►
From YouTube: Education and Workforce WG:Domain Specific Graduate Training & Human Tumor Atlas Network Data Center
Description
Talk #1:
Dr. David Beck
Talk Title: Domain Specific Graduate Training in Data Science in the Context of a University Wide Approach
Bio: David Beck is a Research Associate Professor in Chemical Engineering at the University of Washington (UW), a Senior Data Science Fellow and Director of Research and Education for the eScience Institute. Founded in 2008, UW’s eScience Institute serves to advance data-intensive discovery in all fields from the humanities to science, engineering and medicine. Dr. Beck’s research explores the intersection of molecules and data science, or molecular data science, and its applications to energy, environment and health. Beck is the associate director of the NSF Research Traineeship Data Intensive Research Enabling Clean Tech (DIRECT) and co-director of the Engineering Data Science Institute (EDSI) at UW. He is an adjunct faculty in the Paul G. Allen School for Computer Science and the Department of Environmental and Occupational Health Sciences and coordinates the UW Data Science Seminar.
Talk #2 at 21:21:
Dr. David Gutman
Talk Title: The Human Tumor Atlas Network Data Coordinating Center
Bio: Dr. David Gutman received his PhD in Neuroscience and MD from Emory University School of Medicine and is now an associate professor of Neurology. Dr. Gutman is also a staff Physician at the Atlanta VA Medical Center on the Psychiatric Inpatient Unit. Dr. Gutman's research focuses on developing informatics technologies to process and analyze large imaging data sets, and is leading the imaging informatics efforts to integrate data for the Human Tumor Atlas Network.
A
To
introduce
our
speakers
today
and
the
goals
of
this
group
are
in
the
ether
pad,
but
they're
also
good
to
say
you
know
we
want
to
hear
from
the
different
data
science
programs
that
are
being
happening
also
what
people
are
teaching,
how
they're
looking
at
data
science
funded
research
and
what
that
is,
and
what
the
implications
of
that
for
education
are
in
this
group.
So
today,
we'll
have
two
talks
to
david's,
so
we'll
have
david
and
david
and
david
beck
come
before.
B
Super
thank
you
well
as
much
as
my
favorite
topic
is
to
talk
about
myself.
I
think
I'll
skip
and
go
right
into
some
of
the
material.
Let
me
start
my
timer
here,
okay,
so
I
am
in
chemical
engineering
as
renata
mentioned,
and
I'm
also
part
of
the
e
science
institute.
This
is
our
data
science
institute
on
the
university
of
washington
campus
and
we
were
formed
about
2008.
B
So
if
we
all
put
our
way
back,
hats
on
what
everybody
was
talking
about
in
2008
was
big
data,
so
the
words
data
science
hadn't
really
made
it
into
the
common
lexicon.
At
that
point,
and
all
across
the
university
of
washington
campus,
there
were
groups
that
were
confronted
with
large
or
complex
or
difficult
to
manage.
Data
sets
and
the
challenge
wasn't
generating
the
data
that
had
actually
been
done.
B
What
the
challenge
was
was
how
to
extract
knowledge
and
actual
information
from
those
data,
so
we
convened
a
group
of
folks
from
across
campus
and
really
started
to
think
about
how
do
we
move
beyond
just
data
collection
and
into
use
use
maximum
use
of
those
data?
So
we
sort
of
broke
up
our
tasks
into
three
things:
education,
research,
and
how
to
build
a
community
of
practice,
and
so
I
want
to
talk
about
the
education
component
first,
fun
fact.
B
However,
this
is
the
origin
story
for
the
logo,
for
the
e
science
institute
is
where
folks
were
coming
from,
so
all
right,
so
what
we
decided
or
what?
What
what
arose
out
of
the
these
discussions
was
that
there
are
a
set
of
core
skills
that
would
constitute
what
would
become
data
science
at
the
university
of
washington
and
the
first
one
being
data
management
right.
If
data
is
your
currency,
then
managing
those
data
is
obviously
an
important
component.
B
I'm
actually
going
to
downplay
that
here,
because
it,
I
think
it's
over
represented
in
the
data
science
skill
set,
of
course
visualization.
So
if
you
have
large
and
complex
data
or
complex
results
coming
out
of
black
boxes,
like
some
machine
learning
algorithms,
how
do
you
communicate
those
results
in
a
visual
way
to
your
stakeholders,
to
policy
makers,
to
your
peers
and
so
forth?
B
Well,
when
I
say
specific,
I
really
mean
every
domain
area
on
campus,
whether
it's
arts,
humanities,
engineering
or
sciences,
so
that
we
really
got
started
with
the
formal
program
in
2013
with
an
igert,
and
I
was
the
co-pi
representing
kemi
on
that,
and
so
we
had
sort
of
the
abcs
of
science
and
engineering
here,
along
with
cs
and
statistics,
and
the
approach
was
to
have
chemi
phds.
Take,
for
example,
take
cse,
grad
level,
machine
learning
classes
and
we
produced
our
first
chemical
engineering
phd
the
university
of
washington.
B
B
Unfortunately,
though,
having
chemis,
chemi
phds
or
astronomy
or
biology
phds
take
grad
level,
cse
machine
learning
classes
or
your
grad
level
stats
statistics
classes
is,
is
a
high
prerequisite
barrier,
and
so
our
participation
rate
is
really
quite
low
even
today.
So
this
we
maintain
this
program
and
I'm
happy
to
talk
about
sustainability
as
well,
but
our
participation
rate
is
about
ten
percent
of
our
phd
students
and
zero
percent
of
our
masters
students,
and
so
that
was
another
problem
with
the
that
original
model.
So
in
2016
we
received
an
nrt.
B
The
nrts
are
the
spiritual
successors
for
the
iger.
The
concept
is
to
do
more
with
less
so
whether
the
aggregates
were
about
funding
trainees
directly.
The
nrts
included
some
trainee
funding
directly,
but
it
was
larger
more
about
building
these
sustainable
programs
and
all
of
these
participating
units.
You
can
see
here
that
they
have
a
common
theme
and
that
they're
molecularly
focused.
B
So
the
goal
of
this
program
was
to
have
our
students
be
fluent
in
data
science
methods,
best
practices
and
tool
developments,
so
they
should
know
how
to
choose
a
neural
network
architecture
from
a
suite
of
available
ones.
How
to
use
database
query
languages,
how
to
write
software
using
test
driven
development
and
do
things
like
perform
code
reviews
right
so
we're
not
turning
them
into
computer
scientists
or
statisticians
we're
turning
them
into
data
science,
fluent
chemical
engineers
or
molecular
engineers
and
so
forth.
Again,
no
prerequisites!
B
This
is
super
important
to
our
program
and
we
wanted
it
accessible
to
our
masters
and
phd
students.
So
we
have
it
down
into
a
format
where
they
can
complete
it
in
six
months,
which
is
basically
two
quarters
we're
in
a
quarter
system
at
uw
and
that's
three
courses
and
it's
a
project-based
approach
to
teaching
these
five
things
that
I
talked
about
because
all
the
departments
have
this
molecular
focus.
B
We
can
focus
on
teaching
the
students
things
like
machine
learning
based
on
molecular
properties
and
not
predicting
frenemy
networks,
so
it's
highly
contextualized
for
them,
and
that
provides
a
lot
more
utility.
We
think
at
least
the
evaluations
seem
to
hint
at
that
we
leverage
the
cohort
effect
really
heavily.
