Uncharted Territory


RSS     Archives


AIの問題として個人データの扱いがあります。先ほどのNatureの記事でもPrivacy and consentが懸念事項として挙がっていました。個人データは本人の了承なく収集しないようにすべきだと真っ当な指摘をしていましたが、プライバシーを守りつつ有益な情報を吸い取るFederated learningというGoogleの新しい試みも紹介していました。

Protecting privacy: Federated learning
When technology companies use machine learning to improve their software, they typically gather user information on their servers to analyse how a particular service is being used and then train new algorithms on the aggregated data. Researchers at Google are experimenting with an alternative method of artificial-intelligence training called federated learning. Here, the teaching process happens locally on each user's device without the data being centralized: the lessons aggregated from the data (for instance, the knowledge that the word 'weekly' can be used as an adjective and an adverb) are sent back to Google's servers, but the actual e-mails, texts and so on remain on the user's own phone. Other groups are exploring similar ideas. Thus, information systems with improved designs could be used to enhance users' ownership and privacy over their personal data, while still enabling valuable computations to be performed on those data.

何もプライバシーの問題はAIだけではありませんね。現在の監視テクノロジーのすごさを実感させられたのが今月のNational Geographicのカバーストーリー。自分はFresh Airを聞いていてこの特集を知りました。特集のタイトルはこのようなトピックでおきまりのBig Brother。この記事では監視カメラCCTVの進んだロンドンをメインに、貿易や密猟などで使われる様々な監視カメラの現状を教えてくれ、記事の最後では衛星カメラで地球全域を追いかけられるようになっていることも取り上げています。

Technology and our increasing demand for security have put us all under surveillance. Is privacy becoming just a memory?
 By Robert Draper

In 1949, amid the specter of European authoritarianism, the British novelist George Orwell published his dystopian masterpiece 1984, with its grim admonition: “Big Brother is watching you.” As unsettling as this notion may have been, “watching” was a quaintly circumscribed undertaking back then. That very year, 1949, an American company released the first commercially available CCTV system. Two years later, in 1951, Kodak introduced its Brownie portable movie camera to an awestruck public.

Today more than 2.5 trillion images are shared or stored on the Internet annually—to say nothing of the billions more photographs and videos people keep to themselves. By 2020, one telecommunications company estimates, 6.1 billion people will have phones with picture-taking capabilities. Meanwhile, in a single year an estimated 106 million new surveillance cameras are sold. More than three million ATMs around the planet stare back at their customers. Tens of thousands of cameras known as automatic number plate recognition devices, or ANPRs, hover over roadways—to catch speeding motorists or parking violators but also, in the case of the United Kingdom, to track the comings and goings of suspected criminals. The untallied but growing number of people wearing body cameras now includes not just police but also hospital workers and others who aren’t law enforcement officers. Proliferating as well are personal monitoring devices—dash cams, cyclist helmet cameras to record collisions, doorbells equipped with lenses to catch package thieves—that are fast becoming a part of many a city dweller’s everyday arsenal. Even less quantifiable, but far more vexing, are the billions of images of unsuspecting citizens captured by facial-recognition technology and stored in law enforcement and private-sector databases over which our control is practically nonexistent.


As David Omand, the former director of the Government Communications Headquarters—one of the British intelligence agencies shown by Snowden to be collecting bulk data—put it to me: “On the whole we see our government as efficient and benign. It runs the National Health Service, public education, and social security. And thank God, we haven’t been through the experience of the man in the brown leather trench coat knocking on the door at four in the morning. So when we talk about government surveillance, the resonance is different here.”

That’s not by any means to say that a country like the United States, with its more skeptical view of big government, is wholly immune to surveillance creep. Most of its police departments are now using or considering using body cameras—a development that, thus far at least, has been cheered by civil liberties groups as a means of curbing law enforcement abuses. ANPR cameras are in many major American cities as traffic and parking enforcement tools. In the wake of the September 11 attacks, New York City ramped up its CCTV network and today has roughly 20,000 officially run cameras in Manhattan alone. Meanwhile, Chicago has invested heavily in its network of 32,000 CCTV devices to help combat the murder epidemic in its inner city.


