通用情报 (General Intelligence)

The agency that runs the United States’ airport and border facial recognition program has failed to properly tell the public about how it works, a new report has found. In whole, the report reads like a major red flag: The U.S. government is charging ahead with the adoption of this questionable technology, and it’s not informing the public or keeping proper tabs on accuracy.

运行美国的机场和边境面部识别节目T他机构未能正确地告诉它是如何工作的公开，一份新的报告发现。 总体而言，该报告读起来像是一个重大危险信号：美国政府正在推动采用这一可疑技术，而不是向公众进行通报或保持准确性。

By law, the Customs and Border Patrol is supposed to inform the public when facial recognition is being used by putting up clear, legible signs telling people that their faces are being scanned and how they can opt out. The department is also supposed to put accurate, up-to-date information online about its facial recognition, and provide information through its call center.

根据法律规定，海关和边境Patro L是应该通知时，正在使用的搭建明确，清晰的迹象告诉人们，他们的脸被扫描面部识别的公众，以及他们如何可以选择退出。 该部门还应该在线发布有关其面部识别的准确，最新信息，并通过其呼叫中心提供信息。

Signs disclosing the use of facial recognition were hidden behind bigger signs at airports, and some contained outdated information.

公开使用面部识别的标志被隐藏在机场较大的标志后面，其中一些包含过时的信息。

But a new report from the federal Government Accountability Office (GAO) found CBP lacking in all these regards. Signs disclosing the use of facial recognition were hidden behind bigger signs at airports, and some contained outdated information. Some signs didn’t tell people how to opt out, or what would happen if they asked to opt out.

但是联邦政府问责局(GAO)的一份新报告发现CBP在所有这些方面都缺乏。 公开使用面部识别的标志被隐藏在机场较大的标志后面，其中一些包含过时的信息。 有些迹象没有告诉人们如何选择退出，或者如果他们要求退出，会发生什么。

Not having these signs means that people were less likely to opt out or question whether they needed to submit to a facial scan. Privacy organizations told GAO that CBP discourages opting out, with the justification that doing so “would lead to additional security scrutiny, increased wait times, and could be grounds to deny boarding.” GOA found that CBP officers weren’t present to address opt-out requests at airports, meaning that travelers who opted out would have to wait while additional staff was called in to address concerns.

没有这些迹象意味着人们不太可能选择退出或质疑他们是否需要接受面部扫描。 隐私权组织告诉GAO，CBP不鼓励退出，理由是这样做“将导致额外的安全审查，增加等待时间，并可能成为拒绝登机的理由。” GOA发现，CBP官员没有在场解决机场的退出请求，这意味着选择退出的旅行者将不得不等待额外的工作人员来解决问题。

Investigators also couldn’t always reach the CBP call center to ask about opt-out information. When they did reach staff at the call center, the operators didn’t know information about the facial recognition programs. (Welcome to the club, GAO.)

调查人员也不能总是到达CBP呼叫中心询问有关退出信息。 当他们到达呼叫中心的工作人员时，操作员不知道有关面部识别程序的信息。 (欢迎来到GAO俱乐部。)

The CBP, in turn, blamed airports, airlines, and cruise operators for failing to communicate with passengers. The CBP relies on private companies to post privacy signs. But the CBP has done little to make sure these signs are present: The department only checked on one out of 20 places where the facial recognition tests were taking place. CBP has no plans to make sure such signs are present at other locations.

反过来，CBP则指责机场，航空公司和邮轮运营商未能与旅客沟通。 CBP依靠私人公司张贴隐私标志。 但是CBP几乎没有做任何事情来确保这些迹象的存在：该部门只检查了进行面部识别测试的20个地点中的一个。 美国海关边防总署没有计划确保在其他地方有这种迹象。

The report also contained some interesting statistics about the effectiveness of the facial recognition programs in catching individuals traveling under false identities. In airports, CBP has scanned more than 16 million passengers arriving in the United States up to May 2020, and stopped a total of seven imposters. At the southern border, facial recognition was also used to scan 4.4 million pedestrians crossing into the United States between September 2018 and December 2019, and stopped 215 imposters.

该报告还包含一些有趣的统计数据，说明面部识别程序在捕获以虚假身份旅行的个人方面的有效性。 到2020年5月，美国海关边防总署已在机场对超过1600万名到达美国的旅客进行了扫描，并制止了7名冒名顶替者。 在南部边界，面部识别技术还被用于扫描2018年9月至2019年12月之间进入美国的440万行人，并制止了215名冒名顶替者。

The CBP is also using facial recognition for commercial vehicles in Brownsville, Texas, and at the northern border at Peace Bridge in Buffalo, New York. However, no imposters were reported to be caught at those locations.