B
When
we
were
in
person,
we
were
making
use
of
the
active
learning
classrooms
and
there's
some
semblance
of
this
with
zooms
and
breakouts,
but
you
can
see
a
pod
of
students
can
work
together
and
share
their
screens
here
they
can
share
back
to
the
whole
room
and
I
can
project
each
one
of
these,
and
so
they've
been
really
useful
for
this
kind
of
teaching.
Also
jupiter
notebooks
have
been
tremendously
helpful.
This
is
a
great
way
for
folks
that
have
very
little
computing
experience
to
skill
up
quickly
and
do
it
in
a
very
safe
environment.
B
That's
not
intimidating.
We
have
these
transcriptable
options.
I
mentioned
them
briefly,
but
basically,
if
you
complete
this
program,
you
get
a
modifier
your
degree,
so
it'll
be
a
master's
in
chemical
engineering,
with
a
data
science
option
or
a
master's
in
materials
science
with
the
data
science
option
or
phd
as
well,
and
we
chose
python
for
a
number
of
reasons.
I'm
happy
to
discuss
separately.
B
When
you
have
a
no
prereq
program
like
this,
you
definitely
get
a
bi-little
distribution.
We've
observed
this
in
a
couple
other
settings,
so
you
get
the
folks
who
were
chemistry.
Undergrads
and
minored
in
computer
science
and
then
the
folks
who
were
chemie
undergrads
and
never
took
any
computer
science
classes,
and
so
those
are
the
kind
of
folks
that
are
coming
in
here
and
having
the
ability
to
use
the
digital
tools
for
doing
peer
learning
and
peer
support
has
been
really
great,
and
then
they
do.
B
B
B
That
gives
us
a
lot
of
flexibility
to
kind
of
front
load,
a
lot
of
the
programming
and
software
engineering,
so
that
by
the
time
they
start
doing
the
data
science
pieces.
They
have
a
good
set
of
skills
and
a
good
set
of
software
hygiene
in
place,
so
they're
not
producing
messy
notebooks
they're
thinking
through
the
software
engineering
and
design
principles,
while
they're
doing
the
machine
learning.
B
So
afterwards
we
have
a
big
poster
session.
Again.
This
is
from
2019.
Usually
we
we've
done
this
in
in
in
person,
but
these
last
couple
years
it's
been
all
virtual
and
digital
and
it's
amazing
the
things
that
the
students
will
do
when
they
have.
You
know
the
zoom
breakout
rooms
are
dynamic
right
and
they
don't
require
a
static
poster
and
it's
amazing
what
the
students
will
do
with
that
move
a
little
bit
past
these.
B
So
then
the
capstone
again
is
the
the
holistic
integration
of
what
they've
done
in
the
winter
quarter
and,
alongside
of
it,
we're
doing
soft
skill
preparation,
so
communication
project
management,
stakeholder
engagement,
really
helping
the
students
build
out
their
research
and
and
professional
networks.
The
capstones
come
from
uw
national
labs
ngos
and
companies.
B
And
again
our
visual
syllabus
looks
like
this,
so
we
intersperse
professional
development
topics
along
with
stand-ups,
which
are
a
rapid
way
to
communicate
about
their
progress
in
the
software
that
they're
building
I'll
again,
I'm
gonna
skip
over
these.
Suffice
it
to
say
that
almost
all
of
these
have
been
published
these
posters
that
I've
shown
you
have
turned
into
software
products
that
have
been
published.
This
one
was
shown
at
a
couple
of
national
meetings
and
was
also
published
so
really
quickly.
In
closing,
where
do
our
folks
go?
B
So
these
are
my
students
that
have
gone
through
the
program.
We
see
that
there's
this
typical
group
of
our
masters,
students
that
go
to
industry
or
to
go
to
phd
somewhere
else.
So
that's
that's
sort
of
static
and
happens
whether
or
not
they
participate
in
this.
There
are
some
that
I
didn't
have
data
for
when
I
made
this
figure,
but
this
is
the
interesting
thing
right.
B
This
is
what
was
enabled
there's
this
whole
new
sort
of
group
of
folks
that
are
going
to
become
data
scientists
molecular
data
scientists
in
a
chemistry
or
chemistry,
related
field.
This
is
new
as
well.
We
lose
some,
they
decide.
They
want
to
go
work
for
amazon
or
google
instead
of
saying
in
science
and
that's
great,
because
we
want
optimal
outcomes
for
the
students,
not
just
to
try
to
keep
them
in
science.
B
A
C
Sure
it's
actually
a
really
simple
question.
Can
I
get
a
copy
of
that
picture
of
your
collaborative
classroom,
we're
working
on
designing
a
space
at
trinity
and
I
would
like
some
more
ideas
to
help
share
with
our
architects?
Oh
also,
since
you
shared
all
the
github
stuff
like
I'll,
go
look
at
the
github
core
stuff,
I'm
excited
about
that
that
works
in
the
second
question,
but.
B
C
B
B
A
A
Transfer
the
links
that
are
in
the
chat
into
the
ether
pad
too
and
we're
you
know
we
have
plans
to
make
those
available,
but
we
can
just
send
it
out
as
well.
So
if
you,
if
you
want
to
do
that,
also
the
github
link
that
was
in
the
last
slide,
would
you
paste
that
into
the
ether
pad
too
just
so
that
people
could
directly.
A
B
D
B
I
I
think
I
mean
there's,
there's,
there's
apples
and
oranges
comparison
there,
but
it's
much
easier
to
teach
the
students
in
through
the
nrt
the
program
that
I
just
described,
because
one
we
have
the
cohort
effects
two.
We
have
them
for
eight
hours
a
week
effectively
for
an
entire
quarter.
So
it's
really
intensive
and-
and
you
know,
it's
also
really
focused
right.
So
the
education
that
they're
getting
is
exquisitely
targeted
and
many
of
the
projects
that
they
do
is
their
own
research.
B
So
I
actually
think
that
it's
much
easier
to
do
it
this
way
and
that's
again,
the
philosophy
at
uw
to
contextualize
the
data
science
education
in
the
domain.
It's
much
easier
to
do
that
than
it
is
to
take
the
folks
out
of
the
domain
and
put
them
in
cse,
at
least
that
that's
been
sort
of
my
observation
and
my
experience.
D
D
B
Yeah,
I
think
that's
a
great
great
sort
of
way
to
describe
it.
I
mean
you
do
need,
of
course,
particularly
with
statistics
and
things
like
bias,
variance
trade-off,
and
you
know
some
of
those
kinds
of
concepts.
You
have
to
do
some
motivation
through
theory
and
in
fact
that
helps
the
students
who's,
whose
learning
style
is
more
sort
of
algorithmic,
and
so
it's
important
to
do
that.
B
But
I
think
you're
right
that
they're,
the
hands-on
component
is
something
that
really
benefits
the
students
and,
seeing
you
know
going
as
quickly
as
possible
to
a
decision
tree
for
computing
boiling
points
or
something
like
that
or
predicting
boiling
points
is,
is
something
that
really
resonates
and
then
traps
them
and
gets
them
excited
about.
Oh
now,
let's
do
this
with
a
neural
network,
or
let's
do
this
with
some
other
tool,
and
so
I
think
you're
right.
It's
that
getting
to
the
hands-on
as
quickly
as
possible
is
something
that's
really
really
useful.
