Placeholder Image

字幕表 動画を再生する

  • [MUSIC PLAYING]

  • JEN GENNAI: I'm an operations manager,

  • so my role is to ensure that we're making our considerations

  • around ethically AI deliberate, actionable,

  • and scalable across the whole organization in Google.

  • So one of the first things to think about

  • if you're a business leader or a developer

  • is ensuring that people understand what you stand for.

  • What does ethics mean to you?

  • For us, that meant setting values-driven principles

  • as a company.

  • These value-driven principles, for us,

  • are known as our AI principles.

  • And last year, we announced them in June.

  • So these are seven guidelines around AI development

  • and deployment, which assigned to us how

  • we want to develop AI.

  • We want to ensure that we're not creating or reinforcing bias.

  • We want to make sure that we're building technology

  • that's accountable to people.

  • And we have five others here that you can read.

  • It's available on our website.

  • But at the same time that we announce

  • these aspirational principles for the company,

  • we also identified four areas that we

  • have considered our red lines.

  • So these are technologies that we will not pursue.

  • These cover things like weapons technology.

  • We will not build or deploy weapons.

  • We will also not build or deploy technologies

  • that we feel violate international human rights.

  • So if you're a business leader or a developer,

  • we'd also encourage you to understand what

  • are your aspirational goals.

  • But at the same time, what are your guardrails?

  • What point are you're not going to cross?

  • It's the most important thing to do, is to know what is

  • your definition of ethical AI development.

  • After you've set your AI principles,

  • the next thing is, how do you make them real?

  • How do you make sure that you're aligning with those principles?

  • So here, there are three main things

  • I'd suggest keeping in mind.

  • The first one is you need an accountable and authoritative

  • body.

  • So for us in Google, this means that we have senior executives

  • across the whole company who have the authority

  • to approve or decline a launch.

  • So they have to wrestle with some

  • of these very complex ethical questions

  • to ensure that we are launching things

  • that we do believe will lead to fair and ethical outcomes.

  • So they provide the authority and the accountability

  • to make some really tough decisions.

  • Secondly, you have to make sure that the decision-makers have

  • the right information.

  • This involves talking to diverse people within the company,

  • but also listening to your external users,

  • external stakeholders, and feeding that

  • into your decision-making criteria.

  • Jamila will talk more about engaging

  • with external communites in a moment.

  • And then the third key part of building governance

  • and accountability is having operations.

  • Who's going to do the work?

  • What are the structures and frameworks

  • that are repeatable, that are transparent,

  • and that are understood by the people who

  • are making these decisions?

  • So for that, in Google, we've established a central team

  • that's not based in our engineering and product teams

  • to ensure that there's a level of objectivity here.

  • So the same people who are building the products

  • are not the only people who are looking

  • to make sure that those products are fair and ethical.

  • So now you have your principles that you're

  • trying to ensure that people understand

  • what does ethics mean for you.

  • We're talking about establishing governance structure

  • to make sure that you're achieving those goals,

  • and the next thing to do is to ensure that you're encouraging

  • everyone within your company or the people that you work with

  • and for are aligned on those goals.

  • So making sure, one, that you've set overall goals in alignment

  • with ethical AI--

  • so how are you going to achieve ethical development

  • and deployment of technology?

  • Next, you want to make sure that you're training people

  • to think about these issues from the start.

  • You don't want to catch some ethical consideration

  • late in the product development lifecycle.

  • You want to make sure that you're

  • starting that as early as possible-- so getting

  • people trained to think about these types of issues.

  • Then we have rewards.

  • You have to make sure if you're holding people

  • accountable to ethical development and deployment,

  • you may have to accept that that might slow down

  • some development in order to get to the right outcomes--

  • making sure people feel rewarded for thinking

  • about ethical development and deployment.

  • And then, finally, making sure that you're hiring people

  • and developing people who are helping you

  • achieve those goals.

  • Next, you've established your frameworks,

  • you've hired the right people, you're rewarding them.

  • How do you know you're achieving your goals?

  • So we think about this as validating and testing.

  • So an example here is replicating

  • a user's experience.

  • Who are your users?

  • How do you make sure that you're thinking

  • about a representative sample of your users?

  • So you think about trying to test different experiences,

  • mostly from your core subgroups.

  • But you also want to be thinking about,

  • who are your marginalized users?

  • Who might be underrepresented in your workforce?

  • And therefore, you might have to pay additional attention to

  • to get it right.

  • We also think about, what are the failure modes?

