►
From YouTube: NAB Python 3 Part 1
Description
Broadcasted live on Twitch -- Watch live at https://www.twitch.tv/rhyolight_
A
A
A
A
So,
if
your
so,
the
problem
is
with
an
a
benchmark
like
this,
you
want
to
be
able
to
run
it
in
any
environment,
so
I
might
so
I
might
make
an
R
solution
for
it.
I
might
make
a
Python
solution,
a
C++
pollution
solution,
Lisp
I,
don't
know
whatever
elixir
and
you
can't
you
basically
to
get
your
contribution
or
your
detector
into
the
system.
A
Let's
do
a
quick
review
of
this,
so
the
immense
anomaly
benchmark
contains
data
and
scripts.
This
is
a
new
benchmark
in
the
difference
between
this
and
most
machine
learning.
Benchmarks
is
this:
is
a
time-based
benchmark.
It's
an
unsupervised,
that's
assuming
you're
unsupervised,
so
we're
we're
testing.
How
well
learning
algorithms
can
predict
anomalies
an
unlabeled
streaming
sequential
data,
which
is,
of
course,
what
is
pretty
good
at
so
we
wouldn't,
and
there
were
no
benchmarks
that
we
could
compare
ourselves
with
anyone
else.
So
we
created
this
so
there's
over
50,
labeled,
real-world
and
artificial
time.
A
Series
data
sets
and
they're
all
scalar
values
over
time,
and
so
in
this
repository,
are
the
tools
allow
you
to
easily
run
this
benchmark
on
your
using
your
own
anomaly,
detection,
algorithms,
and
what
we're
gonna
do
is
separate
this,
because
this
is
written
in
Python,
2
and
I.
Think
we
made
some
assumptions
about
the
PI.
There
would
be
a
Python
2
runtime
for
running
these
detectors.
So
if
you
want
to
create
an
entry
in
the
benchmark,
you
create
a
detector.
I
haven't
done
this,
but
but
there
there's
a
research
paper
on
this
here.
A
A
Oh
and
here's
the
link
to
the
NAB
white
paper,
I,
don't
know
what
this
is
exactly.
Is
this
this
I
thought
that
was
the
NAMM
boy
paper
paper?
No,
no!
This
is
different,
so
this
is
probably
very
similar,
but
I
think
it
specifically
tells
you
how
NAB
works.
I
don't
want
to
get
too
too
deep
into
understanding
that
I
have
a
general
under
idea
of
how
it
works.
But
what
I
want
to
do
is
review
the
work
that
Yoon
has
done
to
split
this
up
into
two
repositories.
A
A
A
If
someone
submits
a
new
entry,
a
new
detector,
then
we
essentially
have
to
I
mean
they
have
to
run
it
in
their
environment,
create
these
files
that
represent
the
scores
or
represent
the
anomalies
that
they
found
in
all
of
these
datasets
and
then
run
nab,
which
has
a
scoring
code
that
will
score
all
of
these
and
do
its
logic.
So
nabs
got
specific,
scoring
like
for
false.
False
positives
can
be
treated
differently
than
false
negatives,
so
it
has
some
specific
rules
about
scoring
so
there's
the
standard
profile
at
which
conservative
weights,
positives
and
negatives
they'll.
A
Do
these
all
weight,
false
positives,
false
negatives
differently
and
then
make
true
negatives
right
right.
The
other
true
pop,
true
negatives,
correct
predictions
of
versus,
missed,
predict,
missed
anomalies,
and
so
then
there's
like
a
lower
reward
for
false
positive
and
then
a
low
reward
reward
low
for
false
negative
version
of
these,
and
this
is
the
scoreboard.
A
So,
okay,
this
this
is
the
inspected
update
to
the
readme
okay,
so
there's
the
main
paper:
okay,
that's
I'm,
supervised
the
science
of
direct
one
main
paper
covering
nab
in
the
Mintos
HTM
based
anomaly,
detection
algorithm.
They
have
white
paper
and
then
the
original
publication.
There's
like
how
many
publications
here,
one
two
three
four
got
this
one:
this
one,
which
is
the
wit,
read
me
that
points
to
these.
Actually
going
to
this
evaluating
real-time
anomaly
detection,
this
paper
unsupervised.
Okay,
this
is
another.
This
is
the
same
one.
This
is
science
direct
yeah.
A
A
A
Okay,
we
encourage
you
to
publish
your
results
on
nab
and
share
them
with
us.
Please
cite
the
following
publication:
I
guess
these
sort
of
need
this
so
I,
guess
that's
why
the
extra
paper
like
is
there
because,
but
maybe
this
would
be
better
if
it
were
I.
Don't
know,
scoreboard
caveats.
Look
at
all
these!
A
Please
see
the
wiki
section
on
contributing
algorithms
for
us
to
consider
adding
your
algorithm
to
the
nav
repo.
It
must
meet
the
following
criteria:
open
source
work
with
streaming
data
processed
data
in
real
time.
So
that's
the
that's
important
thing
in
in
real
time.
You
can't
batch
it.
It's
like
one
bit
of
the
data
at
a
time
and
you
get
the
next
time
step
and
then
you
give
us
an
anomaly
indication
right
then,
next
time
step
not
only
indication,
it's
so
that's
different
than
most
deep
learning
systems.
A
A
A
A
Extensive
live,
and
so
if
it
decides
to
apply
an
aggregation
and
update
its
neck
for
its
next
print
update
its
model
for
its
next
prediction.
It's
gonna
have
to
go
back
and
process
and
you
know
keep
access
to
that
data
and
reprocess
it
in
a
batch
format.
You
know
so
it
would
be
a
pipeline
that
produces
the
answer
after
the
pipeline
processes.
It
well
I
guess
you
could
say
it
could
it
can.
It
can
go
back
and
reapply
a
certain
aggregation
and
relearn
I.
A
Don't
think,
there's
anything
that
keeps
any
of
these
algorithms
from
storing
data
and
then
and
then
running
back
through
it
as
just
to
update
its
model
as
it
goes
along
and
I.
Think
that's
essentially
what
LST
M
does
it
runs
batches
and
then
it
adjusts
in
real-time
and
reruns
the
batches
so
that
they
can
update
its
model
in
real
time.
A
Just
it
can't
run
in
bash,
that's
the
things
necessary.
The
algorithms
are
computationally
efficient
to
processing
streaming
data.
Oh
then,
the
following
algorithms
have
been
tested
on
NAB
and
do
not
meet
this
criteria.
Oh
wow
I
have
a
stand-up
meeting
almost
forgot
about
I
can't
believe
it's
only
10
o'clock
I
have
a
stand-up
meeting
in
10
minutes,
so
bear
with
me.