►
From YouTube: 10 - Fairness and Ethics in ML - Emily Denton
Description
Deep Learning for Science School 2019 - Lawrence Berkeley National Lab
Agenda and talk slides are available at: https://dl4sci-school.lbl.gov/agenda
A
Folks
now
we
have
the
last
lecture
of
the
day
we
have
with
us
Emily
Denton
from
Google's
research
and
machine
intelligence
group,
so
Emily
works
on
developing
tools
and
techniques
to
promote
fair,
inclusive
and
ethical.
Ai
is
actually
very,
very
generous
and
agreed
to
give
two
talks
at
this
at
this
school.
So
we
have
our
today
talking
about
this
and
then
is
it
tomorrow.
You're
talking
about
generative
models,
awesome
so
just
a
little
bit
more
about
Emily,
so
she's,
particularly
interested
in
detecting
and
mitigating
a
harmful
bias
in
computer
vision
systems.
A
So
when
you're
doing
machine
learning
for
people
and
social
kind
of
things,
this
kind
of
stuff
is
very
important,
but
you
can
always
think
about
how
these
things
and
I
know
analogs
in
the
kind
of
science
work
we're
going
to
do.
We
want
models
that
are
robust
against
issues
with
our
data
and
so
like
this,
so
Emily
got
our
PhD
from
a
Quran
Institute
and
in
NYU
in
machine
learning,
and
there
her
research
focused
on
unsupervised,
learning,
generative
modeling
so
like
that.
So
thank
you.
Emily.
B
So
when
machine
learning
is
typically
taught,
you
know,
we
tend
to
focus
on
a
very
narrow
range
of
metrics,
and
so
the
goal
is
typically,
you
know
pick
a
metric
optimized
for
it.
You
know,
in
that
case
model
C
is
what
we
should
pick
here
in
this
talk.
We're
gonna
be
thinking
about
other
issues
that
might
come
into
play
when
we're
thinking
about
developing
models
that
are
going
to
be
in
the
world.
B
So
I'm
just
gonna
read
this
quote
from
the
Guardian
this
from
a
couple
of
years
ago,
while
the
neural
networks
might
be
said
to
write
their
own
programs,
they
do
so
towards
goals
set
by
humans,
using
data
collected
for
human
purposes.
If
the
data
is
skewed
even
by
accident,
the
computers
will
amplify
injustice.
So
at
a
high
level.
B
Algorithmic
unfairness
refers
to
the
myriad
of
ways
in
which
harmful
societal
biases
can
get
embedded
in
algorithms
and
lead
to
unjust
or
prejudicial
treatment
of
people
based
on
all
sorts
of
different
sensitive
characteristics
like
race
or
gender
or
sexual
orientation
or
disability
status,
and
so
on
and
so
forth,
and
so
with
that
would
sort
of
really
careful
consideration
of
the
societal
context
within
which
these
systems
are
being
built.
Patterns
of
structural
and
equity
that
are
reflected
in
datasets
can
easily
become
embedded
in
these
models.
So
I'm
just
gonna
go
through
a
couple
examples
of
life.
B
What
I
mean
when
I
say
this,
but
before
that
I
want
to
just
like
make
a
real
clarification
point.
You
know,
there's
a
distinction
between
the
sort
of
phrase
biases
used
in
statistics
and
the
type
of
bias
that
I'm
referring
to
here
right.
Machine
learning
is
sort
of
at
the
core
about
discrimination
and
bias
throughout
this
talk,
I'm
going
to
be
using
bias
to
refer
to
sort
of
unjustified
discrimination.
Not
the
kind
of
you
know
core
machine
learning
use
of
the
term.
So,
okay.
So
here's
a
really
good
example.
B
B
This
was
a
proprietary
software
developed
by
company
Northpoint,
and
this
algorithm
used
information
about
defendants
or
socioeconomic
status
and
family
background
neighborhood
crime
and
other
sorts
of
things
to
reach
a
supposed
prediction
of
their
individual
risk,
and
so
ProPublica
did
an
audit
of
this
algorithm
and
they
found
that
the
tool
correctly
predicted
recidivism
about
61%
of
the
time.
But
there
was
a
clearly
racialized
pattern
in
terms
of
the
errors
that
were
made,
so
the
tool
was
twice
as
likely
to
falsely
flag
a
black
defendant
as
a
future
criminal
and
wrongly
labeling.
B
And
all
of
these
things
like
this
is
a
serious
potential
to
exacerbate
existing
inequality
in
our
society.
