
basics.
B
B
So
of
course,
while
I'm
gonna
give
you
the
basics,
we
cannot
possibly
cover
all
of
machine
learning
in
an
hour,
so
I've
conned
I
mean
talked
Manning
publications
into
giving
away
free
access
to
my
upcoming
video
course
machine
learning
for
mere
mortals
everyone
who's
here
live
today.
If
you're
catching
the
replay
I'm
going
to
try
and
do
something
for
you
too,
but
if
you're
here
you're
live,
you
will
get
access
to
the
Manning
early
access
program,
which
means
that
you
get
to
see
the
draft
videos
as
they're
completed.
B
So
please
go
easy
on
me
if
there's
still
some
glitches,
so
you
can
hit
that
link
here
and
fill
it
out.
We'll
also
put
the
link
back
up
at
the
end,
all
right.
So
a
quick
look
at
women
talk
about
here,
we'll
start
out
with
a
quick
introduction
to
machine
learning
itself
and
the
different
types
that
are
out
there
and
then
we'll
move
on
to
the
various
issues
that
face
the
data
center
and
how
machine
learning
and
AI
techniques
can
help
solve
them.
B
Then
I'm
going
to
show
you
how
easy
it
is
to
actually
do
this
stuff
by
working
a
simple
example
with
you're
using
tensorflow
and
a
little
bit
of
Python,
you
are
not
going
to
believe
you
how
straightforward
this
is.
I
promise
you
finally
we'll
tie
it
all
together
with
a
look
at
how
to
get
started
and
we'll
go
to
QA
all
right.
So,
let's
start
by
me
be
sure
that
we're
all
on
the
same
page
and
look
at
just
what
we're
talking
about
when
we
say
machine
whirring
all
right.
B
So
here
we
have
the
obligatory
definition.
Machine
learning
is
the
study
of
enabling
a
computer
to
perform
in
situations
it
may
not
have
seen
before.
So
what
does
that
actually
mean?
Let's
look
at
it
this
way.
If
you
go
to
your
browser
and
you
type
in
kubernetes
I/o,
the
browser
knows
exactly
what
to
do.
It
goes
to
the
kubernetes
website.
That's
not
machine,
learn
if
you're
just
typing
kubernetes,
that's
probably
not
really
machine
learning
either
because
Google
know
Google
knows
exactly
what
you're
searching
for
it
because
you
typed
it.
B
On
the
other
hand,
if
you
open
up
Siri
or
Google
assistant
or
relax
on
your
phone
and
say
what
is
kubernetes,
that's
a
whole
other
story.
First,
whichever
assistant
it
is,
has
to
translate
the
sounds.
You've
made
into
specific
words,
and
then
it
has
to
interpret
what
you
meant
by
those
words
now,
of
course,
your
phone
doesn't
have
labeled
sound
files
of
you
speaking
every
word
in
the
world,
so
it's
going
to
have
to
interpret
what
you
said
by
comparing
it
to
sound
files
that
it
does
recognize
or
probably
more
specifically,
features
of
those
files.
B
The
point
is
that
the
system
learns
to
do
things
that
it
hasn't
explicitly
been
trained
to
do
by
extrapolating
from
knowledge
that
it
already
has
now,
while
we
may
think
of
AI
and
machine
learning
as
something
really
new.
In
fact,
it's
been
around
for
a
fairly
long
time,
and
these
days,
it's
probably
pretty
difficult
to
get
through
even
a
day
without
being
affected
by
it
in
some
way
or
another.
B
Some
of
the
more
common
places
it's
used
would
be
in
analytics
to
help
interpret
data
advertising,
of
course,
to
help
target
people
who
are
most
likely
to
be
receptive
to
your
message.
It's
to
make
them
sound
more
like
humans,
general
pattern,
recognition
such
as
identifying
specific
types
of
customers
or
even
identifying,
where
there's
a
particularly
need
for
a
new
product,
natural
language
processing.
For
things
like
to
hear
your
balance,
say
balance
or
to
determine
what
people
are
talking
about
online
and
and
of
course,
search
now.
B
More
recently,
we've
seen
it
in
image
recognition,
so
you
can
go
to
Google,
Images
and
type
in
pretty
much
anything
and
you'll
get
a
pretty
good
selection,
based
on
a
text
request
which
is
pretty
cool
to
a
lesser
extent.
The
supplies
to
video
as
well,
though,
of
course,
this
isn't
quite
as
advanced
yet
machine
learning
is
also
showing
up
in
places
like
cybersecurity.
B
Ibm's
watson
has
been
pushing
hard
in
the
medical
space
using
machine
learning
to
help
predict
outcomes
to
try
to
help
doctors,
make
better
decisions
and
then,
of
course,
there's
self-driving
cars,
which
is
the
machine
learning
task
that
probably
would
have
seemed
unthinkable
even
a
few
years
ago.
You
know
the
Jetsons
not
with
thinning
now
before
I
move
on
here.
I
want
to
make
something
clear:
I
have
a
friend
who
referred
to
this
as
the
Oprah
slide.
You.
B
Get
everybody
I
don't
want
to
give
the
impression
here
that
you
need
to
do
everything
at
once:
I'm
just
showing
you
some
of
the
places
that
it
is
being
used
out
there
going
forward.
You
are
better
off
just
kind
of
picking
one
little
piece
to
start
with
alright,
but
getting
back
to
what
I
was
saying.
