►
From YouTube: Machine Learning Applications to Collider Physics
Description
Vinicius Mikuni (NERSC)
Machine Learning Applications to Collider Physics
A
Looking
at
this
animation,
where
these
new
particles
that
are
generated,
they
encode
the
information
on
how
these
protons
they
interact
with
each
other
and
our
knowledge
about
how
the
particles
they
interact
with
one
another
and
in
order
to
be
able
to
take
theories
that
we
have
currently
that
can
describe
how
these
particles
interact.
We
need
to
have
a
very
large
amount
of
experimental
data
of
the
sparkle
collisions
in
order
to
have
a
good
way
to
compare
the
two
together.
A
So
we
do
that
with
simulations
and
we
take
those
simulations.
We
also
pass
them
through
a
simulation
of
the
detector
material
and
how
the
detector
would
measure
each
of
the
particles
that
we
predict
in
your
theory
and
then
the
same
data
analysis
methods
that
we
use
with
real
data.
We
apply
that
to
the
simulation,
then
codes,
our
knowledge,
and
we
compare
the
two
to
see
how
well
can
we
do
the
comparison
between
well?
We
know
before
we
actually
observe
so
the
topic
that
I
want
to
talk
today.
A
They
are
going
to
cover
three
different
ideas.
The
first
one
that
is
called
here
on
the
top-
this
is
called
a
surrogate
model,
is
how
we
can
make
simulations
of
the
detector
response
that
are
faster
compared
to
the
full
simulation
chain
that
we
use
currently
in
your
experiments.
The
second
application
is
called
the
unfolding,
which
is
roughly
the
opposite
directions
that
the
convolution
problem,
where
we
basically
want
to
take
data
that
has
been
measured
by
our
detector
and
converts
it
back
to
this
realm
of
theory.
A
Predictions
where
we
have
particles
that
are
predicted
by
the
theories
before
any
detector
effects
and
the
last
application
that
I
want
to
talk
about
is
on
the
parts
of
the
data
analysis
and
how
we
can
take
the
data
that
we
observe
from
particle
collisions
and
try
to
identify
things
there
that
we
do
not
know
what
it
is.
But
we
just
know
that
something
might
be
there
that
we
don't
understand,
which
is
basically
a
normal
detection
and
how
to
try
to
find
new
physics.
Even
though
you
don't
know
how
new
physics
should
look
like.
A
So
you
have
to
do
a
precise
image
of
these
interactions,
which
corresponds
to
thousands
up
to
millions
of
detector
without
channels
that
we
need
to
do
the
full
simulation
of
to
give
you
a
perspective,
if
take
the
two
main
experiments
at
the
LHC,
the
atlas,
experiments
and
the
CMS
experiments,
and
if
you
take
their
previous
run
times
on
taking
data
collected
by
in
the
years
of
2016
to
2018,
you'll,
see
the
40
of
the
whole
computing
power
of
the
whole
carburation
was
dedicated
to
simulate
events
and
in
particular
the
part
that
takes
the
longest
is
simulates
the
response
of
their
detectors
and
that
already
takes
a
lot
of
computing
power.
A
But
if
you
look
into
the
future,
where
we
plan
to
change
our
accelerator
facilities
in
order
to
have
even
more
particle
collisions
in
a
shorter
amount
of
time,
then
that
means
that
we
have
even
more
data
that
is
going
to
be
collected
and
we
need
to
select
even
more
a
particle
collisions
in
order
to
match
the
amount
of
experimental
observations
that
we
have.
And
if
you
just
try
to
extrapolate
the
budgets
they
currently
have
for
these
simulations
into
the
future.
A
Before
you
actually
need
to
accomplish,
you'd
see
that
roughly
in
this
2026
time
scale
here
where
we
are
going
to
really
do
an
upgrade
of
the
detector
facility,
so
even
more
particles
colliding
at
the
same
time,
you
see
that
basically
or
correct
methods.