Again.
Machine
translation.
This
is
another
example
where
sort
of
social
biases
get
embedded
into
the
system,
and
in
this
example
here
we
see
that
sort
of
stereotypically
male
words
get
translated
with
a
male
pronoun,
whereas
stereotypically
female
words
get
defaulted
to
female
pronouns,
and
this
is
only
Google
actually
fixed
earlier
this
year.
So
that's
kind
of
good.
Ok,
so
there's
lots
of
different
ways
of
characterizing
the
different
types
of
harm
that
can
result
from
ML
systems.
B
Sort
of
you
know
material
harm.
Basically,
if
you
have
a
hiring
algorithm
that
is
disproportionately
recommending
different
groups
of
people
to
be
hired.
This
would,
you
know,
fall
into
this
category
and
then
representational
harm
occurs
when
technology
like
reinforces
stereotypes
or
diminishes
specific
groups,
and
so
thinking
about
this
harm
is
really
important,
because
you
know
it
highlights
the
fact
that
machine
learning
plays
a
really
important
role
in
representations
of
identity
in
our
society.
B
And
then
another
thing,
I
think
is
often
under
looked
in
the
m/l
fairness
world
is
the
sort
of
tendency
for
Silicon
Valley
generally
to
just
propose
technical
fixes
to
social
problems,
often
without
fully
understanding
the
social
problem
and
the
kind
of
structural
conditions
underlying
it.
And
you
know
this
is
you
know,
we're
all
most
of
us.
B
I
think
come
from
science
and
engineering
backgrounds,
and
so
we're
taught
to
like
abstract
away
the
problem
and
really
you
know,
get
to
the
meat
of
it
and
try
and
come
up
with
a
nice,
simple
solution
and
often
you
know
this
can
be
part
of
a
solution.
But
this
isn't
the
whole
thing,
and
so
I
think
it's
really
important
to
you
know
understand
when
we're
falling
into
these
kinds
of
patterns-
and
this
is
often
even
worse
when
data
driven
prediction
systems
are
like
touted
as
more
objective
and
more
neutral
because
they
are
based
on
data.
B
This
is
a
form
of
exclusion
and
subordination
built
into
the
various
ways
in
which
priorities
are
established
and
solutions
are
defined
in
the
tech
industry
and
then,
similarly,
the
sort
of
allure
of
objectivity,
x'
is
really
dangerous,
and
so
he
or
she
says
that
when
bias
is
routed
through
technoscience
and
coded
as
scientific
and
objective,
then
it
becomes
even
more
difficult
to
challenge
and
hold
individuals
and
institutions
accountable.
So
this
is
why
it's
like
really
really
important
from
the
beginning
of
the
design
process
to
understand
you
know
what
are
the
problems
we're
trying
to
fix?
B
Is
there
a
textual
solution?
Is
that
the
entire
solution,
you
know?
Are
you?
You
know
kind
of
engaging
with
the
relevant
stakeholders
and
so
on
and
so
forth.
So
just
really
quick
thing.
So
why
is
this
important
to
all
of
you,
because
you're
all
coming
from
a
different
sort
of
science
backgrounds
and
a
lot
of
the
stuff
I'm
going
to
go
through
today?
A
lot
of
the
harms
come
from
models
being
built
on
social
data
and
being
deployed
in
a
social
setting
and
I
know
that
you
know
immediately.
B
B
So
first,
this
is
a
little
cliche,
but,
like
with
great
power,
comes
great
responsibility,
you're
learning,
you
know
a
lot
of
really
really
powerful
tools
this
week,
and
you
know
you
might
initially
think
you're
going
to
leverage
these
tools
in
your
you
know
very
sort
of
narrow
discipline,
but
machine
learning
is
increasingly
touching
and
sort
of
every
aspect
of
our
lives
and
so
I
think
it's
really
important
that,
as
you
become
aware
of
these
skills,
you
understand
you
know
how
are
these
systems
being
deployed
and
and
also
just
kind
of
understanding
that
this
is
you
know
it's
a
it's
a
pretty
unregulated
area,
there's
a
lot
of
different
ways
in
which
machine
learning
models
are
being
used
and
just
kind
of
be
conscious
as
the
broader
ethical
considerations.
B
Even
if
you
don't
think
they
immediately
apply
to
your
work
right
now,
but
also
they
might
so.