These
are
not
just
weird
little
companies
you've
never
heard
of
working
on
this
stuff.
If
you
go
to
indeed.com
and
you
look
up
machine
learning
jobs,
you
can
see
the
companies
that
are
looking
for
people
who
understand
this
stuff.
B
The
day
I
was
putting
this
tag
together.
These
were
the
top
companies,
some
of
them
some
of
them.
You
know
you
expect
like
Amazon
and
Apple
and
Google
and
Facebook
and
Oracle
and
IBM,
and
even
the
staffing
companies
or
a
big
surprise,
but
look
at
this
number
two
and
number
three
on
this
list
are
JP
Morgan,
Chase
and
Goldman
Sachs.
B
You
can
go
and
look
in
the
AI
category
and
there
are
literally
hundreds
of
them
there
and-
and
why
is
that?
Certainly
not
because
there's
been
an
explosion
in
math,
PhDs
or
too
much
time
on
their
hands
now,
the
difference
is
that
there
is
now
there
are
now
easy
to
use
tool
kits
like
tensor
flow
and
scikit-learn
that
make
it
possible
for
anybody.
Who's
got
a
little
bit
of
programming
skill
to
put
together
real
AI
applications.
So
that's
why
it's
taking
off
and
because
people,
the
people
most
likely
to
be
interested
are
the
most
technical.
B
The
data
center
is
the
natural
place
for
this
to
be
showing
up
so
before
we
move
on
I
do
want
to
clear
up
one
little
bit
of
potential
confusion.
So
far,
we've
been
using
the
terms,
artificial
intelligence
and
machine
learning,
interchangeably
and
we'll
kind
of
do
that
throughout
the
webinar.
But
it's
not
a
hundred
percent
accurate.
B
But
then
it
also
includes
things
like
you
know:
complicated
decision,
trees
that
aren't
necessarily
machine
learning
which,
which
is
using
algorithms
and
potentially
training
to
teach
the
computer
to
handle
situations
it
hasn't
seen
before,
as
I
was
saying,
and
inside
of
machine
learning
is
another
term,
you
probably
have
heard
deep
learning,
which
is
a
type
of
machine
learning
that
involves
neural
networks,
which
is
another
term.
You
probably
heard
okay.
B
So
what
does
all
of
this
have
to
do
with
cloud
native
computing?
Well,
quite
a
lot.
Actually,
if
you
really
want
to
achieve
the
full
potential
of
what's
possible.
So
let's
take
a
quick
look
at
what
cloud
computing
really
means.
First,
you
have
an
environments
that
are
inherently
dynamic.
Cloud
native
applications
need
to
be
built
in
such
a
way
that
various
resources
can
disappear
and
be
replaced
at
any
time,
and
in
like
that,
humans
can't
be
expected
to
constantly
be
on
top
of
everything's
happening,
especially
once
you
get
beyond
a
certain
size.
B
Everything
has
to
happen
very
quickly
if
a
cluster
goes
down,
it
needs
to
be
replaced
immediately,
that
when
somebody
gets
a
call
and
logs
in
and
looks
around
and
decides
what
to
do,
which
is
why
we
have
continuous
automation.
Problems
have
to
be
corrected
without
human
intervention
and
systems
need
to
be
optimized,
even
as
they're
changing,
and
all
of
this
requires
observability.
B
You
have
to
know
what's
going
on
and
in
order
for
either
a
person
or
the
system
to
make
the
decisions
and
I
may
be
making
it
sound
like
there's,
not
enough
information,
but
that's
absolutely
the
opposite.
In
most
cases,
there's
too
much
information
and
that's
another
reason
why
we
need
machine
learning
and
other
AI
techniques
to
interpret
it.
B
And,
of
course
all
this
needs
to
be
done
efficiently
in
terms
of
performance
and
resource
usage
and
cost,
and,
and
that
also
requires
powerful
analysis
now
am
I
saying
that
none
of
this
can
be
done
without
machine
learning?
Absolutely
not.
You
could
do
all
of
this
without
machine
learning,
but
without
it
the
computer
is
only
as
smart
as
we
teach
it
to
be,
and,
as
these
environments
get
more
and
more
complicated,
that's
going
to
have
the
potential
to
become
a
problem.
B
Ok,
so
if
you
do
add
AI
and
machine
learning
elements
to
your
infrastructure,
what
do
you
get?
Well
you've
heard
of
continuous
delivery
infrastructure
where
you
can
get
continuous
improvements
to
your
cellphone.
We
like
to
think
of
this
as
intelligent
delivery
infrastructure
which
takes
things
to
the
next
level.
So,
let's
take
a
quick
look
at
what
that
means
now,
first,
off
an
intelligent
delivery.
Infrastructure
means
that
you
have
a
defined
architecture.
B
This
might
be
salt
manifests
or
puppet
scripts
or
whatever
it
takes,
but
we
are
definitely
talking
about
infrastructure
as
code
here,
so
the
so
that
these
automated
routines
know
what
they're
trying
to
achieve,
or,
at
the
very
least
what
they're
starting
out
with
it
also
gives
you
accountability
and
auditability.
You
also
get
an
architecture,
that's
flexible,
you
or
your
automation
can
change
it
at
any
time,
but
it's
controllable.