They
do
not
scale
to
the
same
level
as
what
we
actually
need
in
order
to
take
advantage
of
this
full
data
set.
A
That
is
based
on
diffusion
models,
where
the
core
idea
is
that,
if
you
have
a
way
that
you
can
control
diffuse
the
sample
into
some
noise,
so,
for
instance,
you
can
imagine
that
we
have
this
dog
here
and
through
some
equation
that
evolves
over
time.
You
can
transform
this
dog
into
a
noise
distribution.
Then
you
can
also
do
the
opposite.
You
can
train
your
network
that
learns
how
to
do
this
inverse
the
fusion
process,
organizing
process,
where,
basically,
you
can
stock
analyze
distribution
and
keep
denoising
until
you
get
something
back.
A
So
the
idea
is
that
you
can
train
a
method
that
we
start
by
taking
some
random
noise
and
use
machine
learning
in
order
to
convert
this
noise
into
something
that
is
actually
useful.
In
our
case,
something
that
looks
like
the
right
picture
here
where
here
I'm
showing
an
election
interacting
with
Detective
material
and
each
pixel
in
this
picture,
corresponds
to
energies
that
have
been
deposited
by
this
box
by
this
particle
and
to
train
this
machine
learning
method.
A
I
use
the
perimeter
certain
computer,
which
has,
as
you
all
know,
lots
of
gpus
and
are
really
good
for
projects
that
take
advantage
of
machine
learning.
Libraries
like
tensorflow
in
this
case
I've
been
using
16
GPU,
so
not
something
incredibly
large
I'll
distributed
using
horovod
and
using
some
different
test.
Data
sets
just
to
see
if
this
concept
of
training
designated
model
is
powerful
enough.
In
order
to
give
you
a
selfie
that
will
actually
be
user
for
data
analysis.
A
So
in
this
case,
I
have
three
different
data
sets
that
corresponds
to
different
detector
layouts,
and
the
main
difference
between
these
data
sets
just
the
amounts
of
the
pixels
that
they
possess.
That
I
need
to
simulate.
So
in
this
case,
you
can
see
that
in
this
first
data
sets
I
have
about
200
or
400
Dimensions.
They
need
to
be
simulated
to
a
data
set
three
where
I
have
46
000
channels,
then
now
the
machine
learning
needs
to
learn
how
to
properly
simulate.
A
That
I
need
to
do
and
if
you
just
compare
the
time
that
it
takes
to
generate
100
particle
interactions
and
if
you
look
at
the
column
in
the
middle
called
column
score,
you
see
the
for
data
set
one
that
would
take
about
four
seconds
using
the
machine
learning
method
which
Compares
a
full
simulation
can
take
up
to
100
seconds
to
1000
seconds.
So
a
couple
order
of
magnitude
faster,
just
because
in
the
full
simulation
depends
on
the
energy
of
the
particle.
The
time
that
it
takes
to
do
this.
A
A
A
Do
our
data
analysis
methods
after
the
detector
effects
and
then
apply
a
correction
over
those
results
in
order
to
convert
these
results
after
detective
effects
to
how
they
should
look
like
before
they
interacts
with
the
detector
and
that's
the
idea
of
unfolding,
and
that's
also
seen
as
the
convolution
problem
and
currently
main
the
main
methods
that
do
die
in
high
energy
physics.
They
need
histograms.
A
And
if
you
go
into
the
future,
where
we
have
more
data
and
you'll.
Look
back
and
you
think
well,
can
I
reuse
the
same
histogram
or
can
I
change
the
histogram,
then
you
already
lost
the
information
since
after
you
put
things
in
a
histogram,
you
cannot
go
back
to
how
they
used
to
look
like
before
they
were
put
in
the
histogram.
A
That's
why
you
can
also
use
machine
learning
to
do
this
process
of
trying
to
identify
how
physics
events
look
before
other
detector
Effects
by
calculating
a
re-awading
function.
So
the
idea
is
that
if
you
have
the
response
of
some
Physics
observable
before
and
after
interacting
with
a
detector,
then
you
can
train
a
machine
learning
that
learns
how
to
relate
one
distribution
to
the
other.
