How bad data keeps us from good AI | Mainak Mazumdar

Transcriber: Leslie Gauthier Reviewer: Joanna Pietrulewicz

翻译人员: Yip Yan Yeung 校对人员: Helen Chang

人工智能（AI） 会在接下来的十年

为全球经济注入 16 万亿美元。

这样的经济不是由几十亿人

AI could add 16 trillion dollars to the global economy

或者几百万个工厂建造，

in the next 10 years.

而是由电脑和算法创造。

This economy is not going to be built by billions of people

我们已经看到了AI的惊人优势：

or millions of factories,

简化任务、

but by computers and algorithms.

提高效益、

We have already seen amazing benefits of AI

改善生活。

然而，谈到政策 制定的公平公正时，

in simplifying tasks,

bringing efficiencies

and improving our lives.

AI并没有兑现承诺。

However, when it comes to fair and equitable policy decision-making,

AI成为了经济的守门人，

决定谁能拿到工作，

AI has not lived up to its promise.

谁能拿到贷款。

AI is becoming a gatekeeper to the economy,

AI只是在大规模和迅速地

deciding who gets a job

巩固和加速我们

and who gets an access to a loan.

对社会影响的偏见。

AI is only reinforcing and accelerating our bias

那么，AI辜负我们了吗？

我们设计这些算法是为了输出 带有偏见的错误决定吗？

at speed and scale

with societal implications.

So, is AI failing us?

作为一个数据科学家， 我可以告诉你，

Are we designing these algorithms to deliver biased and wrong decisions?

不是算法的问题，

而是偏置数据，

它们得对这些决定负责。

As a data scientist, I'm here to tell you,

为了让AI可以被 人类和社会使用，

it's not the algorithm,

but the biased data

我们急需一次重置。

that's responsible for these decisions.

我们该关注的不是算法，

To make AI possible for humanity and society,

而是数据。

我们花了大量时间和金钱，

we need an urgent reset.

以设计和收集高质量的背景数据 为代价，扩大AI的规模。

Instead of algorithms,

we need to focus on the data.

We're spending time and money to scale AI

我们需要停下这些数据， 或我们已有的偏置数据，

at the expense of designing and collecting high-quality and contextual data.

去关注三件事：

数据基础设施、

We need to stop the data, or the biased data that we already have,

数据质量

和数据素养。

今年六月，

and focus on three things:

我们看到了一个杜克大学的

data infrastructure,

data quality

AI模型，PULSE， 产生了令人尴尬的偏差，

and data literacy.

它增强了一张模糊的图像，

In June of this year,

we saw embarrassing bias in the Duke University AI model

使它呈现了一张可辨人像。

called PULSE,

这个算法错误地将一张非白人 图像增强至了一张白人像。

which enhanced a blurry image

into a recognizable photograph of a person.

在训练集中，没有足够的 非裔美国人的图像，

This algorithm incorrectly enhanced a nonwhite image into a Caucasian image.

导致了错误的结果和预测。

African-American images were underrepresented in the training set,

也许这不是你第一次

看见AI误判黑人的图像。

leading to wrong decisions and predictions.

即使AI方法已经改进过了，

Probably this is not the first time

you have seen an AI misidentify a Black person's image.

对有些种族的忽视

依旧会产生有偏差的结果。

Despite an improved AI methodology,

研究是学术性的，

the underrepresentation of racial and ethnic populations

但是，不是所有 数据偏差都是学术性的。

still left us with biased results.

偏差有真实的后果。

This research is academic,

比如 2020 年 美国人口普查。

however, not all data biases are academic.

人口普查是

Biases have real consequences.

很多社会和经济政策制定的基础，

Take the 2020 US Census.

所以人口普查应当记录

The census is the foundation

for many social and economic policy decisions,

美国人口的 100%。

然而，由于新冠疫情

therefore the census is required to count 100 percent of the population

和国籍背后的政治问题，

in the United States.

实际上非常有可能 少算了少数群体。

However, with the pandemic

我预计会有大量少数群体的少计，

and the politics of the citizenship question,

很难定位、联系、说服、 采访他们加入人口普查。

undercounting of minorities is a real possibility.

I expect significant undercounting of minority groups

少计会带来偏差，

who are hard to locate, contact, persuade and interview for the census.