You
always
know
what
you
have
and
where
it
is
now
that
observability
becomes
intelligent
oversight.
B
You
can
have
systems
that
watch
over
what's
going
on
and
make
sure
that
issues
don't
become
problems
provide
optimization
and
healing
operations
as
appropriate
and
so
on.
And,
finally,
with
all
of
that
in
one
place,
it's
much
easier
for
you
to
get
appropriate
security,
because
AI
routines
can
catch
outliers
that
point
to
attacks
before
they
caused
damage.
So
that's
a
very
quick
version
and
and
a
very
quick
vision
of
what
we
could
get
by
incorporating
machine
learning
into
our
infrastructures.
B
Now,
let's
get
a
better
idea
of
how
machine
learning
works,
and
then
we
can
look
at
how
it
can
actually
do
those
things.
So,
let's
start
by
talking
about
some
of
the
different
kinds
of
machine
learning
that
are
out
there.
So
probably
the
simplest
to
understand
is
supervised
learning,
which
is
when
you
train
the
computer
using
data-
that's
already
labeled,
so
that
it
knows
how
to
recognize
similar
data
that
it
hasn't
seen.
Yet.
B
Typically,
there
are
two
types
of
problems
here:
there
are
regression
problems
where
you're
trying
to
determine
the
value
of
something
like
the
price
of
a
stock
on
a
given
day
in
the
future
or
classification
problems
where
you're
trying
to
determine
what
something
is
like.
Is
this
a
cancer
cell
or
not?
B
Some
examples
of
supervised
learning
include
image,
recognition,
speech,
recognition
and
medical
outcome
prediction.
Spam
filtering
stock
market
and
crypto
pricing
is
another.
One.
You'll
also
notice
that
all
of
these
are
situations
where
we
feed
the
system,
lots
and
lots
and
lots
of
training
data
so
that
it
knows
say
what
a
picture
of
a
bus
looks
like
and
then
they
can
recognize
other
buses,
but
this
training
can
be
extremely
extensive
and,
frankly,
quite
a
hassle.
So
there
are
also
a
couple
of
other
kinds
of
machine
learning
now
unsupervised.
B
Learning
is
where
we
just
give
the
computer
the
data
and
let
it
figure
out
the
patterns
without
us
now.
Obviously,
if
we
gave
an
untrained
routine,
a
pile
of
photos,
it's
not
going
to
be
able
to
figure
out
what
a
penguin
looks
like,
but
it
might
give
you
back
groupings
of
birds
versus
fish
or
it
might
give
you
back
black
animals
versus
white
animals
versus
multicolored
animals.
B
The
point
here
is
that
it
may
make
unexpected
connections,
which
is
good
because
sometimes
there's
situations
where
we
might
not
even
know
what
it
is
that
we're
looking
for
some
examples
of
unsupervised
learning
include
clustering,
where
the
system
tries
to
figure
out
what
various
bits
of
data
have
in
common,
such
as
trying
to
group
customers
into
categories,
so
you
can
sell
or
advertise
to
them
effectively
or
chemical
compounds,
so
you
might
be
looking
for
a
new
drug
or
figuring
out.
What's
going
on
in
a
new
environment
unsupervised.
B
B
So
reinforcement
learning
is
where,
rather
than
training
the
routine
using
label
data,
it
trains
itself
via
trial
and
error.
So
some
examples
are
like
I
said
just
about
anything
that
humans
do
or
self-driving
cars.
Obviously,
that's
a
simple
goal:
don't
crash
into
anything,
so
they
let
the
car
run
various
scenarios
and
figure
it
figures
out.
You
know
if
I'm
drifting
towards
a
light
pole,
I
should
probably
change
direction
or
a
car
stopping
in
front
of
me
I
should
stop
to
one
place.
B
It's
not
just
gaming.
There
are
more
practical
uses
for
reinforcement,
learning
such
as
scheduling
classically.
They
refer
to
this
as
the
Traveling
Salesman
problem,
when
you
need
to
figure
out
the
best
route
for
the
salesman
to
travel,
to
minimize
his
expenses
and
maximize
his
or
her
profits.
But
you
can
also
think
of
this
in
terms
of
network
latency,
since
we're
trying
to
talk
specifically
about
the
data
center.
C
B
A
B
Correct
and
and
I
was
gonna
mention
that
a
little
bit
a
little
bit
later,
this
we
have
so
much
data
coming
in
from
like
the
Internet
of
Things,
and
all
of
these
metrics,
like
that.
We,
as
a
storage
expert
friend
of
mine,
said
absolutely
drowning
in
this
data.
If
we
don't
figure
out
ways
to
eliminate
things
that
we
do
not
need.
C
B
Yes,
absolutely
absolutely
crucial.
We
don't
need
to
store
everything
and,
as
I
was
going
to
say
later,
it
scares
the
heck
out
of
me
because
I
am
somewhat
of
a
date
or
a
hoarder
and
I
hate
to
let
anything
go
because
I'm
sure
that
the
minute
I
do
I'm
gonna
need
it,
but
we
have
to
we
have
to
and
and
that's
some
place
where
AI
can
really
a
I.
Am
machine.
Learning
can
really
shine.
B
Right
so,
let's
talk
about
some
of
the
difficulties
in
running
an
efficient
data
center
and
how
machine
learning
and
AI
techniques
can
help
us.