E
Sure
I
was
just
I'm
sorry,
I
missed
the
very
beginning
of
your
presentation,
david,
but
you
know
you're
really
focused
on
any
sort
of
major
that
has
a
molecular
molecular
component
and-
and
I
think
that
I
think
it's
great-
I
just
my
question-
is
how
do
you
decide
which
area
of
application
to
focus
on
and-
and
you
know,
thinking
of
scaling
up
to
other
areas
of
application?
Do
we
wind
up
basically
duplicating
our
efforts?
E
You
know,
we've
got
one
one
set,
of
course
working
classes
for
those
in
molecular
sciences,
another
for
social
sciences,
another
for
arts
and
theater
right.
I
mean
it's
just
it's
very
difficult
to,
like
you
said,
take
someone
outside
of
their
domain
and
get
them
to
come,
take
classes
and
say
cs
or
statistics,
or
you
know
another
discipline
where
it's
much
easier
to
meet
them
where
they're
at.
But
then
we
wind
up
meeting
everyone
where
they're
at
and
that's
that's
a
lot
of
effort.
E
B
But
you
definitely
no
that's
a
great
plan,
and
so
you
can
definitely
run
into
a
scalability
issue
there.
Ben
marwick
earlier
in
the
series
and
hunter
shafer
talked
about
some
how
we're
doing
that
at
the
undergraduate
level,
how
to
support
the
diversity
of
programs.
That
way,
I
I
in
in
our
chat
here,
I
posted
a
link
for
future
rivers,
which
is
another
program
that
we
have
at
the
university
of
washington.
B
That
is
similar
in
spirit,
in
that
it
has
the
data
science,
education,
but
the
common
theme
there
is
folks
working
at
the
urban
water
nexus
right.
So
I
think
I
think
that
the
first
thing
I
would
say
is
that
when
you
can
build
up
a
critical
mass
within
an
area,
then
we
figured
out
how
to
do
that
in
a
way
that
we
can
replicate
it
again
in
into
a
new
area
where
there's
a
critical
mass.
B
But
it
is
hard
to
bootstrap
that
entirely
from
nothing,
and
so
there
are
some
shared
components
of
these
programs
that
I
didn't
really
talk
about,
facilitated
by
the
e-science
institute.
So
there's
some
seminars
and
some
other
less
formal
activities
that
you
you
can
port
from
each
one
of
these
that
reduce
the
sort
of
cost
of
scaling
out
to
new
disciplines.
B
But
you
know
you
are
right
that
if
you
want
to
support
by
the
way
I
didn't
mention
it,
we
have
50
to
60
folks
going
through
this
program
every
year.
If
you
want
to
support
50
or
60
folks,
you
know
that
takes
four
faculty
approximately
three
to
four
faculty
or
two
faculty
and
some
tas
to
do
and
that's
a
heavy
lift
to
do
from
nothing.
A
That's
very
valuable
yeah.
We
should
put
that
as
a
I
want
to
move
to
our
next
speaker,
but
if
there
are
other
questions,
if
there's
other
questions
I
see
in
the
chat,
can
you
put
them
in
the
ether
pad
and
david?
Would
you
mind
just
looking
in
there
and
answering
keep
answering
quotes
if
you
see
them
in
there
perfect
so
david
governor?
Can
you
want
to
share
your
screen.
F
Weird
every
time
I
share
my
screen,
it
mutes
me
feature
so
my
name's
dave
gottman,
I'm
actually
going
to
talk
to
you
about
free
data.
You
guys
can
probably
use
for
your
courses,
so
I
won't
go
too
much
about
my
background,
but
I'm
actually
a
psychiatrist,
but
I
spend
most
of
my
time
doing
cancer
data
science
and
data
harmonization
and
that's
a
question
that
requires
alcohol
to
explain
so
oops
wrong
button.
F
So
I
originally,
I
don't
know
10
years
ago,
got
involved
in
something
called
the
cancer
genome
atlas
which
basically
the
nci
released.
I
don't
know
half
a
petabyte
to
a
petabyte
of
free
data,
essentially
for
people
to
play
with.
I
got
excited
about
it
and
started
playing
with
digital
pathology
or
essentially
they
can
scan
biopsies.
F
So
I
am
now
involved
in
something
called
the
human
tumor
atlas
network,
which
is
another
nci
national
cancer
institute
project
that
was
funded
off
of
the
bo
biden.
Moonshot
cancer
moonshot
initiative
from
a
couple
of
years
and
we're
now
starting
to
slowly
see
some
of
the
fruits
of
that
investment,
and
so,
as
a
data
scientist,
my
job
in
this
network
is
the
nci
funded
10,
different
research,
centers,
five
pre-cancer
atlases
and
five
human
tumor
atlases
and
a
data
coordinating
center.
F
So,
basically,
all
of
these
different
universities
and
groups
generate
many
millions
of
dollars
of
data,
and
one
of
the
goals
of
the
moon
shot
was
to
make
that
data
release
to
the
public
as
quickly
as
possible.
Harmonize
it
and
clean
it
up,
and
that's
really
where
a
lot
of
my
work,
fortunately
or
unfortunately
focuses
I
like
it.
So
I
guess
it's
fortunate.
F
So
there's
I'm
actually
just
a
co-investigator
on
this
project
and
we
a
couple
of
the
we
were
one
dana
farber
sage,
memorial
zone,
kettering
and
isb
in
seattle-
are
the
kind
of
the
major
hubs
for
that.
Although
I've
taken
on
a
pretty
active
role
so
yeah,
you
know
not
shockingly.
The
goal
of
the
human
tumor
atlas
network
in
a
cancer
moon
shot
is
to
basically
bring
increased
focus
onto
figuring
out
what
drives
cancer.
You
know.
F
So
it
really
is
sort
of
a
big
data
process
and
for
these
human
tumor
atlas
networks,
what's
good
and
also
slightly
irritating,
which
I'll
get
into
in
a
minute
is
unlike
some
of
these
earlier
projects,
where
we
kind
of
described
what
experiments
people
should
do
or
what
kind
of
data
they
should
collect.
This
really
becomes
a
lot
bigger
of
a
data
science
question,
because
a
lot
of
the
technologies
that
we're
using
for
this
big
date
for
the
human
tumor
atlas
network,
basically
don't
exist
yet
or
they're
two
years
old.
F
F
So
it's
been
a
labor
of
love
and
what's
also
fun
about
the
project
is
since
it's
a
bit
of
a
wild
west,
harvard
likes
to
do
these
type
of
assays
and
so
harvard
generates
a
lot
of
these
type
of
assays
stanford
likes
to
do
this
type
of
experiment.
So
then
they
generate
these
type
of
assays
and
sometimes
people
use
the
same
platform
differently,
and
sometimes
they
use
the
same
different
platforms
in
the
same
way.
F
So
we
essentially
have
the
entire
gamut
of
data
available
and
I'll
get
to
my
favorite
analogy
a
little
bit
later
on,
but
the
data
standards.
We
want
this
data
to
be
publicly
accessible,
searchable,
indexable
and
also
there
are
so
many
hipaa
requirements
and
data
cleanup
requirements
that
it
makes
even
some
of
the
call
it
easy,
github
stuff,
a
giant
pain
in
the
butt.
So
like.
I
can't
send
my
friend
a
thumbnail
of
a
biopsy
for
him
to
test
his
clinical
data
portal
because
it
hasn't
been
released
publicly.