Secondly,
I'd
like
to
emphasize
that,
like
no
science
or
technology
is
ever
developed
in
a
vacuum.
Technologies
are
often
depicted
as
being
neutral.
You
know
and
kind
of
developed
outside
of
the
political
and
social
contexts,
but
this
is
I.
Think
in
a
lot
of
scholars.
Argue
just
kind
of
false,
and
you
know
most
science
and
technology.
B
You
know
inadvertently
or
explicitly
embodies
different
sort
of
social
relations,
there's
a
lot
of
classic
examples
of
this
which
I'll
skim
through
for
the
sake
of
time.
But
you
know
we've
seen
that
the
ways
in
which
we
design
the
material
world
has
a
potential
to
reflect
and
reinforce
social
hierarchies
or
also
subvert
them,
and
you
know.
Similarly,
this
is
an
example
of
like
physics,
so
you
know,
people
might
say.
Physics
is
physics,
it's
neutral,
it's
science,
but
the
sort
of
modern
photography
as
an
example.
That
is
very,
not
value
neutral.
B
So
modern
photography
was
developed
with
a
very
white
norm
coated
into
camera
sensors-
and
you
know
this
is
this
is
now
you
know-
affects
a
lot
of
different
sensors
that
are
deployed
in
the
world
and
it
took
a
very
long
time,
for
you
know
the
photography
and
Industry
to
even
take
notice
of
this.
So
that's
another
great
paper,
I'll
direct
you
to
read
by
Ben
green
basic
argument
put
forward
is
that
data
scientists
just
need
to
recognize
themselves
as
political
actors
engaged
in
basically
the
normative
construction
of
different
aspects
of
society,
so,
okay
and
then.
B
Finally,
this
is
important
for
you,
because
a
lot
of
the
best
practices
that
we'll
see
for
sort
of
ethics,
informed
design
and
development
are
also
just
good
practices.
Generally,
you
know
a
lot
of
them.
Things
deal
with
some
accountability
and
transparency
and
interpretability,
and
that's
important
for
everybody
in
machine
learning
cool,
so
typical
ml
model
paradigm
is
kind
of
broken
down
into
data
collection
and
then
model
choices
and
a
common
way
of
thinking
about
this.
That
a
lot
of
people
hold
I
think
is
the
data.
B
You
know
it's
kind
of
like
reflects
the
world
and
then,
if
we
fit
an
algorithm,
the
better
we
fit
it,
the
better
a
model
is
going
to
be
and
I'm
also
going
to
be
focusing
a
little
bit
on
the
kind
of
like
downstream
use
cases
of
all
of
this.
So
starting
with
data,
a
lot
of
people
think
that
data
just
reflects
the
world.
As
I
said,
it's
not
Davis,
never
neutral.
B
It's
always
some
kind
of
representation
of
reality
filtered
through
you
know
different
sort
of
human
processes,
so
I'm
gonna
go
through
just
a
couple
different
types
of
biases
that
might
get
into
your
data
set
again.
A
lot
of
the
harms
are
gonna
result
from
these
are
sort
of
within
the
social
setting,
but
understanding
data
set
bias
is
relevant
sort
of
everywhere
that
we
are
using
data,
so
sampling
bias.
Obviously
you
know
this
occurs
when
a
data
set
is
not
representative
of
the
underlying
population
interest.
B
You
know
we
see
a
lot
of
common
image,
datasets
and
machine
learning
exhibiting
different
sort
of
gender
and
racial
and
Geographic
biases.
You
may
also
have
biases
based
on
the
types
of
instruments
you're
using
to
collect
your
data
sets.
You
know
times
of
day
and
conditions
and
all
of
these
types
of
things
it's
really
important
to
be
cognizant
of
all
of
this
when
you're
developing
your
data
sets.
B
B
This
is
just
another
visualization
that
kind
of
highlights
at
about
60
percent
of
the
data
comes
from
the
six
most
represented
countries
in
North,
America
and
Europe
also
unequal
distribution
of
demographics
within
each
class.
This
is
also
an
important
thing,
so
this
data
set
found
sort
of
significant
gender
biases
with
different
activities
that
were
present
in
the
data
set,
so
human
reporting
bias.
This
is
another
thing.
This
basically
means
the
frequency
with
which
people
sort
of
write
about
actions
and
outcomes
doesn't
reflects
or
real-world
statistics.
B
So
this
is
a
nice
example
of.
Like
word
learning
texts.
