Reimagining Trust in AI
[Music]
you may have heard the same
common sense is not so common it happens
every day around us where people break
simple rules
and exemplify this notion for instance
don’t text while you drive only cross
the street when it’s green light
apparently not so common in new york
which is the very
first thing i learned from there we all
have cognitive biases
as to what common sense is and this
notion
also applies between human intelligence
and artificial intelligence
here’s an example how would you label
the sentiments of this conversation
good morning your flight is rescheduled
to 3 am
tomorrow perfect
is it positive neutral or negative to
you
in natural language processing the
algorithm will put
a positive label because it recognizes
the positive words like
good and perfect but it might not be
able to capture any
negative signal in fact detecting
sarcasm or
irony is one of the most challenging
topics in this field
then when responses from ai machines
don’t make sense
would you trust ai it is time to
re-imagine
the trusting ai to start with
it’s worthy to see what ai really is and
what we should expect from it
let’s do a little experiment think about
the phrase artificial intelligence
what is the first image that pops up in
your mind
is it something like these
every time when we talk about ai
we use images with those
octarian brainy sci-fi robots
that have little to do with ai
we should really stop using those
because
ai actually looks like this
artificial intelligence is artificial
after all
at its core ai is ultimately a process
of optimization
based on given inputs it learns the
patents from given information
and makes optimal decisions through
computations
but when ai is overly humanized or
sometimes
overhyped they seem to have the magic
power
beyond human intelligence however
is it a realistic expectation we’re
approaching but generally
we’re not there yet now let’s take a
step back
and think about why is that let’s
compare how
us humans make decisions versus how ai
does
here’s a decision for you to make say
that you
and your partner have been living in
seattle for five years
today you finally got a job offer from
your dream company
but it requires you to relocate to
boston
how would you decide there’s definitely
a long list to consider
your personal development your
relationship
timing expenses lifestyle and on and on
if you have kids 10 more items added to
your list
as you can imagine decision making is a
fairly complicated process for us
then how do ai algorithms decide
it’s utterly simple finding the value of
either
minimum cost or maximum rewards through
mathematical
equations and for this process to work
all the complexity of human
considerations has to be simplified
and quantified into several metrics or
sometimes
even a single number therefore it’s
really unrealistic
to expect a perfect transition from the
decision making of human beings
to the decision making of machines
so far ai is a hardcore optimizer with
strong computing power
and also with limitations in its
optimization
mechanism now let’s just say we do have
a perfect
algorithm today can you fully trust it
now
don’t forget there’s a condition
everything is processed
upon given inputs for machines these
inputs
are the only learning materials and the
foundation
of any trustworthy outputs
that’s why we need to fit in sufficient
unbiased
and representative data in the first
place which is
quite challenging to be honest because
in reality
we may not have enough data there are
always random factors
the data may now fully describe the real
world and the world is changing
dynamically so naturally there are
limitations
in the given inputs with that being said
with limitations in ai algorithms and in
training data
how could we trust ai well
never just trust but validate
there are a bunch of examples that ai
could bring troubles in real cases
if without validation i noticed the
project in healthcare
the motto was to predict whether a
patient would have
negative or positive diagnosis by
looking at their x-ray images
at first the predictions were extremely
accurate
after a while the doctors find out why
and it’s because
some images were taken from portable
x-ray machines
and some were taken from regular
machines
what does that mean think about what
kind of patients
get x-rayed from portable machines
the patients who couldn’t make it to the
regular machine and those are the
patients who are more likely to have
unfavorable diagnosis
turned out that this model is merely
looking at the type
of the x-ray device rather than the
actual pathology
as a matter of fact it is so common in
healthcare that patients who have severe
symptoms
get different treatments than those who
are less sick but apparently
this should not be the only evidence for
diagnosing
that’s why when given a model we should
always
validate ask what drives the model
predictions
and whether those drivers make sense to
you
and validation exercises can be done by
not only ai practitioners
it can be done by anyone for example my