There
now
remember,
I'm,
not
saying
jump
in
and
do
all
of
these
things
from
the
start,
I'm
just
trying
to
show
you
that
there
are
opportunities
to
improve
things
with
AI
right
under
your
nose.
If
you
want
to
go
that
way,
all
right
so
starting
at
the
beginning
of
the
process.
B
Listen
I've
been
a
developer
for
years
and
I
know
I,
don't
want
me
making
decisions
about
the
best
way
to
set
up
a
data
center,
and
the
situation
is
even
worse
now
that
we
have
Micro
Services,
because
once
you've
got
the
system
itself
set
up,
there
are
still
so
many
theoretical
choices
for
developers.
There
are
tons
of
opportunities
to
shoot
yourself
in
the
foot
and
have
problems
and,
of
course,
with
dynamic
environments.
We
have
lots
of
opportunity
for
configuration
configuration
drift.
B
This
is
either
going
to
get
worse
with
immutable
architecture,
because
it's
all
baked
into
instances
that
are
going
to
break
when
things
change
or
is
going
to
get
better,
because
people
know
in
two
instances
that
are
gonna
break
when
things
change
and
they're
gonna
take
care
of
it
personally.
I
think
it's
gonna
be
worse.
Unless
we
get
better
processes
in
place
and
really
you
can
think
of
a
lot
of
these
things
like
playing
a
game
of
broken
telephone,
you
remember
the
game.
B
Will
you
whisper
something
to
someone
and
they
whisper
it
to
someone
else
and
by
the
end
of
the
line,
you
don't
recognize
what
comes
out
it's
funny
when
it's
a
game?
It's
not
so
funny
when
your
entire
infrastructure
is
in
metaphorical
flames.
So
how
do
we
add
intelligence,
this
process?
Well,
for
one
thing,
a
machine
learning
routine
can
look
at
previous
deployments
and
potentially
figure
out
well.
This
is
similar
to
this
app,
which
did
best
when
it
was
deployed.
B
And
once
things
are
configured,
you
want
to
get
that
optimization
right,
which
can
be
tough
because
in
this
distributed
dynamic
environment,
there
are
ton
of
factors
to
consider,
and
some
of
them
are
probably
not
even
visible
to
the
developer,
much
less
controllable
by
him
or
her,
and
even
if
you
do
get
a
perfectly
optimized
environment
or
deployment
once
things
change
like
a
cluster,
goes
down
or
there's
a
slowdown
somewhere.
Your
optimization
goes
right
out
the
window.
B
B
So,
for
example,
if
your
company
ran
a
Super
Bowl
ad,
the
system
might
take
several
factors
into
account
and
realize
a
slight
increase
during
the
commercial
wasn't
just
the
blip
and
scale
up
before
a
huge
increase
hits
when
the
commercial
zone
so
or
your
system
might
expect
more
people
to
be
watching
that
game.
In
Philadelphia
and
Boston
based
on
patterns,
you
picked
up
beforehand
and
redirects
resources.
Accordingly,
sometimes
the
optimization
isn't
even
about
performance,
but
it's
about
cost
with
the
rise
of
public
cloud
and
it
pay-as-you-go
strategy
making
sure
you're
not
paying
for
something.
B
You
don't
is
really
important.
To
give
you
an
example,
I
myself
recently
got
a
bill
from
AWS
for
a
hundred
and
fifty
six
dollars
for
load,
balancers,
I
didn't
even
realize.
I
had
spun
up
when
testing
something
the
previous
month,
so
there
are
way
more
moving
pieces
than
most
people
can
keep
track
of,
and
even
if
you
do
keep
track
of
them,
the
prices
are
always
changing.
Fortunately,
the
unit
prices
generally
are
going
down,
rather
than
up
as
all
these
companies
fight
to
get
you
locked
into
their
own
cloud.
B
But
the
fact
remains
that
there
are
often
times
when
you
are
not
on
the
most
cost-effective
resources,
even
though
you
were
yesterday,
and
we
can't
overlook
the
obvious.
This
gets
to
Christians
question
with
all
of
this
data
flying
around
there's
more
every
day
and
we
are
running
into
these
huge
storage
costs.
B
There's
one
auto
company
that
generates
so
much
data
on
the
test
drive.
They
can't
transmit
it
to
a
central
location
fast
enough
and
they
have
to
send
hard
drives
in
a
chase
fan
that
just
blows
my
mind
so
in
terms
of
how
machine
learning
can
help.
Of
course,
one
thing
that
it's
very
good
is
keeping
track
of
all
these
multiple
parameters.
So
you
can
figure
out,
if
you're,
better
off
on
this
system
or
in
that
one
based
on
usage
in
ways
that
maybe
aren't
obvious.
B
For
example,
maybe
computers
cheaper
at
provider
a
but
you
don't
use
a
lot
of
computer.
You
use
a
lot
of
storage
and
you
transmit
a
lot
of
it
so
you're
better
off
it
provided
B
and
I'm
again,
not
saying
you
couldn't
code,
something
to
figure
this
out
without
machine
learning.
Of
course
I'm
just
saying
that
it
lends
itself
to
that,
especially
in
these
changing
environments
and
as
far
as
the
storage
issue
we
talked
about
this
earlier,
you
know
just
keep
what
you
need
and
AI
and
ml
can
help
you
figure
out
what
that
is.