You
know
if
you're
just
like
trained
to
model
on
text,
you
would
think
that
people
were
murdered
way
more
than
they
exhaled.
But
obviously
this
is
not
true.
So
again,
like
the
things
that
our
data
are
telling
us
are
not
necessarily
representative
of
what
is
actually
happening
in
the
world.
I
like
this
example.
B
So
you
know,
if
you
ask
you
what
they
see
here,
they'll
typically
say
like
bananas
or
a
bundle
of
bananas,
and
if
you
show
people
this
image
now
they'll
say
green
bananas,
and
so
this
is
because
people
tend
to
sort
of
mention
the
things
that
are
not
prototypical,
so
yellow
is
like
very
parts
of
Bill
banana,
so
it
won't
be
mentioned.
Green
is
like
oh,
this
is
novel
and
interesting
and
sailing
it
to
this
particular
image,
and
so
I'll
mention
it.
So
this
can
also
carry
over
into
sort
of
social
stereotypes
right.
B
B
Furthermore,
so
stereotypes
are
sort
of
internalized
associations
that
you
know
occur
through,
like
natural
processes
of
learning
and
categorization
and
implicit
biases
are
super
pervasive.
They
operate
largely
unconsciously
and
they
can't
automatically
influence
the
ways
in
which
people
see
the
world,
and
these
are
really
really
important
when
we're
having
you
know,
people
annotate
our
datasets,
right
and
I
think
this
isn't
like
really
really
overlooked
in
machine
learning.
Is
you
know
what
types
of
cultural
biases
are
getting
into?
B
Your
data
sets
just
purely
through
the
annotation
process
right
so
again
here
this
might
look
like
captioning
one
of
these
images.
Dr.
and
one
of
these
images.
Nurse
racial
stereotypes
are
also
really
important.
I
mean
again,
it's
really
important
to
like
tie
these
back
to
the
kind
of
social
setting
in
which
these
systems
may
be
deployed
cool.
So,
ultimately,
data
is
not
a
neutral
reflection
of
reality,
just
really
important
to
remember
through.
B
B
Do
some
adversarial
testing
find
some
new
data?
Try
and
understand
these
patterns,
and
again
this
has
happened
before
you
know
we've
kind
of
seen
throughout
history.
You
know
light-skinned
male
body
is
being
taken
kind
of
consistently
as
the
default
time
and
time
again,
we've
seen
this
with
like
camera
sensors,
as
I
mentioned
earlier,
you've
seen
this
with
crash-test
dummies.
We
see
it
in
like
medical
research,
so
the
types
of
things
that
were
encountering
here
you
know
they're,
not
new,
cool,
so
again,
disparities
in
accuracy.
This
is
looking
at.
B
B
And
so
these
examples
here,
like
none
of
these
comments,
are
toxic,
but
the
new
model
predicted,
otherwise,
just
because
the
sources
where
the
model
was
trained
on
there
were
you
know,
sort
of
certain
identities
were
overwhelmingly
referenced
in
offensive
ways
and
under
representative
for
positive
ways:
smart
reply,
again
cool
I'm,
gonna
kind
of
skip
through
some
of
these
just
for
the
sake
of
time,
but
you
can
look
through
the
slides,
there's
a
lot
of
examples
of
this
cool
so
like
some
solutions.
So
again,
many
of
the
solutions
start
at
the
data
level.
B
Collecting
representative
training
data
is
obviously
a
first
step.
This
is
an
example
of
some
studies
that
did
this
IBM.
Releasing
diverse
face
data
testing
models
on
representative
data
to
try
and
find
braking
points
is
kind
of
like
adversarial.
Testing.
Here
is
really
important,
algorithmic
audits.
So
this
is
some
work
where
folks
were
just
kind
of
looking
at
public.
Api
is
and
doing
some
really
thorough
and
rigorous
testing
data
set
transparency
efforts.
So
this
is
really
important
and
I
think
this.
B
This
is
an
example
of
some
that
applies
like
very,
very
broadly
I
love
this
to
sort
of
catch
on
more
widely.
So
there's
a
couple.
Different
frameworks
for
comprehensive
data
set
documentation
that
have
been
proposed,
such
as
data
sheets
for
data,
sets
nutrition
labels
data
statements.
Each
of
these
frameworks
basically
provide
a
slightly
different
framework
for
documenting
data,
set
collection
and
annotation
methodologies,
and
so
the
overarching
goals
of
these
frameworks
are
twofold.