威胁我们数据基础设施的质量。

我们来看 2010 年 人口普查的少计。

Undercounting will introduce bias

在最终统计中， 有 1600 万人被忽略了。

and erode the quality of our data infrastructure.

这相当于亚利桑那州、阿肯色州、

Let's look at undercounts in the 2010 census.

俄克拉何马州、爱荷华州 当年的人口总和。

16 million people were omitted in the final counts.

This is as large as the total population

我们还可以看到在 2010 年 人口普查中，少计了

of Arizona, Arkansas, Oklahoma and Iowa put together for that year.

大约 100 万 低于五岁的儿童。

如今，少数群体的少计

We have also seen about a million kids under the age of five undercounted

在其他国家的 人口普查中也很常见，

in the 2010 Census.

因为很难联系到这些少数群体，

Now, undercounting of minorities

他们不信任政府，

is common in other national censuses,

或者生活在政治动荡区域。

as minorities can be harder to reach,

比如，

2016 年的澳大利亚人口普查

they're mistrustful towards the government

or they live in an area under political unrest.

少算了大约 17.5% 的 澳洲土著

For example,

和托雷斯海峡人口。

the Australian Census in 2016

我们估计 2020 年的少计

undercounted Aboriginals and Torres Strait populations

会远远超过 2010 年，

by about 17.5 percent.

这种偏差可能造成的 影响是深远的。

We estimate undercounting in 2020

我们来看人口普查数据的影响。

to be much higher than 2010,

and the implications of this bias can be massive.

人口普查是最权威的、 可公开获取的丰富数据，

Let's look at the implications of the census data.

它记录了人口组成和特征。

Census is the most trusted, open and publicly available rich data

虽然商家会对顾客收集

私密信息，

但是美国人口调查局 会报告完整的、公开的

on population composition and characteristics.

年龄、性别、民族、

While businesses have proprietary information

on consumers,

种族、就业情况、家庭情况信息，

the Census Bureau reports definitive, public counts

还有地理分布情况，

这是人口数据基础设施的基础。

on age, gender, ethnicity,

race, employment, family status,

当少数群体被少计时，

as well as geographic distribution,

支撑公共交通、

which are the foundation of the population data infrastructure.

住房、医疗、

保险的AI模型

更可能忽视最需要 这些服务的群体。

When minorities are undercounted,

AI models supporting public transportation,

housing, health care,

改进结果的第一步

insurance

就是让每一次人口普查数据的

are likely to overlook the communities that require these services the most.

年龄、性别、民族、种族数据库

First step to improving results

具有代表性。

is to make that database representative

既然人口普查如此的重要，

我们就要不遗余力地 让它记录 100% 。

of age, gender, ethnicity and race

per census data.

在数据质量和准确率上投入精力

Since census is so important,

we have to make every effort to count 100 percent.

对AI的可行性至关重要，

不但是为了少数有特权的人，

Investing in this data quality and accuracy

还是为了社会中的每一个人。

is essential to making AI possible,

大多数的AI系统利用已有的

not for only few and privileged,

或者为其他目的收集的数据，

but for everyone in the society.

因为很方便又便宜。

Most AI systems use the data that's already available

但是数据质量是一个 需要大量投入的领域，

or collected for some other purposes

真正的投入。

because it's convenient and cheap.

对偏差的定义、

Yet data quality is a discipline that requires commitment --

数据采集和偏差测量的关注

在快速、大规模、便利的世界

real commitment.

This attention to the definition,

不仅不受重视，

data collection and measurement of the bias,

还会被无视。

作为尼尔森公司（Nielsen） 数据科学组的一员，

is not only underappreciated --

in the world of speed, scale and convenience,

我实地走访收集了数据，

拜访了上海和班加罗尔 之外的零售店。

it's often ignored.

As part of Nielsen data science team,

拜访的目的是测量 这些店铺的零售业绩。

I went to field visits to collect data,

visiting retail stores outside Shanghai and Bangalore.

我们开出了市区，

发现了这些小店——

The goal of that visit was to measure retail sales from those stores.

不正规、交通不便利。

你可能在想

为什么我们会对这些店感兴趣？

We drove miles outside the city,

found these small stores --

我们可以在市区找一家店，

informal, hard to reach.