Meanwhile, Planet’s marketing team spends its days gazing at photographs, imagining an interested party somewhere out there. An insurance company wanting to track flood damage to homes in the Midwest. A researcher in Norway seeking evidence of glaciers eroding. But what about … a dictator wishing to hunt down a roving dissident army?

Here is where Planet’s own ethical guidelines would come into play. Not only could it refuse to work with a client having malevolent motives, but it also doesn’t allow customers to stake a sole proprietary claim over the images they buy. The other significant constraint is technological. Planet’s surveillance of the world at a resolution of 10 feet is sufficient to discern the grainy outline of a single truck but not the contours of a human. Resolution-wise, the current state of the art of one foot is supplied by another satellite imaging company, DigitalGlobe. But for now, only Planet, with its formidable satellite deployment, is capable of providing daily imagery of Earth’s entire landmass. “We’ve run the proverbial four-minute mile,” Marshall said. “Simply knowing it’s possible doesn’t make it any easier.”


I was pondering the implications of this when a young woman showed me what was on her laptop. Her name was Annie Neligh, an Air Force veteran who now leads “customer solutions engineering” at Planet. One of Neligh’s customers needing a solution was a Texas-based insurance company. The company suspected that it was renewing insurance policies for homeowners who weren’t disclosing that they’d installed swimming pools—a 40 percent loss on each policy for the company. So it had asked Planet to provide satellite imagery of homes in Plano, Texas.

Neligh showed me what she’d found. Looking at a neighborhood of 1,500 properties, we could clearly see the shimmering shapes of 520 small bodies of water—a proportion far in excess of what the insurance company’s customers had claimed. Neligh shrugged and offered a thin smile. “People lie, you know,” she said.

Now her client had the truth. What would it do with this information? Conduct a surprise raid on the somnolent hamlets of Plano? Jack up premiums? Order images that might show construction crews installing new Jacuzzis and Spanish tile roofs? The future is here, and in it, truth is more than a kindly educator. It is a weapon—against timber poachers and burglars and mad bombers and acts of God, but also against the lesser angels of our nature. People lie, you know. The age of transparency is upon us.

英語表現的に面白いと思ったのは最後にあるlesser angels of our natureというもの。ピンカー教授はbetter angels of our natureという本を出していましたが、こちらの記事では人間の悪い面を表現しています。




Brain-computer interfaceは年初のEconomistのTechnology Quarterlyでも取り上げられましたが、昨年のNatureの記事でBrain-computer interfaceについてPrivacy and consent, Agency and identity, Augmentation, Biasという4つの懸念事項をあげています。この懸念はAIにも当てはまるそうです。

Rafael Yuste, Sara Goering, Blaise Agüera y Arcas, Guoqiang Bi, Jose M. Carmena, Adrian Carter, Joseph J. Fins, Phoebe Friesen, Jack Gallant, Jane E. Huggins, Judy Illes, Philipp Kellmeyer, Eran Klein, Adam Marblestone, Christine Mitchell, Erik Parens, Michelle Pham, Alan Rubel, Norihiro Sadato, Laura Specker Sullivan, Mina Teicher, David Wasserman, Anna Wexler, Meredith Whittaker& Jonathan Wolpaw
08 November 2017
Artificial intelligence and brain–computer interfaces must respect and preserve people's privacy, identity, agency and equality, say Rafael Yuste, Sara Goering and colleagues.


Bias. When scientific or technological decisions are based on a narrow set of systemic, structural or social concepts and norms, the resulting technology can privilege certain groups and harm others. A 2015 study12 found that postings for jobs displayed to female users by Google's advertising algorithm pay less well than those displayed to men. Similarly, a ProPublica investigation revealed last year that algorithms used by US law-enforcement agencies wrongly predict that black defendants are more likely to reoffend than white defendants with a similar criminal record (go.nature.com/29aznyw). Such biases could become embedded in neural devices. Indeed, researchers who have examined these kinds of cases have shown that defining fairness in a mathematically rigorous manner is very difficult (go.nature.com/2ztfjt9).