F
F
Sorry
heard
me
so,
like
I
said
some
of
the
data
challenges.
There's
an
open
list
of
assays
we're
not
prescriptive.
People
can
basically
whatever
they
decided
they
like
they
get
to
do.
We
have
longitudinal
data,
which
makes
it
even
more
fun
because
the
data
has
to
be
de-identified,
but
we
also
have
to
be
able
to
track
these
samples
longitudinally
and
the
centers
that
give
us
the
data
have
to
do
a
lot
of
the
time,
shifting
and
stuff.
So
that
adds
a
fun
level
of
complexity.
F
F
F
F
It
can't
it's
not
possible
for
the
data
coordinating
center
to
even
understand
what
some
of
these
new
assays
that
people
are
generating
for.
So
as
we
describe
as
we
generate
these
like
data
models
and
rfcs,
to
get
comments
on
like
what
we
should
even
capture
about
your
experiment.
There's
a
lot
of
back
and
forth
between
these
centers-
and
maybe
this
is
my
psychiatry
hat
speaking.
But
getting
people
to
do
anything
is
like
herding
cats.
So,
even
if
that
you
know
people
are
excited
like
we
uploaded
all
of
our
data,
okay,
cool.
F
Well,
what
is
it
well,
we're
writing
a
manuscript
on
it,
we'll
let
you
know
so
anyway.
That's
the
whole
other
problem,
so
just
just
as
an
example
we're
getting
single
cell
rna
data,
we're
getting
sight,
seek
metabolomics,
proteomics
microbiome
again.
A
lot
of
this
is
just
omic
jargon,
but
some
of
these
technologies
are
like
one
or
two
years
old,
and
I
don't
even
know
what
they
do.
We
then
have
electron
microscopy
data,
which
I
understand
very
well,
because
it's
just
a
lot
of
tiffs,
but
it's
tens
of
thousands
of
tiffs.
F
So
then
you're
talking
about
many
many
many
terabytes
of
data
that
no
one
may
ever
look
for,
but
we
may
have
to
pay
for
in
terms
of
storage.
There's
all
these
fun
fluorescent
stuff.
My
actual
specialty
is
really
imaging
related
stuff.
So
this
is
the
stuff
I'm
interested
about,
but
it's
I
know
a
lot
about
like
the
rest
of
the
stack
as
well
and,
more
importantly,
the
metadata
attributes
are
another
small
disaster,
because
some
of
the
metadata
require
are
not
actually
collected
by
the
instrument
that
generates
it.
F
So
one
of
the
big
things
about
the
way
we're
approaching
this
is
we
have
a
single
bucket
or
actually
we
have
23
buckets,
but
essentially
all
of
the
centers
dump
their
file
contents
and
metadata,
hopefully
into
one
of
our
buckets.
So
we
have
one
kind
of
ingestion
pipeline
that
does
the
cleaning
the
harmonization
and
the
standardization
we
hope.
F
But
then
you
know
with
the
magic
of
rest
apis
and
your
other
favorite
apis.
We
want
to
basically
make
many
different
ways
for
this
data
to
be
accessed,
because
it's
so
rich
and
also
because,
as
many
of
you
know,
a
genomics
database
is
not
great
at
visualizing
images
and
an
image
database
is
not
great
for
for
genomics,
or
so
you
know,
epidemiology
whatever.
F
So,
oh
and
I'm
going
the
wrong
way.
Sorry
so,
like
I
said
one
of
my
areas
of
particular
interest,
obviously
I
don't
use
the
machine
learning
jargon,
but
I'm
particularly
interested
in
from
all
of
these
rich
medical
imaging
data
sets.
We
need
to
get
ground
truth
because,
even
though
I
went
to
medical
school,
when
I
look
at
these
pictures,
I
have
no
idea
what
I'm
looking
at
other
than
it's
probably
tissue
from
something.
F
F
We
need
to
make
sure
that
we're
kind
of
capturing
all
this
information
in
a
on
a
good
way,
and
you
know,
like
I
said,
I'm
not
going
to
go
much
of
deep
learning,
but
a
lot
of
the
excitement
of
these
big
data
sets
right
is.
If
you
have
a
big
enough
data
set,
it's
harmonizing
it's
clean
enough.
You
can,
you
know
feed
these
into
your
pipeline.
You
know
pick
your
pick,
your
widget,
but
you
can
hopefully
come
up
with
new
things,
new
new
observations,
new
understanding,
new
knowledge.
F
So,
like
I
said,
one
of
the
joys
of
big
data
you
know
is
my
data
bigger
than
your
big
data,
one
of
the
the
way
I
view
this.
So
this
is
our
current
dashboard
and
I
love
this
so
one
of
our
favorite
centers
and
apparently
I'm
gonna-
make
fun
of
stanford
right
now.
They've
uploaded
17
terabytes
of
data
into
our
bucket
they've,
provided
no
metadata
so,
and
I
can
even
tell
you
how
much
this
costs
to
store,
because
it's
fun,
but
essentially
I
now
have.
I
always
view
it.
F
As
imagine
someone
just
gave,
went
into
their
garage
and
say:
here's
all
my
photos,
they're
in
some
boxes
somewhere.
Can
you
please
find
my
trip
to
hawaii?
You
know,
essentially
without
any
of
the
metadata
being
standardized,
it's
just
a
lot
of
garbage,
and
so,
unfortunately,
and
probably
most
of
the
people
on
the
call
can
appreciate
this.
There's.
F
A
deep
lack
of
understanding,
except
for
maybe
with
the
nsa,
is
how
important
metadata
is
right
just
having
a
lot
of
data,
that's
poorly
annotated,
poorly
organized,
particularly
for
a
project
where
we
want
to
make
this
data
available
for
the
entire
universe,
because
that's
our
mandate
is
very
challenging
and
I
guess
keeps
data
scientists
employed
one
of
the
other
things
that
is
just
sort
of
tangential,
but
is
sort
of
my
favorite
thing
I
like
playing
with
big
images.
One
image
now
is
80
gigabytes.
F
We
actually
not.
So
if
you
want
to
download
the
file
and
play
with
it
locally,
take
about
11
minutes
and
I
have
very
fast
internet
even
more
entertaining
you-
don't
have
100
gigs
of
extra
storage
sitting
around.
It
also
costs
the
nci
about
eight
bucks
for
you
to
download
that
sample
from
it
from
amazon.
F
I
plug
it
into
my
lab
computer
and
I
say
download
everything
I
can
because
I
want
to
play.
If
I'm,
if
I'm
our
government,
I
guess,
and
now
I
have
to
pay
for
amazon's
egress
costs
which
again
are
not
cheap
at
scale.
You
get
into
all
of
these
very
interesting
things
in
terms
of
where
you
know
doing
computation.
In
the
cloud
figuring
out
how
to
give
people
access
to
your
bucket
in
the
cloud
setting
up,
jupiter,
notebooks,
setting
up
environments,
etc,
etc,
becomes
a
whole
other
kind
of
a
challenge
oops.
F
I
keep
clicking
on
the
wrong
thing
so
yeah,
like
I
said
I
did
this
envelope
calculation
a
while
ago,
and
it's
oh
that
worked
a
minute
ago.
Oh
well,
I
guess
I
want
you
that
video
so,
but
the
thing
that
excites
me
about
this
is
once
you
have
the
data.
You
can
do
really
cool
things
with
it
right.