A
So
you
can
always
move
between
these
two
samples,
one
before
and
one
after
detector
effects,
and
the
interesting
thing
is
that
this
idea
has
also
been
used
in
a
real
experiment
in
the
real
particle
collisions
so
just
show
they
actually
not
only
works
with
other
example,
but
really
works
with
your
barcode
collisions.
So,
in
this
case,
I
use
some
data
set,
that's
been
collected
by
an
experiment
or
running
during
2006
and
2007
so
a
few
years
ago.
A
But
the
series
is
a
really
nice
data
set
that
you
can
use
to
study
particular
interactions
and
to
train
this
model.
We
really
need
a
lot
more
computing
power,
so
in
this
case
I
use
a
128
gpus
just
because
data
set
sizes
that
we
need
to
deal
with
here,
a
lot
bigger
compared
to
what
we
had
before.
Plus
you
also
need
to
take
into
accounts,
other
things
like
uncertainties
and
so
on.
A
So
if
everything
you
see
in
these
slides,
for
instance,
a
physics
of
circles
that
were
measured
in
just
one
pass
of
this
machine
learning
matter
and
the
last
part
that
I
want
to
talk
in
this
talk
is
about
how
we
can
take
the
data
that
we
have
from
particle
collisions
and
try
to
identify
things
that
we
don't
understand
about
it.
Because
there's
a
lot
of
open
questions
in
high
energy
physics
that
we
basically
don't
have
a
good
answer
for.
A
A
And
one
thing
that
is
very
tricky
when
you're
trying
to
think
about
how
to
identify
things
that
you
don't
understand
is
that
let's
say
you
have
some
data
that
you
put
in
a
histogram
or
whatever
data,
visualization
method
that
you
like.
That
is
the
one
on
the
left
and
you
want
to
interpret
how
these
data,
what
you're
actually
looking
at.
A
So
you
can
imagine
that.
Maybe
there
is
something
that
is
very
exciting
and
there
is
some
new
physics
process
that
you
don't
understand
and
that's
why
you
want
to
figure
out,
but
there's
also
the
chance
that
everything
that
you're,
just
looking
at
is
compatible
with
the
predictions
from
the
theory
that
you
have,
that
just
predicts
how
particles
interacts
your
phone
another
and
those
you
understand
relatively
well.
A
So
the
idea
of
this
project
is
to
ask
yourself:
okay,
if
I
have
some
data
sets
that
tells
me
that
might
be
something
interesting
that
I
don't
understand,
can
I
give
you
context,
can
I
identify
how
false
positives
look
like
in
order
to
take
the
response
of
false
positives
and
compare
with
the
data
that
I
have
and
see
if
they
match
or
not,
which
is
exactly
this
idea
of
this
project?
What
the
idea
was
you
estimate,
how
do
false
positives,
look
like
by
using
the
data
but
in
different
regions?
A
The
Learners,
how
to
interpolate
between
these
two
ideas
and
the
nice
thing
is
that
if
you
try
to
use
this
method
in
a
data
set
that
basically
has
nothing
new.
There
is
no
new
physics
and
everything
just
compatible
with
your
background,
then.
Basically,
what
this
method
tells
you
is
that
there's
nothing
there,
which
is
really
good.
So
the
matter
is
really
capable
of
telling
you
that
there
is
nothing
when
there
is
nothing,
but
on
the
other
hand,
if
there
is
something,
then
this
method
is
also
able
to
tell
you
and
say
that.
A
Well,
the
number
of
observations
that
you
see
in
this
region
does
not
match
the
number
of
observations
that
we
should
expect
in
a
force
with
a
background.
Only
hypothesis
and
you
can
use
this
difference
between
or
you
have
to
Observer
what
you
predict
to
give
an
estimates
of
how
many
new
things
are
actually
in
your
sample,
which
is
what
is
shown
here
on
the
right
foreign.