Practical steps to counter bias within technologies are already being discussed in industry and academia. Such ongoing public discussions and debate are necessary to shape definitions of problematic biases and, more generally, of normality.

We advocate that countermeasures to combat bias become the norm for machine learning. We also recommend that probable user groups (especially those who are already marginalized) have input into the design of algorithms and devices as another way to ensure that biases are addressed from the first stages of technology development.


by Julia Angwin, Jeff Larson, Surya Mattu and Lauren Kirchner, ProPublica
May 23, 2016







Research on collective recall takes on new importance in a post-fact world.
Laura Spinney
07 March 2017 Corrected: 08 March 2017

Strange things have been happening in the news lately. Already this year, members of US President Donald Trump's administration have alluded to a 'Bowling Green massacre' and terror attacks in Sweden and Atlanta, Georgia, that never happened.

The misinformation was swiftly corrected, but some historical myths have proved difficult to erase. Since at least 2010, for example, an online community has shared the apparently unshakeable recollection of Nelson Mandela dying in prison in the 1980s, despite the fact that he lived until 2013, leaving prison in 1990 and going on to serve as South Africa's first black president.

Memory is notoriously fallible, but some experts worry that a new phenomenon is emerging. “Memories are shared among groups in novel ways through sites such as Facebook and Instagram, blurring the line between individual and collective memories,” says psychologist Daniel Schacter, who studies memory at Harvard University in Cambridge, Massachusetts. “The development of Internet-based misinformation, such as recently well-publicized fake news sites, has the potential to distort individual and collective memories in disturbing ways.”

実はTOEICにも「銀ブラ問題」に近いものがあります。L&R 1のPart2にある以下のやりとりです。

Can we try that Brazilian café tonight?
- I went there last week.

和訳ではブラジル料理レストランとして訳していますが、Brazil coffeeの喫茶店の可能性はないでしょうか。cafezinhoは冒頭の動画でも出ていましたね。

Starbucks Brazil coffee master Vivi Fonseca explains the long-standing Brazilian coffee tradition of cafezinho. It means "little coffee" in Portuguese, but it's actually a moment to connect, chat and relax over coffee.










English Journalの今月号の特集『\たった150語!/TOEICスコアが劇的にアップする 「コスパ最高」の英単語』のコラムで「ランキング上位から見るTOEICの世界」が紹介されていて、TOEICらしい語としてmailが53位になっていました。


(1) mailにはe-mailも集計されている可能性がある

(2) Questions 181-185 refer to the following e-mails.の部分も集計されている可能性がある。

(1) mailにはe-mailも集計されている可能性がある



3 《U》 (E )メール
• I just want to check my mail.
• You have mail.
同意 email
4 《C》 (E )メール(のメッセージ)
• I got a mail from him this morning.
同意 email


e-mail address(メールアドレス)、 mailing address(郵送先住所)などのように、ほぼ「住所」の意味で使われている。


Please take a moment to fill out the following survey and mail it to us in the enclosed self-addressed, stamped envelope by May 28.

 we can send you a print version in the mail if you prefer



(2) Questions 181-185 refer to the following e-mails.の部分も集計されている可能性がある。

e-mailの使用頻度が高いといっても、少なくともYutaのテキストデータでは、Questions 181-185 refer to the following e-mails.の部分で使われているからなんですよね。テストの問題指示文の扱いは集計には厄介なところです。例えば今度出る銀フレでは設問に出る単語・表現を別に扱っていますね。この当たりもいい悪いではなく著者の方針に委ねられるところです。




English JournalはいつもKindle Unlimitedで読んでいるんですが、3月号の特集が楽しみで今回は単品購入しました。

【特集】\たった150語!/TOEICスコアが劇的にアップする 「コスパ最高」の英単語
今や年間に200万人以上が受験する英語試験の定番、TOEIC L&Rテスト。



それではなぜTOEICスコアが停滞しがちなのか? その原因は特集での西嶋愉一先生のコメントにあるのではないでしょうか。