So
these
are
the
stuff
I
like
looking
at.
You
have
a
history,
a
pathology
image
or
something
like
that.
F
F
There's
I
don't
know
fifty
thousand
to
a
hundred
thousand
cells
on
it
and
each
of
those
cells
expresses
some
amount
of
one
to
a
hundred
markers.
So
now
you
have
these.
Not
only
do
you
have
the
vis,
you
have
the
visualization,
which
is
what
I
like
doing,
but
you
have
all
these
exciting
clustering
and
data
science
and
annotation
all
these
other
fun
things
that
we're
hoping
we
can
enable.
Once
we
put
this
data
in
the
cloud
and
make
the
tools
sane
enough
to
do
it,
you
know
we
have
the
other
things.
F
F
F
A
I
want
to
give
time
for
people
to
answer
questions
and
if
you
have
links,
if
you
can
put
those
in
the
in
the
ether
pad
for
when
your
data
release
or
if
there's
a
page
that
people
should
follow
for
where
to
get
it
going
forward,
that
would
be
great,
but
also
just
saying
you
know
in
the
future
we
do
want
to
highlight
sometimes
in
these
talks
data
sets
that
are
open
for
capstones
or
for
our
coursework
or
for
other
things.
You
know
to
really
highlight
partners
that
are
doing
this
good
work.
A
Okay,
yeah,
I
see
a
request
to
repost
the
etherpad
link
so
for
those
that
join
later
you
can
add
yourself,
you
know
sign
in
and
then
also
you
can
see
the
notes
we
had
so
far.
So
does
anyone
have
a
question?
I
have
a
question,
but
then
I
will
I'll
do
first
so
that
other
people
have
questions
for
david.
E
Oh,
I
just
I
put
it
in
the
in
the
chat
and
I
may
thank
you
for
that
presentation
dave.
It
seems
like
really
really
exciting
work
that
you
guys
are
doing
and-
and
I
actually
have
two
questions
now
so
the
first
one
is-
I
may
have
missed
it.
These
are
all
samples
coming
from
real
cancer
patients.
Is
that
right
and
at
what
cancer
centers?
Are
they
at
specific
cancer
centers
in
the
u.s
and
and
do
you
also
get
scans,
for
example,
from.
F
F
The
the
way
that
worked
is
they
reached
out
to
cancer
centers,
mostly
in
the
u.s
but
all
over
the
world,
and
they
basically
said
I
have
a
couple
of
brain
tumors.
Give
me
a
thousand
bucks
I'll,
send
them
to
you
or
whatever
the
the
money
was,
but
basically
the
centers
where
the
data
was
sent
centrally
and
then
wash
you
did
the
mrna
sequencing
nationwide
children's
did
the
bio
pathology
scanning
the
dna
seque
or
the
the
chip.
Whatever
the
mrna
chips
went
to
seattle,
I
don't
really
remember,
but
basically
the
way
that
worked
is.
F
It
was
a
hub
and
spoke
model
where
the
samples
went
to
a
buyer
repository,
they
did
qc.
We
obviously
track
what
center
it
came
from.
But
then
the
centers
ran
50
000
mrna
chips
right
they
one
center.
Did
it
all?
What
was
made
sense
but
was
also
confusing.
Is
the
technology
has
obviously
changed
over
time
too?
F
So
originally
we
were
using
alpha
metrics
and
then
the
v2
came
out
which
was
cheaper,
and
then
you
have
to
figure
out
how
to
convert
the
old
data
to
the
new
data
and
for
some
things
they
just
kept
using
the
old
chips
for
some
of
them.
They
have
a
little
bit
of
both
and
you
know,
and
then
for
breast
cancer.
They
have
a
thousand
samples
because
that
was
deemed
more.
F
F
F
E
E
F
F
So
we
have
they're,
probably
like
two
or
we're
under
staff,
to
be
honest,
but
we
have
like
two
or
three
ftes
there
doing
the
aws
moving
data
security,
part
of
it
and
that's
been
the
total
disaster
off
the
record,
because
I
get
keep
getting
lot.
Did
anyone
know
that
if
someone
uploads
files
into
your
bucket,
you
can't
look
at
the
files
they
put
in
your
bucket,
unless
you
then
enable
another
permission,
I
didn't
so.
I
couldn't
view
half
the
files
that
were
uploaded
into
the
bucket
anyway.
F
So
there's
like
two
or
three
ftes
there,
mskcc
and
remorsel
kettering.
They
have
probably
two
ftes
they're
doing
more
front
end
visualization
stuff,
dana
farber.
We
have
a
project
manager
there.
We
have
ethan
cerami
who's
the
pi,
but
he
writes
a
lot
of
widgets
and
code
and
then
we
have
a.
I
don't
know
if
I
want
to
call
her
an
ontologist,
but
we
have
a
data
standards,
information
and
then
there's
probably
two
or
three
other
developers
floating
around
in
the
background
as
well
and
then
as
part
of
the
contractor
grant
or
whatever
mechanism.
F
This
is
the
centers
have
to
give
us
the
data.
I
mean.
That's
the
whole
point
of
these
moonshot
grants,
but
you
know
it's
like
I
said
it's
a
moving
target
because
oh
and
then
there's
isp
sorry
the
institute
for
system
biology
who
are
also
big
data
people,
so
they're
kind
of
working
on
big
query
and,
like
you
know,
less
structured
ways
to
again
make
the
data
available
for
people
to
you
know
basically
fuss
around
with
so
probably
15
or
20
people,
but
obviously
not
everyone's
full-time.
You
know.
F
So
here's
a
fun
one.
So
once
data
gets
visualized
by
me
or
my
team,
we
eventually
have
to
submit
it
to
the
imaging
data
commons,
which
is
another
nci
bucket
they've
decided
there
can't
be
any
dates
anywhere
in
any
file,
doesn't
even
matter
what
the
date
is.
It
can
be.
You
know
the
date.
My
scanner
was
turned
on.
It's
a
date
having
dealt
with
this
anytime
there's
a
date.
We
have
these
conversations
about
whether
or
not
it's
a
bad
date
or
a
good
date.
F
So
every
freaking
file
that
comes
in
I
have
to
write
well
automate,
hopefully
but
write
a
scrubber
that
basically
makes
sure
that
this
date
is
not
a
bad
date.
No,
I'm
sorry,
I'm
just
going
to
make
them
0101
whatever
year.
It
was
right
so
we're
like
originally
we're
like
we're
just
going
to
send
all
terabyte
of
data
back
to
stanford
and
make
them
do
it.
I'm
like
that'll,
take
two
years.
F
How
about
you
know
so,
then
we
have
to
do
it
because
it's
probably
faster,
but
then
we
run
into
provenance
where
now
I'm
changing
data
that
someone
else
submitted
and
so
what
the
hell
do
I
do
with
it.
I'm
not
going
to
delete
it,
but
I
also
don't
want
to
spend
again.
I
like,
I
don't
really
like
talking
about
the
money
but
storing
stuff
on
amazon
is
really
expensive
right.
F
So
if
we
have,
I
think
a
petabyte
of
storage
on
amazon
is
like
200
300
000
a
year
I
plus
or
minus
it
might
be
cheaper
now,
but
it's
it's
not
cheap
right.
So
if
I
have
to
duplicate
a
petabyte
of
imaging
data,
because
I
got
rid
of
five
bits
to
get
rid
of
a
date,
it's
a
lot
of
cash.
