How bad data keeps us from good AI | Mainak Mazumdar

Transcriber: Leslie Gauthier Reviewer: Joanna Pietrulewicz

譯者: C Leung 審譯者: 麗玲 辛

未來 10 年，AI 能為全球經濟

增加 16 萬億美元。

這種經濟不是由數十億人

AI could add 16 trillion dollars to the global economy

或數百萬家工廠所建成，

in the next 10 years.

而是由電腦和演算法。

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

我們已經看到 AI 在簡化任務、

or millions of factories,

but by computers and algorithms.

提高效率和改善生活方面 帶來的驚人好處。

We have already seen amazing benefits of AI

從這 15% 難以接觸到的 羣體中收集信息。

in simplifying tasks,

bringing efficiencies

and improving our lives.

AI 還未兌現它的承諾。

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

AI 正成為經濟的守門人，

決定誰能獲得工作，

AI has not lived up to its promise.

誰能獲得貸款。

AI is becoming a gatekeeper to the economy,

AI 只是更快速、更大規模地 加深、放大我們的偏見，

deciding who gets a job

and who gets an access to a loan.

影響社會。

AI is only reinforcing and accelerating our bias

那麼，AI 是否會讓我們失望？

我們是否設計這些演算法， 去作出偏頗及錯誤決定？

at speed and scale

with societal implications.

So, is AI failing us?

作為數據科學家，我在此告訴你，

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

並不是演算法，

而是偏誤的數據

造成這些決定。

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

為了讓 AI 能造福人類社會，

it's not the algorithm,

but the biased data

我們急需重置。

that's responsible for these decisions.

我們要關注的不是演算法，而是數據。

To make AI possible for humanity and society,

我的任務不是趕快建立新的演算法，

we need an urgent reset.

卻忽略了設計和收集 高品質、有脈絡的數據。

Instead of algorithms,

we need to focus on the data.

We're spending time and money to scale AI

我們需要停止使用現有的偏誤數據，

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

而專注於三件事：

數據基礎建設、

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

數據品質

和數據素養。

今年 6 月，

and focus on three things:

我們看到杜克大學，一個名為 PULSE 的 AI 模型尷尬偏誤。

data infrastructure,

data quality

and data literacy.

它將一張模糊的圖像

In June of this year,

we saw embarrassing bias in the Duke University AI model

增強呈現為一張可識別的人物照片。

called PULSE,

該演算法錯誤地將非白人圖像 增強呈現為白人圖像。

which enhanced a blurry image

into a recognizable photograph of a person.

在訓練集中， 非裔美國人圖像的代表性不足，

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

導致誤判和錯的預測。

African-American images were underrepresented in the training set,

這可能不是你首次

看到 AI 誤認黑人圖像。

leading to wrong decisions and predictions.

儘管 AI 方法有所改進，

Probably this is not the first time

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

但種族和民族人口代表性不足

仍然給我們偏誤結果。

Despite an improved AI methodology,

這項研究是學術性的，

the underrepresentation of racial and ethnic populations

但並非所有數據偏誤都是學術性。

still left us with biased results.

偏誤會帶來嚴重後果。

This research is academic,

以 2020 年美國人口普查為例。

however, not all data biases are academic.

人口普查

Biases have real consequences.

是許多社會和經濟決策的基礎，

Take the 2020 US Census.

因此人口普查必須對美國 100% 的人口進行統計。

The census is the foundation

for many social and economic policy decisions,

然而，由於新冠疫情

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

和公民身份問題的政治因素，

in the United States.

少計少數族裔的情況確實存在。

However, with the pandemic

難以找出、聯繫、說服 和訪談的少數族裔羣體，

and the politics of the citizenship question,

我預計他們在人口普查中 數量會被嚴重低估。

undercounting of minorities is a real possibility.