B
Okay,
so
you've
got
everything
configured
you've
got
it
I
optimized
you're,
not
paying
more
for
it
than
you
need
to,
but
things
can
still
break
harder,
goes
down
software
configurations
drift.
We
talked
about
that
earlier.
Sometimes
you
have
data
corruption
going
on
and
you
might
not
even
know
you
might
have
silent
data
corruption
caused
by
bad
firmware
or
even
loud
noises
or
cosmic
radiation
silence.
The
data
corruption
doesn't
get
noticed
by
the
operating
system
and
you
might
not
find
it
until
it's
much
much
too
late.
Remember
that
broken
telephone
analysis.
B
This
could
be
a
huge
issue
before
you
even
know
it.
You
might
need
to
get
a
head
start
on
running
out
of
storage
or
memory
or
other
resources.
So,
of
course,
you
can
do
auto
scaling
without
machine
learning
as
you've
been
saying,
but
by
adding
it
in
you
have
the
ability
to
do
more
sophisticated
versions
of
it.
You
can
also
let
the
system
detect
patterns
that
might
indicate
a
disk
is
going
to
go
bad
in
the
near
future
or
the
hardware
resources
are
gonna.
B
Issues
are
gonna
pop
up,
so
you
can
move
data
or
resources
and
replace
that
Hardware
before
it
becomes
a
problem.
Machine
learning
is
also
good
for
detecting
non
patterns
or
anomalies.
That
indicates
something
isn't
right,
whether
again
Hardware
going
bad
software
data,
that's
been
corrupted
and,
of
course,
there
are
security
issues
which
is
somewhere
that
machine
learning
is
already
in
play.
I
think
we're
all
pretty
familiar
with
distributed.
B
You
might
also
use
it
to
find
situation
where
somebody's
doing
something
is
just
not
right
and,
of
course,
pattern.
Recognition
can
help
you
detect
when
a
DDoS
is
headed
your
way,
so
you
can
take
steps
to
prevent
issues
in
so
much
as
that
is
possible.
It
might
also
help
you
locate
the
source
of
the
attack
and
block
it.
Obviously,
that's
tougher
with
the
distributed
attack
like
the
Mirai
botnet,
were
millions
of
IOT
devices
were
taken
over
and
used
for
this
kind
of
thing,
but
that's
the
whole
point.
B
A
machine
learning
routine
might
quickly
figure
out
that
that's
what's
happening
and
help
you
filter
out
that
traffic.
Okay.
So
now
that
we
talked
about
some
of
the
things
that
machine
learning
can
do
for
I
want
to
show
you
how
easy
it
is
to
actually
do
this
before
I.
Do
that,
though,
I
want
to
tell
you
a
really
quick
story
about
15-20
years
ago.
I
was
writing
for
a
very
large,
very
sophisticated
company
and
I
won't
name
them
for
reasons
that
will
become
obvious
in
a
minute.
B
They
invited
me
down
to
their
offices
in
southwest
to
see
a
new
program
that
they
were
working
on
and
they
were
trying
to
create
software
that
was
self-configuring
self,
optimizing,
self
healing
and
self
protecting.
Now,
as
I
was
typing,
this
into
the
tech,
I
realized
pretty
much.
What
I
was
describing
in
the
last
section,
but
at
time
I
kind
of
looked
at
them
and
said
well,
if
it's
self
configuring
self
optimizing
so
feeling
and
self
protecting?
How
do
you
turn
it
off?
B
B
Fortunately,
as
I've
been
saying
it's
not
as
tough
as
it
sounds,
so,
let's
dive
in
ok.
So
in
this
example,
we're
gonna
try
and
create
a
Python
script
that
uses
machine
learning
to
predict
the
best
approximation
for
a
set
of
points.
In
this
case,
it's
the
relationship
between
the
length
of
an
article
and
the
number
of
shares
that
article
got
and,
of
course
there
are
any
number
of
lines
that
we
can
a
draw
that
approximate
this
data
and
some
are
going
to
be
more
accurate
than
others.
B
But
how
can
we
quantify
how
accurate
each
line
is?
Well,
what
we're
going
to
do
is
we're
going
to
define
the
error
for
each
line.
So
let's
say
we
have
just
these
five
points.
So,
if
we're
going
to
draw
a
line,
each
point
would
be
a
set
distance
from
that
line
represented
by
these
green
lines
and,
of
course,
the
smaller
the
total
distance
or
the
less
green,
the
better
and
approximation
that
one
it
is.
But
how
do
we
figure
out
that
distance?
Well,
this
is
a
simple
XY
coordinate
system,
so
we
have
for
each
point.
B
We
have
the
Y
value
and
we
have
I
predicted,
which
is
what
the
line
thought.
Why
would
be
for
that
particular
x-values,
so
the
length
of
each
of
those
green
lines
is
going
to
be
absolute
value
of
y
minus
y
predicted
we're
using
the
absolute
value
because
we
don't
want
any
negative
distances.
So
the
total
error
for
this
line
is
going
to
be
the
sum
of
all
the
individual
little
green
lines
for
the
sum
of
the
absolute
value
of
y
minus
y
predicted.
But
how
do
we
know
what
Y
predicted
is?
B
B
Now,
of
course,
we
don't
want
to
run
our
data
center
and
just
guessing,
but
fortunately
there's
an
organized
way
to
do
this.