So
then
you
have
to
deal
with
like
well.
Stick
it
in
glacier,
okay!
Well,
then,
we
have
to
build
a
policy
to
put
it
into
glacier,
but
then
right
what?
F
If
I
screw
something
up
and
I
have
to
pull
a
policy
to
get
it
out
of
glacier
okay
well,
we
have
as
soon
as
a
file
shows
up
in
my
system.
I
give
it
a
new,
unique
id
and
everything's
tagged
to
that.
Oh,
but
now
I
change
the
file.
So
now
I
have
to
give
it
a
new
new
id.
You
know,
so
you
want
to
keep
a
data
scientist
or
someone
with
a
huge
amount
of
paranoia
or
any
or
an
accountant
employed,
there's
a
lot
of
kind
of
hidden
mandates
even
like
originally.
F
F
F
Just
slow
so
like
this
is
what
I
wrote
I
don't
know
half
a
decade
ago.
I
thought
I
would
basically
take
50
images,
create
a
new
slide
called
humantumeratlist.org,
dot,
pix
and
then
great
they'd.
Let
people
play
around
with
the
pictures.
Well,
yeah!
I
was
wrong,
so
you
know
we
have
to
de-identify
stuff
the
data
formats
that
people
are
giving
us
don't
exist,
yet
we
have
to
then
clean
up
across
all
of
these
exciting
assays.
Because,
honestly,
this
is
kind
of
you
know.
This
was
fun
to
look
at
you
know
a
decade
ago.
F
A
F
There's
there's
a
lot
of
stuff
that
can
still
be
done.
I
don't
you
know.
Students
want
to
play
with
data.
That's
not
a
problem
so
like
what
I'm
doing,
unfortunately,
for
our
first
data
release.
Is
I'm
just
eating
it
and
manually
going
through
the
first
300
images
myself
and
looking
for
the
tags,
because
other
it's
not
that
hard
for
me
to
do,
I'm
not
doing
it
for
the
other
50
000
images.
You
know
like
there's
sort
of
the
what
we're
going
to
share
publicly
and
what
and
actually
there's
you
know.
F
There's
google
has
tools
to
do
this,
whatever
whatever,
but
you
know
going
back
to
what
david
said
a
lot
is
I'm
not
a
software
engineer.
My
dad
was-
and
I
mean
I'm
weird
because
I'm
a
psychiatrist
who
spends
most
of
his
time
coding.
So
I'm
a
very
bad
software
engineer,
but
there
are
ways
to
do
this.
Well
right.
You
know.
Data
goes
in
scan
metadata,
look
for
scary
stuff,
send
it
to
some
service.
It
gives
you
some
things
back.
F
Okay,
then
we
have
to
write
our
business
rules
so,
like
here's,
a
fun
one,
I
hate
xml
xmls
have
schemas
a
lot
of
the
schemas
have
dates.
My
tool
was
detecting
that
the
schema,
which
I
can't
ever
change,
because
it's
a
schema
has
a
date
in
it.
So
I
have
to
write
a
business
logic
to
say
this
is
a
date,
but
this
is
a
good
date.
Okay,
now
now
it
flags
another
date.
Now
google
said
oh
no.
This
is
another
scary
date,
I'm
like
that's.
A
I
would
also
say
you
know:
this
is
just
a
plug
for
some
of
our
groups
where
we
have
the
cyber
infrastructure
group.
You
know,
there's
a
project
they're
working
on
for
open
data,
open
storage
network
for
high
impact,
scientific
data
or
reuse.
So
now
it's
live
as
of
2021,
it's
through
the
four
there's
four
data
hubs,
and
then
some
other
partners
and
now
it's
live
on
exceeds
so
you
can
request
storage
allocations
and
you
can
also
buy
your
own
pods.
F
F
Again
for
like
this
image
that
I
think
I'm
looking
at
is
70
or
80
gigs
right,
I'm
not
pulling
all
80
gigs
in
real
time,
because
that
would
doesn't
work
very
well.
So
my
tool
goes
in
and
does
magic,
blah
blah.
But
basically
you
know
if
the
performance
on
the
stuff
that's
on
google
versus
on
the
stuff,
that's
local,
or
on
s3
or
whatever
it's
noticeably
slower.
I
don't
know
how
much
noticeably
slower
it
is
because
it's
a
pain
to
characterize.
F
C
So
let
me
let
me
spin
a
slightly
different
question
sort
of
thinking
off
of
what
renato
was
saying
about
this
lots
of
tasks
for
students
to
do
work
and
things
like
this,
and
you
know
one
of
the
things
that
I
always
liked
to
do.
For
my
students,
when
I
teach
these
are
these
intro
classes
and
stuff
like
that
is
like
I
want
their
homework
to
like
have
an
impact.
If
I
can
the
question,
I
guess
the
part
of
this
thought
is
is,
as
some
of
this
data
becomes
openly
accessible.
C
F
Oh
yeah,
so
I
mean
you
know
if
I
have
another
small
grant
where
we're
like
working
on
de-identification
tools,
there's
already
data
in
the
tcga
that
has
identifiers
in
it
and
you
don't
have
to
go
to
medical
school
to
to
see
someone's
name
and
be
like
that's
that
that's
some
dude's
name
or
to
see
or
or
like.
This
is
a
date
you
know
so
a
lot
of
times
when
I
try
to
test
students.
Like
I
I
mean
I,
I
went
to
doctor
school
at
some
point.
F
I
don't
know
what
the
hell
I'm
looking
at.
I
guess
it's
a
it's
a
wd
76845.
I
don't
know
right
and
it
has
a
bunch
of
colors
and
I
think
cd4
is
an
antibody
for
the
immune
system.
So
there's
something
going
on.
I
don't
know
what
I'm
looking
at
right
but
like
when
I
look
at
it.
At
the
same
time,
I
think
I
just
saw
a
weird
blue
dot
right
and
I'd
be
like
that's,
probably
an
artifact
right.
I
don't
someone
drew
with
pen
or
this
is
crap
right.
F
I
don't
know
what
it
is
right.
So
there
are
ways
you
can
basically
define
the
tools
that
as
long
as
someone
has
eyes,
they
don't
know
what
it
is,
but
they
know
it
shouldn't
be
there
right.
So
even
for
like
we'll
call
it
a
de-identification
pipeline,
essentially
they
just
have
to
write
tools,
they
flag,
they
flag,
something,
and
then
I
don't
know
if
everyone
knows
what
regular
expressions
are
right,
but
you
know
you
write
a
date
module
you
go.
Oh
I
found
a
date
when
I
was
going
through
this
before
I
said
okay.
F
Well
now
I
want
to
see
if
I
can
write
something
that
will
automatically
find
this
state
right,
this
date
format
and
then
you
do
it
again
and
you
do
it
again
and
you
do
it
again.
So
basically,
I
mean
most
of
my
life
is
iterative
refinement
of
crappy
code,
I've
written
right
so
yeah
I
mean
there
are
ways,
definitely
to
you
know,
especially
even
for
this
project,
where
we're
just
looking
for
call
them
latent
dates.
A
Outside
you
know
we're
at
1201,
so
if
anybody
has
to
leave,
you
can
go
ahead.
We'll
stick
around
for
another
couple
of
minutes.
If
people
have
questions,
but
if
you
do
have
questions
and
you
have
to
go
put
them
in
the
ether
pad-
and
I
think
david
and
david
have
said
that
they're
willing
to
look
at
those
if
we're
going.