lovely younger sister she’s a dancing
teacher
once i asked her to think about this and
she looked at me
how am i supposed to do is there an
easier way to validate ai
i told her yes there are ways for you to
do it easily and critically
for example validate with corner cases
corner cases adjust the situations that
occur outside of the normal cases
if a technology works in both normal and
corner cases
we know we could trust it to a larger
extent
a couple of months ago i was researching
a very popular
object detection technology it is an
algorithm that can locate and identify
objects
in images or videos so that we can do
image annotation
face recognition and so forth so i was
testing the model by
throwing out some images let’s take a
look
take the typical rush hour afternoon on
new york friday
we can see that in this busy picture the
algorithm captured a lot of items
they are located in bounding boxes with
labels
indicating the categories and the
respective probabilities
essentially those probabilities are
telling us how certain the algorithm is
so here the model is 80 sure that
in this yellow box this is a car it also
detected
the pedestrians traffic lights etc
this model is by no means perfect but it
did a fair job in general
now let’s take it up to another level by
testing some corner cases
what if we apply a picture with a
different angle
for example from an overhead view like
this one
when the shapes of items twist when we
look at them from above
would the model still hold up
now i got a bit confused you can see
that in the center
the people and the chairs were correctly
and confidently detected
while in this corner the motor view that
person as a teddy bear
it also means the flag the picture in
the frame and the table
those are the variance objects that the
model is supposed to detect
but it wasn’t able to identify them in
weird angles
let’s try another case what if we
fit in an image with incomplete
information
for example a partially covered cat face
would the detector figure this out
somehow it totally missed the cute cat
and magically detected
three ties instead i know the kitty
feels the same
astonished as you can see by
thinking critically asking what-if
questions and testing corner
cases it’s almost effortless to validate
and the bonus point here is that being
critical of science
inspires science just like we saw it
didn’t work with the covered cat face
in july 2020 the u.s national institute
of standards and technology
has tested 89 facial recognition
algorithms
and find that the error rate spiked
up to 50 percent for faces wearing musks
plenty of developer teams were actively
working on this during
covet 19 and six months later there was
a huge
improvement in january 2021
the u.s department of homeland security
tested a new algorithm
and it can identify airline passengers
with masks
96 percent of the time
wherein musks is a corner case for
facial recognition
and being aware of this corner case
accelerated
technology advancement
yet still a lot of times ai developers
might not be aware of the corner cases
or the potential consequences
it really relies on everyone to dive
deeper
and validate from different perspectives
it’s not only
one ai expert needing validation from
other ai experts
but also inputs from people outside of
this field
even if your expertise is in something
completely different
if you are a musician a nurse a lawyer
a factory worker or someone in sports
your input
matters start asking questions towards
the ai technologies
that tries to impress you or the data
conclusion that tries to convince you
we’re at a time when everyone is close
to data
close to ai technologies while we’re
also at a time
when people and ai developers are far
apart
so i’m asking for both sides to come
closer and bridge the gap
for ai practitioners in order to collect
more
corner cases and be more aware of the
impact
feedback loops and platforms should be
created and easily accessible to the
audience
the audience on the other hand is also
responsible to participate not only just
as a pure consumer a passenger
but as a feedback provider this teamwork
really matters to help ai make more
sense
and more importantly it matters for
advancing ai
as a product as a service as a policy
and
as a society now let’s go back to the
beginning of this talk
reimagining trust in ai
just like how trust is built up between
us humans
we spend some time getting to know each
other through both glorious and
imperfections
test each other with corner cases
subconsciously or consciously
see each other experiencing fun time and
down time
and work on challenges together
then eventually you know if you trust
that person
it’s the same with ai we build trust
with ai through
understanding what it really is
acknowledging its limitations
through interactions and validations and
more crucially
through collaborations trustworthy ai
is always the result of joint efforts
we need collective human intelligence to
progress
machine intelligence we’re all in this
revolutionary disruption
together and that is the idea
worth spreading thank you