  • And what we mean by that is if people have been negatively

  • affected by a product in the past,

  • we want to make sure they won't be negatively affected

  • in the future.

  • So how do we learn from that and make sure

  • that we're testing deliberately for that in the future?

  • And then the final bit of testing and validation

  • is introducing some of those failures

  • into the product to make sure that you're stress testing,

  • and, again, have some objectivity

  • to stress test a product to make sure it's achieving

  • your fair and ethical goals.

  • And then we think about it's not just you.

  • You're not alone.

  • How do we ensure that we're all sharing information

  • to make us more fair and ethical and to make sure

  • that the products we deliver are fair and ethical?

  • So we encourage the sharing of best practices and guidelines.

  • We do that ourselves in Google by providing

  • our research and best practices on the Google AI site.

  • So these best practices cover everything

  • from ML fairness tools and research

  • that Margaret Mitchell will talk about in a moment,

  • but also best practices and guidelines

  • that any developer or any business leader

  • could follow themselves.

  • So we try to both provide that ourselves, as well

  • as encouraging other people to share their research

  • and learnings also.

  • So with that, as we talk about sharing with external,

  • it's also about bringing voices in.

  • So I'll pass over to Jamila Smith-Loud

  • to talk about understanding human impacts.

  • JAMILA SMITH-LOUD: Thank you.

  • [APPLAUSE]

  • Hi, everyone.

  • I'm going to talk to you a little bit

  • today about understanding, conceptualizing, and assessing

  • human consequences and impacts on real people and communities

  • through the use of tools like social equity impact

  • assessments.

  • Social and equity impact assessments

  • come primarily from the social science discipline

  • and give us a research-based method

  • to assess these questions in a way that is broad enough

  • to be able to apply across products,

  • but also specific enough for us to think about what

  • are tangible product changes and interventions that we can make.

  • So I'll start off with one of the questions

  • that we often start when thinking about these questions.

  • I always like to say that when we're

  • thinking about ethics, when we're thinking about fairness,

  • and even thinking about questions of bias,

  • these are really social problems.

  • And one major entry point into understanding social problems

  • is really thinking about what's the geographic context in which

  • users live, and how does that impact their engagement

  • with the product?

  • So really asking, what experiences

  • do people have that are based solely on where they live

  • and that may differ greatly for other peoples who

  • live in different neighborhoods that are either

  • more resourced, more connected to internet-- all

  • of these different aspects that make regional differences so

  • important?

  • Secondly, we like to ask what happens to people when they're

  • engaging with our products in their families

  • and in their communities.

  • We like to think about, what are economic changes that

  • may come as a part of engagement with this new technology?

  • What are social and cultural changes that really do impact

  • how people view the technology and view their participation

  • in the process?

  • And so I'll start a little bit of talking about our approach.

  • The good thing about utilizing kind

  • of existing frameworks of social and equity impact assessments

  • which come from--

  • if you think about when we do new land development

  • projects or even environmental assessments,

  • there's already the standard of considering social impacts

  • as a part of that process.

  • And so we really do think of employing new technologies

  • in the same way.

  • We should be asking similar questions about how communities

  • are impacted, what are their perceptions,

  • and how are they framing these engagements?

  • And so one of the things that we think about

  • are kind of what is a principled approach to asking

  • these questions?

  • And the first one really is around

  • engaging in the hard questions.

  • When we're talking about fairness,

  • when we're talking about ethics, we're

  • not talking about them separately

  • from issues of racism, social class, homophobia,

  • and all forms of cultural prejudice.

  • We're talking about what are the issues as they

  • overlay in those systems.?

  • And so it really requires us to be

  • OK with those hard questions, and engaging with them,

  • and realizing that our technologies and our products

  • don't exist separately from that world.

  • The next approach is really towards thinking anticipatory.

  • I think the different thing about thinking

  • about social and equity impact assessments

  • from other social science research methods

  • is that the relationships between causal impacts

  • and correlations are going to be a little bit different,

  • and we really are trying to anticipate

  • harms and consequences.

  • And so it requires you to be OK with the fuzzy conversations,

  • but also realize that there's enough research,

  • there's enough data that gives us

  • the understanding of how history and contexts impact outcomes.

  • And so being anticipatory in your process

  • is really, really an important part of it.

  • And lastly, in terms of thinking about the principled approach

  • is really centering the voices and experiences

  • of those communities who often bear the burden

  • of the negative impacts.

  • And that requires understanding how

  • those communities would even conceptualize these problems.