D
F
In
fact,
I
know
it's
really
not
something
we're
scared
about,
but
we
have
a
policy
that
the
person
who
is
going
to
get
the
data
after
me,
we
don't
want
to
have
the
stupid
conversation
about
dates
ever
again
right,
because
every
time
you
see
one
we're
now
getting
into
really
a
cyber
security
problem
where
we
say:
oh
crap,
they
put
an
absolute
date,
but
we
have
relevant
dates.
So
if
I'm
smart
enough-
and
I
have
enough
data-
can
I
use
this
one
date
that
probably
occurred
around
when
they
were
in
the
surgery
plus
three?
F
You
know
it's,
it's
sort
of
gotten
to
the
point
where
it's
annoying
me
where
you
know
this
is
for
nci,
so
we
can't
screw
up
and
release
data,
but
at
the
same
time
the
the
idea
that
there's
perfect
de-identification
is
a
bit
of
unless
you
spend
more
money
than
we
have
and
b.
I'm
pretty
sure
someone
on
this
call
can
tell
me
if
you
have
enough
data,
you
can
identify
anybody
right.
I
look
at
their
death
record,
they
had
a
glioblastoma,
they
lived
in
boston,
okay,
that
was
jane
smith.
F
I
don't
know
you
know
so,
but
again
you
know.
There's
you
know.
The
date
thing
is
a
good
example
of
something
where
you
know
we
need
to
build
a
workflow,
but
it's
called
shareable,
because
the
stuff
that
would
be
exposed
is
very
low
risk
what's
funny
about
it,
though,
is
I
can
guarantee
someone's
already
done
it.
The
electron
microscopy
data
we're
getting.
It's
really
big.
It's
like
many
many
many,
many
many
many
terabytes,
there's
probably
dates
in
everything.
F
Do
I
want
them
to
re-upload
it?
I
can't
make
a
dog
sit,
as
I
say.
Sometimes
right,
so
I
don't
even
know
if
they
will
they'll
say
I
don't
think
it's
a
date.
I
don't
think
it's
important
okay!
Well,
I
can't
make
you
re-upload,
I
don't
do
I
delete
it
all.
Do
I
just
run
a
script
and
do
it
myself?
What
do
I
do
with
the
old
data
right?
I
made
it
clear
on
every
call:
I'm
not
stupid
enough
to
to
be
the
final
source
of
provenance
for
any
crappy
script.
F
I
write
to
get
rid
of
anything
right
anytime.
You
touch
someone's
data,
it
becomes
your
data
so
or
I
might
have
to
rerun
the
script
or
I
learn
something
new
or
you
know
whatever
the
process
is,
so
you
get
into
all
these
like
data
governance
issues
that
are
both
interesting
and
unsolvable
by
a
co-investigator
on
a
much
bigger
project
right,
but
the
real
blockers
right.
F
It's
not
a
technical
problem
for
me
to
delete
half
a
petabyte
of
data,
but
I
usually
pay
my
taxes
and
it
would
piss
me
off
to
give
amazon
an
extra
two
hundred
thousand
dollars
for
just
literally
you
changed
four
bytes
right.
I
mean
oh,
if
it's
oh
f,
my
birthday's
0.509.75,
don't
steal
my
social
security
stuff.
If
I
make
it
oh
101.75
right,
I've
now
changed.
05
I've
changed
two
bytes
right.
The
five
becomes
a
one
and
the
nine
becomes
a
one.
F
C
F
Could
you
I'm
changing,
embedded
metadata
in
the
file,
the
the
the
csv
stuff's
cheap,
like
any
a
json
file
or
any
a
csv
file?
We
change
on
the
fly,
but
some
of
the
stuff
we'll
call
it
by
definition,
is
either
burnt.
Like
a
t
like
a
like
a
png
or
a
tiff
right,
you
get
any
picture
they
put
like
if
from
a
camera,
these
are
magical
cameras,
but
a
normal
camera.
It
actually
can
put
your
gps
coordinates
in
there.
That's
a
bad
one
right.
It
puts
the
date.
F
C
F
F
F
I
think
I
don't
know
or
what
you
know
you
get
into
these
all
these,
like
pragmatic,
whatever
you
want
to
call
it
where,
like
like,
there's
latent
costs
and
everything
right
so
again,
the
biggest
example
is.
I
did
this
for
tcga.
I
downloaded
all
20
terabytes
of
data
a
lot
right
now.
Some
of
the
data
we're
talking
about
is
even
bigger
than
that,
and
you
know
I
have
an
asset
home,
but
I
can't
I
don't
have
that
much
space.
Oh
you
know
it's
like
it's
what
these
are
fun
things
to
think
about.
F
You
know
what
also
happens
is
so
carl.
The
other
problem,
though,
is
then
instead
of
the
all
of
the
data.
I
need
to
look
at
an
image
being
in
a
bucket
that
the
nci
owns.
This
metadata
is
now
stored
in
some
other
magical
database
that,
if
you're
not
probably
some
of
the
people
on
the
call
you
don't
know
how
to
write
the
python
rest
api
script
to
graph.
In
fact
I
mentioned
this.
This
is
what
I
wanted
to
do
initially.
F
Is
you
download
the
bits?
You
then
have
a
magical
renaming
script
that
basically
pulls
the
metadata.
You
want,
puts
it
in
the
file
names
and
makes
it
not
a
total
pain
in
the
ass
to
work
with
right.
Didn't
go
over
very
well
right
and
we
even
have
this
stuff
with
semantics
names
right.
So
essentially,
my
stupid
system
basically
stores
everything
on
my
file
system
just
using
md5
sums,
so
it's
ea
7767,
I'm
like
I'm.
Looking
for
gutman
dash,
not
tumor
healthy
guy
in
the
world
dot
picture
right.
F
I
have
to
look
up
the
md5,
some
of
that
original
file
or
whatever,
but
then
we
have
this.
Other
problem
is
as
soon
as
I
call
it:
gutman's
normal
healthy
tissue
dot,
whatever
I've
now
embedded
semantic
information
in
the
label
and
then
it
turns
out
guttman
wasn't
so
healthy
because
he
has
coven.
So
now
my
file
name,
which
I
was
using
to
have
semantic
meaning,
is
now
because
later
on,
it
turned
out.
I
did
it
wrong,
so
we
actually
had
this
happen
with
the
cancer
genome
atlas,
where
oh,
they
didn't
actually
have
a
brain
tumor.
F
It
was
a
breast
tumor
or
I'm
kidding.
But
you
know
what
I
mean
like
where
you
update
something
you
know
just
because
you
have,
the
bits
doesn't
mean
whatever
you've
associated
with
are
correct
and
then
you
get
into
versioning
and
data,
and
you
know
who's
in
the
know.
To
know
that,
oh,
I
should
have
redownloaded
the
data,
because
you
know
carl.
B
F
F
A
Because
you
know
the
data,
the
importance
of
metadata,
how
to
also
get
this
across.
I
think
this
is
so
important
and
I
think
it's
lost
sometimes
in
the
idea
of
trying
to
make
things
simple
for
people
coming
into
fields
where
you
think,
okay,
you
won't
spend
years
on
these
types
of
problems,
but
these
are
really
kind
of
fundamental
blockers
to
some
of
the
things.
A
In
research
and
like
being
able
to
be
good
at
this,
there
are
practices
there
are,
you
know,
working
groups.