I expect significant undercounting of minority groups

少計會帶來偏誤，

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

也會削弱我們數據基礎設施的品質。

讓我們看看 2010 年 人口普查中的少計情況。

Undercounting will introduce bias

在最終統計有 1600 萬人被忽略，

and erode the quality of our data infrastructure.

相當於亞利桑那州、阿肯色州、

Let's look at undercounts in the 2010 census.

俄克拉荷馬州和愛荷華州 當年人口的總和。

16 million people were omitted in the final counts.

This is as large as the total population

我們還看到在 2010 年人口普查，

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

少計了約 100 萬名五歲以下兒童。

在其他國家的人口普查中，

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

少計少數族裔的情況很常見，

in the 2010 Census.

因為少數族裔或更難被接觸到，

Now, undercounting of minorities

他們不信任政府，

is common in other national censuses,

或者生活在政治動盪的地區。

as minorities can be harder to reach,

例如，

2016 年澳洲人口普查

they're mistrustful towards the government

or they live in an area under political unrest.

少計了約 17.5% 原住民 和托雷斯海峽人口 。

For example,

the Australian Census in 2016

我們估計，2020 年的少計比例

undercounted Aboriginals and Torres Strait populations

將遠高於 2010 年，

by about 17.5 percent.

而這種偏誤的影響可能非常深遠。

We estimate undercounting in 2020

讓我們來看看人口普查數據的影響。

to be much higher than 2010,

and the implications of this bias can be massive.

人口普查是最可靠、最開放和最公開，

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

關於人口組成和特徵的豐富數據。

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

雖然企業擁有關於消費者的

專有信息，

但人口普查局會報告 明確、公開的統計數據，

on population composition and characteristics.

包括了年齡、性別、民族、

While businesses have proprietary information

on consumers,

種族、就業、家庭狀況

the Census Bureau reports definitive, public counts

以及地理分佈，

這些都是人口數據基礎設施的基石。

on age, gender, ethnicity,

race, employment, family status,

當少數羣體被少計時，

as well as geographic distribution,

支援公共交通、

which are the foundation of the population data infrastructure.

住房、醫療保健

和保險的 AI 模型，

很可能會忽略最需要這些服務的社群。

When minorities are undercounted,

AI models supporting public transportation,

housing, health care,

要改善結果的第一步，

insurance

是根據人口普查數據，

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

使數據庫在年齡、性別、 民族和種族方面

First step to improving results

具有代表性。

is to make that database representative

由於人口普查如此重要，

我們必須盡力做到百分百的統計。

of age, gender, ethnicity and race

per census data.

數據品質和準確性投放資源

Since census is so important,

we have to make every effort to count 100 percent.

對促成 AI 非常重要，

這並非為了少數人和特權階層，

Investing in this data quality and accuracy

而是為了社會中的每個人。

is essential to making AI possible,

大多數 AI 系統使用已有的

not for only few and privileged,

或為其他目的而收集的數據，

but for everyone in the society.

因為這樣既方便又便宜。

Most AI systems use the data that's already available

然而數據品質是一門需要投入，

or collected for some other purposes

真真正正投入的學問。

because it's convenient and cheap.

這種對偏誤的定義、

Yet data quality is a discipline that requires commitment --

數據收集和測量的關注

在追求速度、規模和方便的世界裏，

real commitment.

This attention to the definition,

不僅沒有得到重視，

data collection and measurement of the bias,

還常常被忽視。

作為尼爾森數據科學團隊一員，

is not only underappreciated --

in the world of speed, scale and convenience,

我曾實地考察

走訪了上海和班加羅爾的 零售店，收集數據。

it's often ignored.

As part of Nielsen data science team,

考察目的是統計這些商店的零售額。

I went to field visits to collect data,

visiting retail stores outside Shanghai and Bangalore.

我們駕車遠離市外，

找到這些小店，

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

非正規、難以接觸到的小店。

你可能想知道，

我們為何會對這些特定商店感興趣？

We drove miles outside the city,

found these small stores --

我們本可以選擇市內的商店，

informal, hard to reach.