  • I think sometimes we come from a technical standpoint,

  • and we think about the communities

  • as separate from the problem.

  • But if we're ready to center those voices and engaged

  • throughout the whole process, I think

  • it results in better outcomes.

  • So to go a little bit deeper into engaging

  • in the hard questions, what we're really trying to do

  • is be able to assess how a product will impact

  • communities, particularly communities

  • who have been historically and traditionally marginalized.

  • So it requires us to really think

  • about history and context.

  • How is that shaping this issue, and what could we

  • learn from that assessment?

  • It also requires an intersectional approach.

  • If we're thinking about gender equity,

  • if we're thinking about racial equity,

  • these are not issues that live separately.

  • They really do intersect, and being OK

  • with understanding of that intersectional approach

  • allows for a much fuller assessment.

  • And then, lastly, in thinking about new technologies

  • and thinking about new products, how does

  • power influence outcomes and the feasibility of interventions?

  • I think that the question of power and social impact

  • go hand-in-hand, and it requires us

  • to be OK with [? answering. ?] Answering might not

  • get the best answer, but at least

  • asking those hard questions.

  • So our anticipatory process is part of a full process, right?

  • So it's not just us thinking about the social and equity

  • impacts, but it really is thinking about them

  • within the context of the product--

  • so really having domain-specific application of these questions,

  • and then having some assessment of the likelihood

  • of the severity of the risk.

  • And then, lastly, thinking about what are meaningful

  • mitigations for whatever impacts that we have to developed.

  • And so it's a full process.

  • It requires work on our team in terms

  • of understanding in the assessment,

  • but it also requires partnership with our product teams

  • to really do that domain-specific analysis.

  • Centering the assessment.

  • I talked a little bit about this before,

  • but when we're centering this assessment, really,

  • what we're trying to ask is, who's impacted most?

  • So if we're thinking about a problem that

  • may have some economic impact, it

  • would require us to disaggregate the data based

  • on income to see what communities, what populations,

  • are most impacted-- so being OK with thinking about it in very

  • specific population data and understanding

  • who is impacted the most.

  • Another important part is validation.

  • And I think Jen mentioned that a lot, but really

  • thinking about community-based research engagements,

  • whether that's a participatory approach,

  • whether that's focus groups.

  • But really, how do we validate our assessments

  • by engaging communities directly and really centering

  • their framing of the problem as part of our project?

  • And then going through iteration and realizing

  • that it's not going to be perfect the first time, that it

  • requires some pull and tugging from both sides

  • to really get the conversation right.

  • So what types of social problems are we thinking of?

  • We're thinking about income inequality, housing

  • and displacement, health disparities,

  • the digital divide, and food access.

  • We're thinking about these and all different types of ways,

  • but I thought it might be helpful

  • if we thought about a specific example.

  • So let's look at the example of one

  • of the types of social problems that we

  • want to understand in relation to our products and users.

  • The topic of inequity related to food access, which

  • this map shows you--

  • and it's definitely a US context that we're

  • thinking about this question for now,

  • but also always thinking about it from a global way.

  • But I thought that this map was a good way for us

  • to look at it.

  • As you can see, the areas that are shaded darker

  • are the areas where those users might have a significantly

  • different experience when we're thinking about products that

  • give personalization and recommendations maybe

  • for something like restaurants.

  • So we're thinking about questions

  • about how those users are either included or excluded

  • from the product experience, and then we're

  • thinking about going even further and thinking about how

  • small businesses and low resource businesses

  • also impact that type of product.

  • So it requires us to realize that there's

  • a wealth of data that allows us to even go here as

  • deep as the census tract level and understand that there are

  • certain communities who have a significantly

  • different experience than other communities.

  • And so, like I said, this map is looking

  • at communities at a census tract level

  • where there's no car and no supermarket

  • store within a mile.

  • And if we want it to look even deeper,

  • we can overlay this information with income.

  • So thinking about food access and income disparity,

  • which are often connected, gives us

  • a better understanding of how different groups may

  • engage with a product.

  • And so when thinking about a hard social problem like this,

  • it really requires us to think, what's

  • the logical process for us to get

  • towards a big social problem and have very specific outcomes

  • and effects that are meaningful and are making a change?

  • And it requires us to really acknowledge

  • that there's contexts that overlays

  • all parts of this process, from the inputs that we have,

  • from the activities that we do-- which may, in my case,

  • be very much research-based activities--

  • and then thinking about what are meaningful outputs.