Now,
let's
say
I
told
you
that
I
was
thinking
of
a
number
between
one
and
a
hundred,
and
if
you
guessed
I
tell
you
if
you
were
too
high
or
too
low.
So
if
you're
being
logical,
you'd,
probably
start
by
guessing
50
and
I,
tell
you
you
were
too
low,
so
you'd
know
it's
being
51
and
100,
then
you'd
probably
get
slack
or
75,
and
then
I
tell
you
that
you
were
too
low.
B
B
So
if
we
were
going
to
graph
the
error
based
on
the
guess,
you
can
see
here
that
on
the
left,
the
slope
is
negative
because
we're
moving
towards
the
right
answer
and
on
the
right,
it's
positive
as
we're
moving
away
from
the
right
answer
right
at
the
right
answer.
The
slope
of
this
line
is
zero
and
and
that's
what
we're
trying
to
find
we're
trying
to
find
the
place
where
the
slope
would
be
a
zero,
because
the
error
is
at
its
minimum.
B
So
if
I
were
to
vastly
oversimplify
this,
if
you
know
the
math
here,
yes
I
know,
I
am
vastly
oversimplifying
this.
The
error
might
look
something
like
this
and
we're
going
to
try
and
make
sure
that
we're
always
moving
downhill,
so
to
speak
so
that
we
can
get
down
here
to
the
bottom.
One
thing
to
keep
in
mind,
though,
is
the
size
of
the
steps
we're
taking.
If
we
were
taking
steps
that
were
too
small,
they
take
a
really
long
time
to
get
to
the
bottom.
B
If
we
take
steps
that
are
too
big,
we
could
wind
up
missing
the
bottom
altogether
and
just
bounce
back
and
forth
between
the
two
sides.
So
if
we
put
all
that
together,
we
get
a
technique
called
gradient
descent,
which
is
basically
an
algorithm
that
tries
to
find
the
parameters
that
are
going
to
give
us
that
minimum
error
now
to
do
that.
We're
going
to
need
a
couple
of
things
now.
Obviously
I
don't
have
time
to
take
you
through
all
the
setup
and
everything.
B
But
if
you
fill
out
the
former's
for
machine
learning
for
mere
mortals,
I'm,
probably
gonna
put
together
a
blog
post
on
how
to
get
things
that
absolutely
you'll
get
notified
about
that.
So,
though,
the
pieces
are
the
data
of
course,
which
in
this
case
is
just
a
CSV
file,
Python
three
tensorflow,
which
is
an
open
source,
machine
learning
library
and
it's
all
running
in
this
case
on
Ubuntu.
B
So
if
we're,
if
we
look
at
what
we're
actually
going
to
do
here,
it's
pretty
straightforward,
we'll
read
the
data,
then
we'll
define
the
model
which
in
this
case
is
that
y
equals
MX
plus
B
and
we'll
define
the
error
in
the
optimizer
and
then
we'll
repeatedly
run
the
optimizer
until
we
get
as
close
as
we
can
to
the
right
answer,
all
right.
So,
let's
go
over
to
my
bin
to
machine
which
probably
is
not
set
up
the
right
window,
but
we
will
okay.
That's.
A
C
B
So
if
we
go
here,
you
ignore
that
stuff.
Like
I
said
we
will
come
over
here
to
our
file.
This
is
what
I
get
for
not
proofing
the
section
Hana
time.
This
is
how
you
know.
It's
live
folks,
all
right.
So
now,
I
already
I
already
generated
this
data.
I
generated
this
data
myself,
so
I
would
know
if
we
were
getting
the
right
answer.
So
the
shares
in
this
case
are
going
to
be
point
three
seven
times
the
article
length
plus
12
plus
or
minus
some
random
amount.
B
So
we
know
that
the
right
answer
is
going
to
be
well
come
back
here
come
back
here.
The
right
answer
is
going
to
be
y,
equals
point
three:
seven
X
plus
twelve,
so
M
equals
point
three
seven
and
B
equals
tall.
So
let's
actually
go
ahead
and
do
this
all
right,
so
we'll
start
out
by
just
importing
the
pieces
that
we
need
this
bit
here
is
just
to
suppress
some
warning.
So
just
ignore
that
and
then
we'll
go
ahead
and
we
will
load
the
data.
C
B
B
All
right
so
we're
defining
the
model.
You
remember
we're
looking
for
y
equals,
MX
plus
B,
so
we're
going
to
define
our
initial
guesses
for
M
and
B,
and
just
to
show
you
that
I
haven't
fixed
this
in
any
way
we're
going
to
use
random
numbers,
then
what
we're
doing
is
creating
a
single
tensorflow
variable
W,
which
is
basically
an
array
of
M
and
B.
Tensorflow
barriers
are
special
because
tensorflow
knows
that
it
can
update
them,
as
it
runs
the
optimizer
to
try
and
get
closer
to
the
right
answer.
B
Now,
then,
for
the
model
itself,
we
basically
need
three
values.
We
need
the
article
length
the
shares
and
the
number
of
shares.
The
model
predicts
that
we're
going
to
have.
The
first
two
are
tensorflow
placeholders,
which
means
that
when
we
run
the
optimizer
we're
going
to
feed
in
the
data
for
them
and
then
the
predicted
shares
is
just
MX,
plus
B,
so
W
0
times
the
article
length
plus
W
want
okay.