So,
first
for
data
set
creators,
the
aim
is
to
provide
a
process
that
encourages
a
sort
of
reflexive.
B
Look
at
you
know
the
way
in
which
it
was
created
and
distributed
and
maintained,
and
well
as
making
clear
any
sort
of
underlying
assumptions
that
went
into
the
data
set
creation,
and
then
the
data
set
creators
were
also
encouraged
to
kind
of
think
about
harms
and
implications
that
use
some
stuff
like
that,
and
then
on
the
data
set
consumer
side.
This
really
facilitate
sort
of
informed
decision
making.
So
I
think
this
is
something
that
should
just
generally
be
adopted
by
people
who
are
creating
and
putting
out
data
sets.
B
So
basically
in
summary,
you're
working
with
data,
you
know
collecting
data
annotating
data
using
existing
data,
there's
a
wide
range
of
questions
that
I
think
people
should
always
be
asking
such
as
you
know.
Where
did
the
data
come
from?
Who
collected
it
as
the
data
representatives
or
measurement
error,
reporting
bias?
If
there's
annotations,
are
they
subjective?
Is
there?
You
know
stats
on
inter
annotator
agreement
stuff
like
that,
you
know:
could
the
annotations
themselves
be
reflecting
harmful
social
stereotypes?
Does
the
data
itself
reflect
patterns
that
would
be
harmful
to
reproduce
or
amplify?
B
Okay,
so
moving
on
I'm
gonna
talk
a
little
bit
about
sort
of
designing
and
testing
with
fairness
and
inclusion
in
mind.
So
so
the
first
thing
we
could
consider
is
just
looking
at
multiple
different
evaluation
metrics
when
you're
training
and
evaluating
your
model-
and
this
is
important,
because
each
evaluation
metric
really
provides
different
types
of
information
about
your
model
and
understanding
the
implications
of
different
types
of
errors
is
useful.
To
kind
of
you
know,
help
trade
off,
you
know
acceptable
trade-offs
between
false
positives
and
false
negatives,
and
so
on.
B
So,
in
addition
to
looking
at
multiple
different
error,
metrics,
something
that
you
can
do
is
break
down
your
quantitative
analysis
by
different
groups
of
data.
So
within
the
AML
fairness
community,
this
typically
looks
like
breaking
down
your
data
set
along.
Maybe
demographic
lines,
cultural
lines-
you
know
typical
lines,
but
you
could
also
imagine
doing
this
type
of
analysis
broken
down
by
sort
of
conditions
under
which
the
data
was
obtained.
B
Different
sort
of
devices
stuff
like
that
really
just
trying
to
understand
more
fine-grained
patterns
in
in
in
your
in
your
models
performance.
So
this
is
just
a
good
example
of
some
early
work
that
did
this.
This
is
a
gender
shades
paper.
They
adopted
this
approach
and
the
key
thing
is
to
kind
of
look,
not
just
a
unitary
group,
so
here
are
sort
of
gender
based
groups
and
skin
tone
based
groups,
but
also
intersections
of
those
groups.
Because
again,
this
can
kind
of
tell
you
more
than
this
kind
of
aggregated
statistics
can
so
model
cards.
B
This
is
another
really
nice
framework.
It's
kind
of
the
complementary
to
the
data
cards
and
data
set
documentation
frameworks
that
I
was
just
proposing,
and
so
the
idea
here
is
a
formalized
set
of
protocols
that
sort
of
put
forward
model
evaluations
with
fairness
and
inclusion
in
mind
so
actually
before
I
go
on
the
goals
of
this
are
basically
twofold
right.
We
have
the
again.
This
is
very
analogous
to
the
the
data
set
annotations
or
data
set
on
documentation
stuff,
so
for
model
creators
right.
B
It
encourages
this
kind
of
like
thorough
and
critical
evaluation,
then
also
like
a
really
thorough
consideration
of
intended,
uses
and
stuff
like
that,
and
then
for
the
model.
Consumers
again.
This
is
providing
kind
of
informed
if
facilitate
informed
decision-making.
So
basically,
what
this
does
is
we
see
sort
of
model
details
broken
down
intended,
uses
broken
down
and
made
made
clear,
different
sort
of
factors
relevant
to
analysis.
So
again,
this
might
be.
B
You
know,
breaking
down
analysis
by
different
devices,
different
conditions,
different
subgroup
stuff,
like
that,
a
you
know,
kind
of
brief
summary
of
what
the
training
data
looked
like.