I
know
that
when
I
was
in
government
they
I
went
to
all
these
working
groups
for
darpa,
where
they
had
this
exact
same
problem.
So
there's
there's
people
that
have
been
thinking
about
these
things,
but
yet
the
propagation
of
it.
F
Down
I
mean
so
switching
topics
briefly.
So
at
the
va.
I
I'm
I
pretend
I'm
a
psychiatrist,
but
I
mostly
do
data
stuff
now,
because
I
know
how
to
program
so
like
we
now
have
thousands
of
people
we've
called
to
schedule
the
vaccine.
They
actually
got
it
somewhere
else,
but
we
now
have
a
flag
somewhere.
That
says
they
got
it
somewhere
else,
but
it
doesn't
show
up
on
our
numbers
because
it
has
to
be
documented
in
some
other
system.
F
So
now
we
have
to
have
nurses,
call
back
3000
people
and
then
change
another
bit
in
some
other
table
so
that
this
dashboard
gets
updated
nationally.
So
we
get
credit
and
blah
blah
blah
so
yeah.
And
then
that's
not
it's
a
data
science
problem
because
I
know
where
to
find
the
stuff
but
calling
3
000
people
is
not
my
problem.
You
know
so,
like
you
know
you
can.
F
E
F
D
F
A
C
Well,
yeah,
it
also
one
makes
me
wonder
if,
like
part
of
the
challenge
needs
to
be,
you
know,
if
you
think
about
how
get
large
file
storage
works
right,
like
you're,
only
ever
uploading
the
differences
and
then
but
the
problem
is
that
with
you?
If
you
have
like
a
a
difference
file,
that's
tracking!
Somehow
you
could.
You
can
reverse,
and
you
know
someone
could
reverse
engineer
all
the
data
that
you've
removed
out.
If
all
you've
done
is
track.
F
And
that
is
a
big,
so
the
way
aws
works,
or
at
least
the
buckets
right
a
bucket
once
it's
there,
it's
the
object's
immutable
from
what
I
can
understand
so
well.
I
know
that
because,
apparently,
in
order
to
give
me
permission
to
these
files
that
I
couldn't
read,
we
actually
have
to
copy
them
to
the
same
bucket
but
apply
a
different
policy.
F
But
then
you
actually
like
I'm
like
the
performance
sucks,
guys
well,
you're.
You
maybe
carl
appreciate
this,
but
we
put
up
all
of
our
data
and
they
just
mounted
it
with
an
s3
fuse
file
system
and
not
surprisingly,
the
performance
is
terrible
compared
to
it
being
on
a
ssd
drive
in
the
same
server
that
it
used
to
be
on
and
like
everything's
slow.
Now
guys.
I
go
yeah
because
you
did
something
stupid
that
was
not
going
to
work,
but
you
saved
money.
I
don't
know
like.
We
ran
out
of.
C
Disks
they
saved
a
lot
of
money
because
an
ssd
is
way
more
expensive
than
a
a
standard
drive.
Yeah,
there's
a
there's
all
so,
if
you,
if
you
have,
you
know
right
like
there's
a
there's
scales,
the
the
faster
you
go,
the
more
expensive
it
gets
like
yeah
cash
is
super
fast
ram
is
super
fast,
but
it's
also
super
super
expensive
for
for
gigabytes
of
data.
C
C
F
C
Because
what
and-
and
I
watched
it
there's
this
fantastic
video
that
I
used
for
my
computational
linear
algebra
class-
that
talks
about
like
the
different
ways
in
which
you
scan
the
data
depending
upon
the
kinds
of
operations
you
want
to
do,
and
so,
when
you're
trying
to
do
certain
kinds
of
masking
or
certain
sorts
of
convolutions
and
things
like
that,
you
don't
want
to
use
a
standard
data.
Read
that
reads
all
of
the
the
column
of
data,
a
bits
because
you
actually
want
you
need
low.
C
You
need
a
different
locality
and
the
probability
that
well
the
issue
with
that
is
that
until
we
reorganize
how
our
hardware
stores
the
bits
and
data,
even
when
we
reorganize
our
computation
to
improve
locality
on
the
computation,
we
haven't
actually
improved
the
storage
access
to
it.
So
all
we're
doing
is
fixing
the
hot,
the
really
really
fast
memory
usage,
but
we're
not
fixing
the
access
to
the
fast
usage
where,
if
we
were
storing
those
fi
those
image
files
differently
on
the
hard
drive
itself,
then
we
would
actually
be
able
to
improve.
C
A
So
go
ahead,
someone
has,
if
you
wanted
to
talk,
we
have
about
two
more
minutes,
two
or
two
to
five
more
minutes,
and
then
I
put
in
the
link
just
for
you
at
the
openstoragenetwork.org.
You
know.
I
know
you
already
have
a
set
up
with
aws,
but
because
of
the
it's
new
it's
going,
it
went
live
it's
like
a
distributed
system
of
pods
at
different
universities
that
are
leveraging
nsf
cyber
infrastructure
and
is
now
part
of
exceed,
and
but
they
have
from
what
I
can
tell
this.
F
A
A
D
F
So
originally
this
is
this
goes.
This
is
why
it's
not
a
phi
problem,
it's
a
policy
problem.
So
yes,
originally
we
wrote
our
policies
worried
about
phi,
but
then
the
one
of
the
upstream
places
that
will
eventually
be
the
home
of
all
of
the
data
basically
does
not
want
any
date
anywhere
in
any
file.
So
it's.
D
D
Is
it
is
it
because
they
could
not
interpret
correctly
the
hipaa
or
they
are
scared
of
hippa
or
what
is
the
reason
behind
just
scrapping
the
all
the
dates,
because
sometimes
you
might
need
some
days
to
for
especially
for
a
longitudinal
study.
Right
like
the
example,
you
are
giving
only
one
sample
of
that
person
that
many
petabytes
of
pictures,
meaning
it's
not
it's
not
taken
all
in
one
on
one
date
right.
So
no.
F
We're
requiring
them
to
date
shift
and
then
use
relative
dates
or
something
or
like
day
since
diagnosis
or
whatever.
So
I
agree
with
you,
it's
a
bit
of
an
overstatement
but
at
the
same
time,
having
even
dealt
with
irbs,
where
I
got
in
a
fight
like
10
years
ago,
I
just
wrote
my
birthday
and
sodium
144.
F
right
people.
Basically,
you
know
the
people
interpreting
these
policies,
sometimes
just
say
a
date.
The
word
date
is
in
hipaa
and
then
we
started
getting
into
these
philosophical
conversations.
If,
like
I
knew
what
month,
someone's
surgery
was
in,
and
I
knew
the
relative
date
from
some
other
data
sources,
could
I
use
an
absolute
date
and
a
relative
date
to
do
something
else.
You
know
like
it
started
getting
esoteric,
but
at
the
same
time
you
know
these
are
these
are
big
big
money
initiatives
and
if
you've
ever
dealt
with
a
data
breach,
it
sucks.
F
D
Well,
I
I
because
if
there
is
a
mandate
like
hipaa,
we
have
to
make
sure
we
follow
it
unless
it
is
modified
further,
but
I
think
all
these
examples
and
issues-
and
you
know
the
importance
of
having
some
dates
for
longitudinal
research-
should
go
into
that.
You
know
that
policy
meeting
and
slowly
change
it.