So
now
we
need
to
get
the
error.
B
B
First,
we're
going
to
choose
the
step
size
now.
Remember
we
talked
about
that
earlier.
You
can
experiment
with
different
step
sizes,
I'm
going
to
use
point
0
to
5,
because
about
as
big
as
I
can
get
without
having
a
problem,
and
then
we
create
the
optimizer,
which
is
the
gradient
descent
optimizer.
This
is
the
we
talked
about
the
question
earlier.
I
will
get
to
that
later.
We
are
giving
it
the
step
size
and
we're
telling
it.
Then
we
want
to
minimize
the
error
just
make
it
so
you
can
see
more
hair
there.
You
go
alright!
B
So
now
that
we've
created
all
these
variables,
we
want
to
go
ahead
and
define
the
model
as
this
collection
of
variables
and
we're
ready
to
run
it
okay.
So
we
define
the
model
and
we
will
run
it
now.
Remember
we
we
specified
we're
gonna,
create
a
session
to
run
all
this
okay,
so
we've
got
with
session
now.
Remember
we
specified
our
initial
variables
as
random
numbers,
so
it
would
help
if
we
printed
them
out.
So
we
know
where
we're
starting.
B
Okay,
we're
creating
this
data
to
use.
Variable
right
here
and
then
we're
feeding
it
the
data
that
we
got
from
our
CSV
file.
Okay,
so
when
all
that's
done,
we
will
print
out
the
value
that
we
got
and
then
down
at
the
bottom.
I
have
a
routine
that
goes
ahead
and
prints
out
the
graph.
So
let's
go
ahead
and
run
that
if
I
can
get
to
the
right
screen.
B
C
B
It's
supposed
to
be
terrible
at
this
point
because
we
are
supposed
to
have
only
one
guest
but
apparently
we
ran
through
and
we
did
all
of
them.
Yes,
we
did
remember
I,
said
I,
didn't
rehearse
this
part
there
you
go
so
if
you
only
ran
one
guest,
it
would
be
terrible.
Okay,
you
would
not
get
a
good
answer
at
all,
so
instead,
what
you
want
to
do
is
you
want
to
go
ahead
and
run
it
multiple
times.
B
So
that
is
what
we
are
doing.
We
are
actually
running
it
through
20,000
times,
as
you
can
see,
and
so
we've
we've
set
the
data.
We
are
running
the
optimizer,
the
error
and
the
weight
and
then
we're
out
we're
out
putting
them
here,
but
we
don't
need
the
optimizer.
So
that's
what
that
underscore
is
it's
just
a
place
only
and
we're
feeding
it.
Data
to
use
and
every
thousand
iterations
were
printing
it
out
and
we
get
the
answer
and
you
notice
it
came
to
the
answer
really
really
fast.
B
B
B
Okay,
we
started
out
with
a
predicted
model
of
sixteen
point:
seven
one,
six
plus
14
a
huge
error,
the
weights
less
gradually
than
normally
changed,
and
we
wound
up
with
where
we
needed
to
be
okay.
So
let
us
see
where
we
are
so
basically
just
a
few
steps
here
really
and
we
did
some
actual
real
machine
learning.
You
saw,
we
ran
it
multiple
times
it
can't
kept
coming
up
with
the
answer.
So
how
do
you
put
this?
To
use?
B
Okay
main
thing
to
remember
here:
is
you
need
to
take
small
steps?
You
can't
go
from.
You
know
black
and
white
tintypes
to
VR
in
in
one
step.
You
know
you
want
to
take.
You
want
to
take
little
steps
in
between,
and
one
great
place
to
start
is
with
analytics,
because
you
are
probably
already
doing
some
work
there.
You've
already
got
the
data.
You've
already
got
processes
in
place
to
determine
what
you're
looking
for
and
what
to
do
once
you
find
it
and
you
can
get
some
real
results.
B
The
most
important
thing
is
to
know
what
it
is
you're
trying
to
accomplish.
Just
the
best
machine
learning
routine
in
the
world
isn't
going
to
help
you.
If
you
don't
know
what
you
wanted
to
do,
alright
and
that's
it,
so
we
are
ready
for
Q&A.
So
if
you
have
not
gotten
your
questions
in
now
is
now
is
the
time
also
you
don't
forget
that
you
can
go
ahead
and
grab
grab.
That
course,
alright.
So
remember,
if
you're
not
doing
this
live
click
the
link
anyway,
please
hopefully
we'll
have
something
for
you
there
as
well.
A
C
A
How
do
we
account
for
the
ways
that
machine
learning
determines
how
to
do
things,
but
into
which
we
have
no
visibility?
Given
that
machine
learning
has
been
documents,
it's
come
up
with
surprisingly
biased,
even
racist
rules
when
learning
how
to
do
things,
this
would
see
be
something
we
should
address
so
I.
B
B
Well,
we'll
put
it
in
with
with
the
notes
afterwards,
but
it's
very
true
and
I
think
it's
really
important
that
researchers
and
even
people
who
are
just
working
with
the
data
are
confident
of
this,
because
your
your
data
is
very
much
going
to
affect
whether
you
have
these
problems.
I'll
give
you
I'll,
give
you
an
example.