What
the
testing
data
looked
like
different
metrics
are
gonna
be
proposed.
Why
are
these
relevant
metrics?
What
are
the
trade-offs,
the
different
metrics
and
then
the
actual
kind
of
quantitative
analysis
itself
is
proposed
and
then
there's
additionally,
an
ethical
considerations
section
which
I
think
is
also
you
know
increasingly
important
to
just
kind
of
put
in
any
kind
of
model.
B
So,
okay,
so
now
I'm
going
to
get
into
some
things
called
fairness
definitions.
So
there's
been
a
real
flurry
of
work
recently
on
translating
complex
social
notions
of
fairness
into
precise
mathematical
formulations.
These
are
frequently
referred
to
as
fairness,
definitions
and
I
have
fairness
in
quotes
here,
because
I
think
it's
like
really
important
to
recognize
that
these
you
know
these
are
mathematical
like
abstractions
and
formulations,
and
they
are
in
attempts
to
capture
different
social
notions
of
fairness.
B
So,
generally
speaking,
these
definitions
assume
that
each
data
instance
is
associated
with
some
protected
attribute.
These
attributes
will
frequently
reflect
different
protected
groups
in
society
that
have
often
been
historically
marginalized,
and
these
attributes
are
sort
of
leveraged
in
different
ways.
Throughout
and
many
of
the
definitions,
try
and
capture
something
related
to
a
couple:
different
notions
of
social
fairness,
disparate
treatment
and
disparate
impact.
Disparate
treatment
is
often
thought
of
as
like
explicit
discrimination
based
on
some
sensitive
attribute.
Where,
as
disparate
impact
is,
is
this
sort
of
more
like
long
term?
B
B
So
one
thing
that
people
have
tried
is
like
fairness
through
unawareness,
and
this
basically
just
means
don't
let
your
predictor
have
access
directly
to
any
of
the
sensitive
attributes
like
race
or
gender,
and
so
on.
This
framework
has
pretty
severe
limitations.
Obviously,
just
because
sensitive
attributes
have
a
lot
of
proxies.
So,
for
example,
you
know:
do
the
history
of
segregation
in
this
country?
Zip
code
is
a
really
really
good.
B
So
demographic
parity
is
another
definition
that
people
have
put
forward
and
this
basically
asks
that
the
rate
of
positive
predictions
across
different
groups
again,
these
groups
might
be
defined
along
racial
or
gender
lines.
Just
ask
the
the
rate
of
positive
predictions
is
equal,
regardless
of
that
sensitive
variable,
and
so
you
know
if
we
had
like
a
system
that
was
like
I,
don't
know
deciding
whether
or
not
to
hire
people
for
the
job
of
a
CEO,
then
this
definition
would
say
you
have
to
hire.
B
You
know
equal
numbers
of
men
and
women
for
example,
so
a
quality
of
odds
is
a
another
condition.
So
this
is
basically
mathematically
the
same
as
the
last
one,
except
now
we're
conditioning
on
the
true
value
of
the
target
variable
and
so
okay,
so
yeah
I
believe
there's
an
entire
talk
later
this
week
on
interpretability
methods,
so
I'm
not
going
to
go
into
too
much
detail.
That's
kind
of
like
speak
about
a
few
things.
Basically,
but
basically,
you
know
girl
networks,
their
black
boxes,
they've
taken
inputs,
they
produce
outputs.
B
This
is
cool,
they're
really
powerful,
but
it
also
means
that,
if
we're
using
these
systems
in
ways
that
are
going
to
be
affecting
people
and
like
really
significant
ways,
we
need
to
kind
of
understand
like
why
are
predictions
being
made.
You
know
this
is
important
for
in
terms
of
understanding
causes
of
different
types
of
biases,
but
also
giving
people
sort
of
a
feeling
of
control
and
a
mechanism
of
like
you
know,
coming
back
and
saying:
hey
I
feel
like
this
decision
was
unfair
and
just
so
tea
cab
is
a
really
nice
method.
B
I
believe
that
the
creator
of
tea
cab
is
coming
tomorrow,
so
I
think
she'll
go
into
a
lot
further
depth,
but
I
I
really
like
this
method.
Basically,
it
kind
of
lets
you
ask
the
question
of
like
you
know
how
important
is
the
concept
of
gender
for
this
dr.
classifier
and
it's
an
interpretability
method
that
works
in
a
kind
of
high
level
conceptual
space,
as
opposed
to
providing
very
low
level
like
pixel
explanations,
and
so
it's
quite
nice
in
terms
of
understanding
bias
of
classifiers
counterfactual
methods
are
another
really
nice
method.