  • And so to go in a little bit deeper

  • in kind of this logic model way of thinking about it,

  • we have a purpose now, in thinking about the food access

  • example, to reduce negative unintended consequences

  • in areas where access to quality food is an issue.

  • We're also very aware of the context.

  • So we're thinking about the context of food access,

  • but we're also thinking about questions of gentrification.

  • We're thinking about displacement.

  • We're thinking about community distrust.

  • So we realize that this question has

  • many other issues that inform the context, not just

  • access to food.

  • But as part of the process, we're identifying resources.

  • We're thinking, where are there multidisciplinary research

  • teams that can help us think through?

  • What are our external stakeholders that

  • can help us frame the problem?

  • And then, what are the cross-functional relationships

  • that we need to build to really be

  • able to solve this kind of problem,

  • while acknowledging what our constraints are?

  • Oftentimes, time is a huge constraint,

  • and then gaps just in knowledge and comfort

  • in being able to talk about these hard problems.

  • Some of the activities and inputs

  • that we are thinking about can help

  • us get to some answers are really

  • thinking about case studies, thinking about surveys,

  • thinking about user research where we're asking user

  • perception about this issue.

  • How does engagement based on your geography

  • differ in being able to do that analysis?

  • And then creating tangible outputs,

  • some that are product interventions and really

  • focused on how we can make changes to the product,

  • but also really community-based mitigations in thinking about

  • are there ways in which we're engaging

  • with the community, ways in which we're pulling data

  • that we can really use to create a fuller set of solutions.

  • And really, it's always towards aspiring for positive effects

  • in principle and practice.

  • So this is one of those areas where

  • you can feel like you have a very principled approach,

  • but it really is about being able to put them into practice.

  • And so some of the things that I'll leave you

  • with today in thinking about understanding

  • these human impacts are really being able to apply them

  • and thinking about applying them in specific technical

  • applications, building trust through

  • equitable collaboration-- so really thinking about,

  • when you're engaging with external stakeholders,

  • how do you make it feel equitable

  • and that we're both sharing knowledge

  • and experiences in ways that are meaningful--

  • and then validating the knowledge generation.

  • When we're engaging with different communities,

  • we really have to be OK that information, data, and the way

  • that we frame this can come from multiple different sources,

  • and it's really important.

  • And then really thinking about, within your organization,

  • within your team, what are change agents

  • and what are change instruments that really

  • make it a meaningful process?

  • Thank you.

  • Now Margaret will talk more about the machine learning

  • pipeline.

  • [APPLAUSE]

  • MARGARET MITCHELL: Great.

  • Thanks, Jamila.

  • So I'll be talking a bit about fairness and transparency

  • and some frameworks and approaches for developing

  • ethical AI.

  • So in a typical machine learning development pipeline,

  • the starting point for developers is often the data.

  • Training data is first collected and annotated.

  • From there, a model can be trained.

  • The model can then be used to output content

  • such as predictions or rankings, and then downstream users

  • will see the output.

  • And we often see this approach as

  • if it's a relatively clean pipeline that

  • provides objective information that we can act on.

  • However, from the beginning of this pipeline,

  • human bias has already shaped the data that's collected.

  • Human bias then further shapes what we collect

  • and how we annotate it.

  • Here are some of the human biases that commonly contribute

  • to problematic biases and data, and in the interpretation

  • of model outputs.

  • Things like reporting bias-- where we tend to remark

  • on things that are noticeable to us,

  • as opposed to things that are typical--

  • things like out-group homogeneity bias--

  • where we tend to see people outside of our social group

  • as somehow being less nuanced or less

  • complex than people within the group that we work with--

  • and things like automation bias--

  • where we tend to favor the outputs of systems

  • that are automated over the outputs of what humans actually

  • say even when there's contradictory information.

  • So rather than this straightforward, clean,

  • end-to-end pipeline, we have human bias

  • coming in at the start of the cycle,

  • and then being propagated throughout the rest

  • of the system.

  • And this creates a feedback loop where,

  • as users see the output of biased systems and start

  • to click or start to interact with those outputs,

  • this then feeds data that is further trained

  • on-- that's already been biased in this way--

  • creating problematic feedback loops

  • where biases can get worse and worse.

  • We call this a sort of bias network effect,

  • or bias "laundering."

  • And a lot of our work seeks to disrupt this cycle

  • so that we can bring the best kind of output possible.

  • So some of the questions we consider

  • is, who is at the table?

  • What are the priorities in what we're working on?

  • Should we be thinking about different aspects

  • of the problem and different perspectives as we develop?

  • How is the data that we're working with collected?

