Presentation at NERSC Data Day 2017 at Lawrence Berkeley National Laboratory on the potential of Generative Adversarial Networks to speed up scientific simulation and empower scientists and researchers.
2. The Rise of AI
2
• Deep
Learning
research
has
achieved
impressive
results
both
in
academic
and
industry
contexts
Finding
extreme
weather
events
in
climate
simulations
Classifying
new
physics
events
at
the
Large
Hadron
Collider
Computer
Vision
Sentiment
analysis
Language
understanding
• Deep
Learning
applications
have
the
potential
to
revolutionize
science
4. Generative Modeling for Science
4
Why study generative models?
“After
all,
when
applied
to
images,
such
models
seem
to
merely
provide
more
images,
and
the
world
has
no
shortage
of
images.”
5. Generative Modeling for Science
4
Why study generative models?
“After
all,
when
applied
to
images,
such
models
seem
to
merely
provide
more
images,
and
the
world
has
no
shortage
of
images.” except
that
we
do!
ATLAS
grid
consumption
LSST-‐DESC
cosmological
simulation
run
on
16384
of
Titan's
GPU-‐enhanced
nodes
8. Taxonomy of Generative Models
5
From
I.
Goodfellow
Maximum
Likelihood
Explicit
density
…
Tractable
density
Fully visible belief nets:
NADE
MADE
PixelRNN
Change of variables
models (nonlinear ICA)
9. Taxonomy of Generative Models
5
From
I.
Goodfellow
Maximum
Likelihood
Explicit
density
…
Tractable
density
Fully visible belief nets:
NADE
MADE
PixelRNN
Change of variables
models (nonlinear ICA)
Approximate
density
Variational
VAE
Markov
Chain
Boltzmann machine
10. Taxonomy of Generative Models
5
From
I.
Goodfellow
Maximum
Likelihood
Explicit
density Implicit
density
…
Tractable
density
Fully visible belief nets:
NADE
MADE
PixelRNN
Change of variables
models (nonlinear ICA)
Approximate
density
Variational
VAE
Markov
Chain
Boltzmann machine
11. Taxonomy of Generative Models
5
From
I.
Goodfellow
Maximum
Likelihood
Explicit
density Implicit
density
…
Tractable
density
Fully visible belief nets:
NADE
MADE
PixelRNN
Change of variables
models (nonlinear ICA)
Approximate
density
Variational
VAE
Markov
Chain
Boltzmann machine
Markov
Chain
Direct
GSN
GAN
12. Taxonomy of Generative Models
5
From
I.
Goodfellow
Maximum
Likelihood
Explicit
density Implicit
density
…
Tractable
density
Fully visible belief nets:
NADE
MADE
PixelRNN
Change of variables
models (nonlinear ICA)
Approximate
density
Variational
VAE
Markov
Chain
Boltzmann machine
Markov
Chain
Direct
GSN
GAN
15. CaloGAN
7
• Fast
&
accurate
simulation
of
calorimeter
showers
in
a
three-‐layer
heterogeneously
segmented
LAr+lead
detector
• Ad-‐hoc
design
to
fit
Physics
data:
• sparsity
• high
dynamic
range
• highly
location-‐dependent
features
• Soon
to
be
integrated
in
Geant
(Paganini
et
al.,
arXiv:1705.02355)
3x96
12x12
12x6
16. CaloGAN
8
Energy
(Paganini
et
al.,
arXiv:1705.02355)
Average
energy
deposition
per
calorimeter
layer
in
the
GEANT4
training
dataset
(top)
and
in
the
GAN
generated
dataset
(bottom)
Ten
positron
showers
generated
by
varying
shower
energy
in
equal
intervals
while
holding
all
other
latent
codes
fixed.
The
three
rows
are
the
shower
representations
in
the
three
calorimeter
layers.
The
energies
of
showers
in
the
green
box
were
within
the
range
of
the
training
dataset,
while
the
ones
in
the
red
box
are
in
the
extrapolation
regime.
• Realistic
average
and
individual
images
• Diverse
samples
• Conditional
generation
based
on
physical
attributes
• Parameter
interpolation
and
extrapolation
17. CosmoGAN
9
• Generate
convergence
maps
of
a
particular
ΛCDM
model
• Match
summary
statistics
of
training
dataset:
• 1st
order
statistic
(pixel
intensity)
• 2nd
order
correlations
• measurement
of
the
non-‐
Gaussian
corrections
(Mustafa
et
al.,
arXiv:1706.02390)
18. Conclusions & Outlook
10
• GANs:
promising
technology
to
accelerate
scientific
simulation
• Potential
to
save
millions
of
CPU
hours
and
dollars
• In
scientific
applications:
• Unique
considerations
and
constraints
• Unique
evaluation
methods
for
generative
models
• Outstanding
challenges
require
continued
R&D
to
meet
fidelity
demands