CBP还在德克萨斯州布朗斯维尔以及纽约州布法罗市和平桥北部边界的商用车上使用面部识别。 但是，据报在这些地方没有冒名顶替者。

The CBP highlighted one success story from the southern border to GOA: The agency used facial recognition to catch a pedestrian trying to enter the United States wearing Halloween makeup.

CBP强调了从南部边界到GOA的一个成功案例：该机构使用面部识别技术来捕捉一名试图穿着万圣节化装进入美国的行人。

Perhaps the most glaring oversight that GAO noted, however, was that CBP doesn’t regularly monitor the accuracy of its facial recognition. The border agency looks at data from two flights per airport per week to make sure the images are good quality and accuracy is above the acceptable threshold of 90%. But according to the report, this process doesn’t alert CBP to ongoing daily issues until days or weeks after they start to occur.

但是，GAO指出的最明显的疏忽可能是CBP没有定期监视其面部识别的准确性。 边防局每周检查每个机场两次航班的数据，以确保图像质量良好且准确性高于可接受的90％阈值。 但是根据该报告，此过程直到CBP开始发生后的几天或几周才提醒CBP日常的日常问题。

The GAO report reflects the confusing, uneven use of facial recognition technologies at border crossings today. While the number of cameras pointing at our faces continues to grow, we know less and less about why they’re scanning us or what databases they’re matching us against.

GAO报告反映了当今边境口岸人脸识别技术的使用混乱，不平衡。 尽管面对我们的相机数量不断增长，但我们对他们为什么扫描我们或与我们匹配的数据库了解的越来越少。

And now, here’s some of the most interesting A.I. research of the week:

现在，这是本周最有趣的AI研究：

Facebook探索3D照片 (Facebook explores 3D photos)

New research from Facebook looks at how to convert a 2D photo to 3D, by using an algorithm that guesses what the scene would look like from all angles and then fills in the gaps. An example picture is shown of a person standing in the snow. The algorithm uses the context of the snow around them to fill in the gaps, and creates the ability to view the photo from an entirely new perspective.

Facebook的一项新研究着眼于如何通过使用一种算法从各个角度猜测场景的外观然后填补空白的算法，将2D照片转换为3D。 示例图片显示了一个站在雪地上的人。 该算法利用周围雪域的背景来填补空白，并具有从全新的角度查看照片的能力。

介绍最糟糕的恶梦 (Introducing your worst Flightmare)

A team from Zurich University has released a new quadcopter simulator called Flightmare, which allows researchers to insert hundreds of digital quadcopters and have them learn to fly. The simulator’s main strength is its flexibility — researchers can have their algorithms make decisions up to 20,000 times per second, meaning they can learn to fly by making incredibly granular adjustments.

苏黎世大学的一个团队发布了一种名为Flightmare的新型四轴飞行器模拟器，该模拟器可使研究人员插入数百个数字四轴飞行器，并使他们学会飞行。 模拟器的主要优势在于灵活性-研究人员可以使他们的算法每秒做出20,000次决策，这意味着他们可以通过进行难以置信的细微调整来学习飞行。

关于新型AI助手的建议 (A proposal for a new kind of A.I. assistant)

Google Assistant and Siri are cool, but a paper from the University of Alberta suggests that a better way to make intelligence assistants would be to train them in creating and editing text. They argue that going beyond mere transcription of dictation is a small enough jump that algorithms would still be able to understand the domain, and that commands like “remove that apostrophe” are clear but require understanding of simple context that the algorithm can see.

Google助手和Siri很酷，但是阿尔伯塔大学的一篇论文建议，使智能助手更好的方法是训练他们创建和编辑文本。 他们认为，仅听写命令的转录就足够小了，算法仍然可以理解该域，并且“删除撇号”之类的命令很清楚，但需要了解算法可以看到的简单上下文。

适用于动漫书呆子的更多AI (More A.I. for anime nerds)

There is so much A.I. research that centers around anime and manga, I can only conclude that every A.I. researcher is a massive nerd. This paper is for the cosplayers, though. Research from Doshisha University and Zozo Technologies in Japan attempts to generate images of real clothing from screenshots of anime. They say this would allow realistic costumes to be automatically generated, making them easier to recreate in real life.

关于动漫和漫画的AI研究太多了，我只能得出结论，每个AI研究者都是一个书呆子。 不过，本文适用于扮演者。 日本同志社大学和日本Zozo Technologies的研究试图从动漫的屏幕截图生成真实服装的图像 。 他们说，这将允许自动生成逼真的服装，使它们在现实生活中更容易重现。