B
The
idea
here
is
to
try
and
isolate
the
causal
effects
by
asking
questions
of
the
form
you
know
if
what
this
one
specific
thing
had
changed,
all
else
being
equal,
how
might
the
prediction
be
different
and
there's
a
huge
sort
of
history
of
these
types
of
analyses
being
done
outside
of
the
machine
learning
context,
so
these
are
referred
to
as
like
audit
studies,
so
people
have,
you
know,
looked
at
bias
and
hiring
decisions,
for
example
by
like
sending
out
equal
sets
of
resumes
with
names,
just
change,
so
you
would
have
like
a
you
know:
stereotypically
female
name
or
a
male
name
or
names
that
correlate
with
different
racial
groups
and
then
looking
at
the
kind
of
callback
rates
where
everything
else
in
the
resume
stays.
B
B
These
papers
are
sort
of
interesting
to
read
if
you're
interested
in
machine
learning
fairness,
but
I
would
follow
them
immediately
by
this,
with
this
paper
by
ISA,
who
talks
a
lot
about
how
race
and
gender
are
not
really
entities
that
can
have
counterfactual
causality,
and
this
has
been
kind
of
looked
at
in
social
statistics-
a
lot
there's
a
huge
history
of
this.
It
sort
of
really
only
makes
sense
to
talk
about
something
having
counterfactual
causality.
B
If
it
is
like
meaningful
to
talk
about
like
two
units
that
are
identical,
except
for
the
one
thing
and
just
the
ways
in
which,
like
race
and
gender
structure,
every
aspect
of
our
society
like
this,
this
is
kind
of
a
nonsensical
thing
to
talk
about
so
again
interesting
papers,
but
I
think
that
needs
to
be
followed
up
with,
like
a
much
more
nuanced
understanding
of
of
how
sort
of
racial
and
gender
based
oppression
work
in
our
society,
so
counterfactual
fairness
and
text.
This
is
a
nice
thing.
B
Basically,
you
know
if
you
flip
certain
tokens
and
text.
How
does
a
models
prediction
change,
and
this
is
another
model-
that
kind
of
uses,
generative
techniques,
basically
to
manipulate
images
and
see
how
the
classifiers
response
changes,
cool,
so
visualizations
and
other
exploratory
tools
can
also
answer
questions
and
you
know,
point
to
surprising
issues
and
machine
learning
systems.
So
this
is
a
tool
called
facets.
It's
an
open
source,
visualization
tool
that
I
believe
chaotic,
Google
and
I
should
know
that
pretty
sure
came
out
of
Google
and
it
can.
B
It
can
basically
help
you
understand
and
analyze
machine
data
sets.
So
go.
Look
at
that
and
then
tensor
flow
model
analysis,
TF,
MA,
it's
another
tool.
This
provides
scalable
slice
or
full
pass
metrics
lets
you
look
at
your
model,
performance,
broken
down
by
different
subgroups
and
there's
really
sort
of
nice
aggregate
statistics
or
another
useful
tool.
B
So
another
important
thing
to
consider
is
sort
of
you
know
throughout
all
of
these
questions
like
who
to
say
and
who
was
consulted
in
the
design
and
development
process-
and
you
know,
as
we've
seen,
machine
learning
models
they
learned
from
historically
collected
data
sets,
and
so
populations
that
have
you
know
experienced
individual
or
structural
biases
in
the
past
are
often
the
most
vulnerable
to
harm
resulting
from
these
models.
And
so
you
know
it's
really
important
to
include
you
know:
voices
of
domain
experts,
but
also
you
know,
sort
of
perspectives
of
people
who
lived
experiences.
B
So
this
is
just
a
nice
paper,
because
I
focused
a
lot
on
kind
of
vision
and
LP
applications
so
far,
and
so
this
is
sort
of
looking
at
machine
learning
used
within
a
clinical
setting,
and
you
know
we
see
machine
learning
being
increasingly
used
to
improve
diagnosis
and
treatment
and
health
system
efficiency.
Basically,
and
this
model
really
looks
at
how
sorts
of
model
design
and
data
collection
and
all
of
these
things
could
you
know,
be
done
in
a
really
like
proactive
way
to
increase
health
equity
cool.
So
so
we've
looked
at
data.