將電子數據輕鬆地整合到數據管道中，

And you may be wondering --

why are we interested in these specific stores?

便宜、方便、簡單。

We could have selected a store in the city

為何我們對這些商店的數據品質

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

和準確性如此着迷？

cheap, convenient and easy.

答案很簡單：

因為這些農村商店的數據舉足輕重。

Why are we so obsessed with the quality

and accuracy of the data from these stores?

根據國際勞工組織的資料，

The answer is simple:

because the data from these rural stores matter.

40% 中國人

和 65% 印度人生活在農村地區。

According to the International Labour Organization,

試想一下，

如果印度 65% 的消費力 被排除在模型之外，決䇿上的偏誤，

40 percent Chinese

and 65 percent of Indians live in rural areas.

代表着決策會看重城市多於農村。

Imagine the bias in decision

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

如果沒有這種城鄉背景

以及有關生計、生活方式、 經濟和價值觀的信號，

meaning the decision will favor the urban over the rural.

零售品牌就會在定價、 廣告和營銷方面做出錯誤投資。

Without this rural-urban context

and signals on livelihood, lifestyle, economy and values,

甚或城市偏誤將導致

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

有關農村衛生和 其他資源投放的錯誤決策。

Or the urban bias will lead to wrong rural policy decisions

錯誤決策並非 AI 演算法的問題，

with regards to health and other investments.

而是數據問題，

因它首先排除了想測量的領域。

Wrong decisions are not the problem with the AI algorithm.

It's a problem of the data

要優先考慮的是背景數據，

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

而不是演算法。

我們再來看一個例子。

我走訪了俄勒岡州 這些偏遠的拖車公園住宅

The data in the context is a priority,

not the algorithms.

和紐約市的公寓，

Let's look at another example.

邀請這些家庭參加尼爾森小組。

I visited these remote, trailer park homes in Oregon state

我們邀請這些家庭 在一段時間內參與統計，

and New York City apartments

小組是具有統計代表性的家庭樣本。

to invite these homes to participate in Nielsen panels.

Panels are statistically representative samples of homes

我們的任務是把每人都算進統計，

從這些有收看地面電視習慣的

that we invite to participate in the measurement

over a period of time.

Our mission to include everybody in the measurement

西班牙裔和非洲裔家庭中收集數據。

led us to collect data from these Hispanic and African homes

根據尼爾森的數據，

這些家庭佔美國家庭總數的 15%，

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

約有 4500 萬人。

Per Nielsen data,

these homes constitute 15 percent of US households,

對品質的承諾和關注，

代表我們要盡一切努力，

從這 15% 難以接觸到的羣體中收集信息。

which is about 45 million people.

Commitment and focus on quality means we made every effort

它為什麼要緊？

to collect information

這是一個相當大的羣體，

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

對營銷人員、品牌

和媒體公司都非常重要。

Why does it matter?

如沒有這些數據，

This is a sizeable group

營銷人員、品牌和他們的數據模型

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

就無法接觸到這些人，

as well as the media companies.

也無法向這些非常重要的 少數羣體投放廣告。

Without the data,

the marketers and brands and their models

沒有廣告收入，

would not be able to reach these folks,

Telemundo 或 Univision 等廣播公司

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

就無法提供免費內容，

And without the ad revenue,

包括對我們的民主 至關重要的新聞媒體。

the broadcasters such as Telemundo or Univision,

would not be able to deliver free content,

這些數據對企業和社會都不可或缺。

including news media,

which is so foundational to our democracy.

我們有個千載難逢機會， 去減少 AI 中人類的偏見，

This data is essential for businesses and society.

始於數據。

我的任務不是趕快建立新的演算法，

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

而是建立更好的數據基礎設施，

starts with the data.

使符合道德規範的 AI 成為可能。

Instead of racing to build new algorithms,

我希望你也能加入我的任務。

my mission is to build a better data infrastructure

謝謝。

that makes ethical AI possible.

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

Thank you.