电子数据可以轻松地 接入数据流程，

And you may be wondering --

why are we interested in these specific stores?

便宜、方便又简单。

We could have selected a store in the city

为什么我们对这些店的数据质量

where the electronic data could be easily integrated into a data pipeline --

和准确率如此感兴趣？

cheap, convenient and easy.

答案很简单：

因为这些乡镇小店的数据很重要。

Why are we so obsessed with the quality

and accuracy of the data from these stores?

国际劳工组织表示，

The answer is simple:

because the data from these rural stores matter.

40% 的中国人

和 65% 的印度人 生活在乡村。

According to the International Labour Organization,

想像印度消费的 65%

40 percent Chinese

被排除在模型之外， 政策制定的偏差会怎样，

and 65 percent of Indians live in rural areas.

意味着这些决策会偏向 对城市比对乡村更为有利。

Imagine the bias in decision

when 65 percent of consumption in India is excluded in models,

没有这些城乡背景信息

和民生、生活方式、 经济和商品价值的信号，

meaning the decision will favor the urban over the rural.

零售品牌会在定价、广告 和营销上做出错误的决定。

Without this rural-urban context

and signals on livelihood, lifestyle, economy and values,

倾向城市的偏差会导致 错误的乡村政策制定，

retail brands will make wrong investments on pricing, advertising and marketing.

包括医疗和其他方面。

Or the urban bias will lead to wrong rural policy decisions

错误的决定不是AI算法的问题。

with regards to health and other investments.

这是数据的问题，

排除了那些早就该 被统计的地区的数据。

Wrong decisions are not the problem with the AI algorithm.

It's a problem of the data

这些背景数据才该被首先关注，

that excludes areas intended to be measured in the first place.

而不是算法。

我们来看另一个例子。

我拜访了一些俄勒冈州 偏远的房车营地，

The data in the context is a priority,

not the algorithms.

和纽约的公寓，

Let's look at another example.

邀请这些家庭加入 一些尼尔森论坛。

I visited these remote, trailer park homes in Oregon state

我们在这段时间邀请了一些家庭

and New York City apartments

参与论坛的统计，

to invite these homes to participate in Nielsen panels.

数据样本具有代表性。

Panels are statistically representative samples of homes

我们的目标是邀请所有人加入统计，

让我们可以收集这些西班牙裔 和非洲裔家庭的数据，

that we invite to participate in the measurement

over a period of time.

Our mission to include everybody in the measurement

他们都会收看有线电视。

led us to collect data from these Hispanic and African homes

根据尼尔森的数据，

这些家庭占美国 家庭的 15% ，

who use over-the-air TV reception to an antenna.

大约 4500 万人。

Per Nielsen data,

these homes constitute 15 percent of US households,

对质量的投入和关注意味着

我们要尽全力

which is about 45 million people.

从这 15% 难以 触及的人群收集信息。

Commitment and focus on quality means we made every effort

为什么这很重要？

to collect information

这是个人数可观的群体，

from these 15 percent, hard-to-reach groups.

对市场人员、品牌、

传媒公司都非常非常重要。

Why does it matter?

没有这些数据，

This is a sizeable group

市场人员、品牌和他们的模型

that's very, very important to the marketers, brands,

无法接触到这些人，

as well as the media companies.

也无法对这些非常非常 重要的少数群体做广告。

Without the data,

the marketers and brands and their models

没有广告收入，

would not be able to reach these folks,

像Telemundo和Univision 这样的电视台

as well as show ads to these very, very important minority populations.

就无法播送免费的内容，

And without the ad revenue,

包括新闻媒体，

the broadcasters such as Telemundo or Univision,

我们民主制度的基础。

would not be able to deliver free content,

这些数据对商家 和社会都十分重要。

including news media,

which is so foundational to our democracy.

我们减少AI人为偏差的 千载难逢的机会

This data is essential for businesses and society.

始于数据。

我的目标不是 争先恐后设计新算法，

Our once-in-a-lifetime opportunity to reduce human bias in AI

而是建设更好的数据基础设施，

starts with the data.

让AI更合乎伦理。

Instead of racing to build new algorithms,

我希望你能加入我的使命。

my mission is to build a better data infrastructure

谢谢。

that makes ethical AI possible.

I hope you will join me in my mission as well.

Thank you.