  • What kind of things does it represent?

  • Are there problematic correlations in the data?

  • Or are some kinds of subgroups underrepresented in a way

  • that will lead to disproportionate errors

  • downstream?

  • What are some foreseeable risks?

  • So actually thinking with foresight

  • and anticipating possible negative consequences

  • of everything that we work on in order to better understand

  • how we should prioritize.

  • What constraints and supplements should be in place?

  • Beyond a basic machine learning system,

  • what can we do to ensure that we can account

  • for the kinds of risks that we've anticipated

  • and can foresee?

  • And then what can we share with you, the public,

  • about this process?

  • We aim to be transparent as we can about this

  • in order to bring about information about how we're

  • focusing on this and make it clear that this is part

  • of our development lifecycle.

  • I'm going to briefly talk about some technical approaches.

  • This is in the research world.

  • You can look at papers on this, if you're interested,

  • for more details.

  • So there are two sorts of ML--

  • Machine Learning-- techniques that we've

  • found to be relatively useful.

  • One is bias mitigation, and the other one we've

  • been broadly calling inclusion.

  • So bias mitigation focuses on removing a signal

  • for problematic variables.

  • So for example, say you're working

  • on a system that is supposed to predict whether or not

  • someone should be promoted.

  • You want to make sure that that system is not

  • keying on something like gender, which we know is correlated

  • with promotion decisions.

  • In particular, women are less likely to be promoted

  • or are promoted less quickly than men in a lot of places,

  • including in tech.

  • We can do this using an adversarial multi-task learning

  • framework where, while we predict something

  • like getting promoted, we also try and predict

  • the subgroup that we'd like to make sure isn't affecting

  • the decision and discourage the model

  • from being able to see that, removing the representation

  • by basically reversing the gradient and backpropagating.

  • When we work on inclusion, we're working

  • on adding signal for something-- trying to make sure

  • that there are subgroups that are accounted for,

  • even if they're not well-represented in the data.

  • And one of the approaches that works really well for this

  • is transfer learning.

  • So we might take a pre-trained network

  • with some understanding of gender,

  • for example, or some understanding of skin tone,

  • and use that in order to influence

  • the decisions of another network that

  • is able to key on these representations in order

  • to better understand nuances in the world that it's looking at.

  • This is a little bit of an example of one

  • of the projects I was working on where we were able to increase

  • how well we could detect whether or not someone was smiling

  • based on working with some consented gender-identified

  • individuals and having representations of what

  • these gender presentations looked like, using that

  • within the model that then predicted whether or not

  • someone was smiling.

  • Some of the transparency approaches

  • that we've been working on help to further explain to you

  • and also help keep us accountable for doing

  • good work here.

  • So one of them is model cards.

  • In model cards, we're focusing on reporting

  • what model performance is, disaggregating

  • across various subgroups, and making it clear that we've

  • taken ethical considerations into account,

  • making it clear what the intended

  • applications of the model or the API is,

  • and sharing, generally, different kinds

  • of considerations that developers should keep in mind

  • as they work with the models.

  • Another one is data cards.

  • And this provides evaluation data about,

  • when we report numbers, what is this based on?

  • Who is represented when we decide a model can be used--

  • that it's safe for use?

  • These kinds of things are useful for learners-- so people who

  • generally want to better understand

  • how models are working and what are the sort of things

  • that are affecting model performance for third party

  • users.

  • So non-ML professionals who just want

  • to have a better understanding of their data

  • sets that they're working with or what the representation

  • is in different data sets that machine learning models are

  • based on or evaluated on, as well as machine

  • learning researchers.

  • So people like me, who want to compare model performance, they

  • want to understand what needs to be improved,

  • what is already doing well, and help

  • be able to sort of benchmark and make progress

  • in a way that's sensitive to the nuanced differences

  • in different kinds of populations.

  • Our commitment to you, working on

  • fair and ethical artificial intelligence and machine

  • learning, is to continue to measure, to improve,

  • and to share real-world impact related to ethical AI

  • development.

  • Thanks.

  • [APPLAUSE]

[MUSIC PLAYING]

字幕と単語

ワンタップで英和辞典検索 単語をクリックすると、意味が表示されます

B1 中級

公平・倫理的な人工知能・機械学習のためのプレイブックを書く(Google I/O'19 (Writing the Playbook for Fair & Ethical Artificial Intelligence & Machine Learning (Google I/O'19))

  • 1 0
    林宜悉 に公開 2021 年 01 月 14 日
動画の中の単語