B
So
yeah
earlier
I
talked
about
some
facial
analysis
systems
and
how
a
lot
of
different
studies
have
been
done
to
show
that
they
failed
to
recognize
darker-skinned
individuals,
and
so,
while
understanding
these
kinds
of
differential
patterns
and
performance
is
important,
there
is
a
lot
of
examples
of
technologies.
We're
just
equalizing
statistics
across
different
groups
is
not
going
to
be
sufficient
to
solving
the
problem
so
face.
Recognition
is
an
example
where
the
technology
is
dangerous.
You
know
if
it
works,
really
well
for
everybody
and
it's
dangerous.
B
B
B
Oh
there's
no
hits,
but
this
guy
looks
a
lot
like
Woody
Harrelson,
and
so
they
go
off
and
they
google
Woody
Harrelson
and
they
pull
a
photo
of
Woody
Harrelson
and
they
feed
that
photo
into
the
face
recognition
system
get
a
hit
from
that
and
then
go
investigate
that
lead,
and
so
obviously
this
is
not
the
way
machine
learning
models
work.
This
is
like
dark.
This
is
the
title.
B
Scenarios
in
which
you
might
be
used
is
ultimately
if
these
tools
are
gonna
be
released
to
the
world
like
we
need
to
be
thinking
about
these
things
like
from
the
beginning,
so
gender
classification.
This
is
another
example
like
there's,
there's
no
way
to
like
make
gender
classification
like
good,
really
like
this,
isn't
an
issue
of,
like
you
know,
minimizing
area
statistics
across
different
groups
and
again
some
like
further
reading
here,
sexual
orientation
classification.
Another
thing
that,
like
just
don't,
build
it.
This
is
this
a
couple
papers
that,
like
just
won't
die.
They
keep.
B
You
know
resurfacing
simcha.
Some
colleagues
of
mine
wrote
this
great
medium
post
last
year
that
just
kind
of
like
debunked
one
of
these
papers
and
yeah
read
that
it's
good
and
yeah.
This
is
a
startup
that
markets
itself
as
being
the
first
to
technology
on
first,
a
market
with
proprietary,
computer
vision
and
machine
learning,
technology
for
profiling,
people
and
revealing
their
personality
is
based
only
on
their
facial
image
and
it'll-it'll
predict
things
like
high
IQ
and
white-collar
offender
and
terrorists
just
from
an
image
so
again,
like
just
think
of
it,
if
you're
building
something.
B
B
So
pretty
yeah!
This
is
another,
this
isn't
from
the
same
startup.
But
this
is
another
paper
that
came
out
predicting
criminality.
Again
some
colleagues
of
mine.
They
wrote
a
blog
post,
basically
where
they,
you
know
we're
kind
of
showing
that,
like
this
article
is
all
about
like
all
like
the
angles
between
like
the
nose
and
the
corners,
your
lips,
like
that's
predictive,
but
that's
also
like
predictive
of
rounding
at
smiling.
So
like
really,
what
are
you
doing
here
so
yeah?
B
B
So,
just
like
be
aware
of
the
history
of
the
types
of
problems
that
you're
trying
to
solve
and
the
ultimate
impact
that
it
could
have
and
yeah.
This
is
a
good
blog
post.
This
is
a
blog
post
in
response
to
that,
like
predicting
criminality,
but
it
also
kind
of
like
goes
through
a
whole
history
here,
so
I
would
recommend
deleting
that
as
well
and
yeah.
Overall
social
context
is
important.
B
So
a
lot
of
machine
learning
fairness
work
focuses
very
specifically
on
equalizing
arrow
statistics
across
different
groups,
and
this
is
like
a
really
important
part
of
understanding
how
technologies
are
working
and
a
really
important
thing
to
consider.
When
you're
you
know,
building
and
designing
machine
learning
technologies,
but
kind
of
assessing
fairness,
purely
an
algorithmic
level
is
insufficient.
Ultimately,
you
know
we
need
to
be
understanding.
How
are
these
systems
going
to
be
used?
B
What
are
the
social
structures
that
they're
going
to
slot
into,
and
and
really
thinking
about,
that
sort
of
fully
contextualized
understanding
when
we're
trying
to
build
fair
and
equitable,
just
inclusive
machine
learning
models
and
yeah?
So
I
kind
of
like
initially
phrase
this
like
this
right?
You
had
your
data,
you
had
your
model
and
you
might
think
about
intended
uses
and
hopefully
you're
documenting
all
of
this,
and
hopefully
you're.