B
If
you
have
a
machine
learning
routine,
that
compares
that
compares
people
who
are
applying
for
a
for
a
let's
say
for
a
promotion
at
a
big
company
and
you've
got
a
woman
and
you've
got
a
man
and
the
machine
learning
routine,
compares
these
two
candidates
and
decides
that
the
man
would
be
more
successful
and
the
reason
that
it
does.
That
is
because
it
has
more
examples
of
successful
men,
but
it's
not
taking
into
account
the
fact
that
that's
because
it's
been
difficult
for
women
to
get
into
those
positions
in
the
first
place.
B
So
this
is
something
that
humans
are
definitely
going
to
have
to
really
take
in
hand
to
make
sure
that
we
are
keeping
these
things
in
in
check.
I
I
can't
I
can't
stress
that
enough,
they're
using
machine
learning
now
for
sentencing
in
criminal
cases
and
it's
exactly
the
same
problem.
So
if
you're
looking
for
me
to
say,
oh,
that's
not
going
to
happen,
I
I'm
not
going
to
do
that
because
it's
it's
happening,
that's
why
we
all
need
to
understand
this
now.
A
Excellent
okay:
we
got
one
more
opening
up.
Let's
have
a
look,
Oh
lots
and
lots
of
people
asking
to
share
the
slides,
we're
not
gonna,
sir
two
slides
with
your
email
addresses
because,
as
you
can
probably
figure
out,
that's
gonna
be
a
real
pain.
So
instead,
what
we're
gonna
do
is
take
the
recording
of
this
and
put
it
online
at
that
URL
that
I've
just
dropped
into
the
chat
and
next
to
it
we
will
attach
a
PDF
of
the
slides.
A
B
B
Some
techniques
involve
momentum,
so
like
the
steeper,
the
hill,
so
to
speak,
the
faster
you
go
which
allows
you
to
get
past
what
are
called
these
local
minima,
where
it's
not
really
the
the
minimum,
but
you
know
it
looks
like
it
kind
of
where
you
are.
There
are
techniques
that
let
you
kind
of
look
forward.
The
problem
with
that
is
now
you've
got
momentum.
You
might
wind
up
overshooting
the
real
minimum.
B
A
Cool,
thank
you.
I
see
some
people
raising
hands.
I,
put
a
comment
in
the
chat
earlier.
I
can't
give
you
the
mic
with
a
group
of
this
size.
So
if
you're
raising
your
hand,
please
turn
that
into
a
question
and
stick
in
the
chat
we'll
make
sure
that
it
gets
answered
all
right
next
questions,
let's
sort
of
look
I
guess
we
can
go
to
the
top.
Now,
with
some
questions
from
earlier
I
guess
you
could
choose
yourself,
there's
three
or
four
or
five
in
there.
We've
got
about
six
minutes
left
all.
B
How
do
you
keep
algorithms
from
increasingly
self
validating
when
anomalies
might
actually
be
beneficial
to
making
your
dataset
more
accurate?
There
is
a
technique
called
cross-validation
winner,
these
okay.
First
off,
there
is
a
balance
between
relying
too
much
on
the
anomalies
and
doing
what's
called
overfitting,
where
you
are
to
specifically
on
your
training
data,
and
you
don't
generalize
well
and
underfitting,
where
you
don't
enough
of
the
details
are
not
very
and
one
way
to
kind
of
get
between.
Those
two
things
is
to
do
what
they
normally.
What
you
do
is
you
would
like
maybe
take.
B
A
B
B
A
B
How
do
you
know
what
step
size
to
use
yeah
so
there's
a
couple
of
ways
that
you
can
do
that
things
like
step
size
and
and
when
you
get
into
like
normal
networks,
the
the
size
and
shape
of
your
neural
network
and
all
that
stuff.
Those
are
called
hyper
parameters
because
there
are
parameters
about
your
parameters
and
there
are.
There
are
different
techniques
that
you
can
use
to
find
the
best
hyper
parameters.
B
You
know
I
mean
you
could
just
guess,
and
you
could
do
trial
and
error.
You
can
do
what
they
call
a
grid,
search
where
you
basically
just
run
them
all
all
the
different
combinations
and
you
know
see
which
one
is
best.
Obviously,
that's
not
going
to
work
well
with
a
big
problem.
You
can
do
random
search,
which
lets
you
just
sort
of
run
a
few
and
pick
the
best
one
and
then
keep
running
more
of
the
simulations
and
while
you're
actually
running
your
problem,
let's
see
if
you
can
find
something
better.
B
So
there
are,
there
are
ways
to
kind
of
work
on
that
problem.
Let's
see
I
see
someone
is
asking.
Does
time-based
database
is
common
on
the
machine
learning
I'm
basic
I
assume
the
question
is,
actually,
can
you
do
machine
learning
with
time
based
databases,
and
the
answer
is
yes,
there
is
a
technique
called
recurrent
neural
networks
which
take
into
account
things
that
happen
in
the
past,
and
that
is
that's
the
kind
of
thing
you
would
use
for
like
stock
prediction
and
crypto
pricing.
B
B
When
should
we
be
using
radiant
anytime,
you
have
your
you're
trying
to
predict
like
a
value.
You
are
probably
going
to
be
using
gradient
descent
at
some
point,
even
within
even
within
neural
networks.
Typically
behind
the
scenes,
it's
some
form
of
gradient
descent.
In
that
all
there
is
no
there.
There
are
other
techniques
as
well,
but
it's
it's
extremely
common.