Data can be scarcer than you think, and full of traps
数据可能比你想象的要稀缺，而且充满陷阱
AMAZON’S “GO” STORES are impressive places. The cashier-less shops, which first opened in Seattle in 2018, allow app-wielding customers to pick up items and simply walk out with them. The system uses many sensors, but the bulk of the magic is performed by cameras connected to an AI system that tracks items as they are taken from shelves. Once the shoppers leave with their goods, the bill is calculated and they are automatically charged.
亚马逊的“GO”商店令人眼前一亮。这些不设收银员的店铺2018年首次在西雅图开业，顾客只要亮出手机应用，就可以拿了商品直接走人。该系统使用大量传感器，但其魔法主要是由连接到AI系统的摄像头完成的。AI系统会追踪商品从架子上被取走的过程。一旦顾客拿着商品离店，账单就结算完毕，自动向他们收费。
Doing that in a crowded shop is not easy. The system must handle crowded stores, in which people disappear from view behind other customers. It must recognise individual customers as well as friends or family groups (if a child puts an item into a family basket, the system must realise that it should charge the parents). And it must do all that in real-time, and to a high degree of accuracy.
在一个拥挤的商店里做到这一点并不容易。系统要能够应付人员密集的环境：摄像头可能被其他顾客阻挡而看不到某些人的动作。它必须能识别单个顾客，还有同行的朋友或是全家出动。如果一个孩子把一件商品放进自家购物篮，系统必须意识到应该向他的父母收费。而且它必须实时又高度准确地完成这一切。
Teaching the machines required showing them a lot of “training data” in the form of videos of customers browsing shelves, picking up items, putting them back and the like. For standardised tasks like image recognition, AI developers can use public training datasets, each containing thousands of pictures. But there was no such training set featuring people browsing in shops.
为指导机器做这些，需要向它们展示大量“训练数据”：顾客浏览货架上的商品、拿取商品、把商品放回货架等各种行为的视频。对于像图像识别这样的标准化任务，AI开发人员可以使用公用训练数据集，每个都包含成千上万张图片。但记录人们逛商店的公用训练集尚不存在。
Some data could be generated by Amazon’s own staff, who were allowed in to test versions of the shops. But that approach took the firm only so far. There are many ways in which a human might take a product from a shelf and then decide to choose it, put it back immediately or return it later. To work in the real world, the system would have to cover as many of those as possible.
有些数据可由亚马逊自己的员工生成，公司此前让他们进入测试版店铺中。但这么做有其局限。人们有各种各样的方式从架子上取走一件商品并决定买下它、立即把它放回架子，还是稍后再放回。要在现实世界中真正奏效，系统必须涵盖尽可能多的可能性。
In theory, the world is awash with data, the lifeblood of modern AI. IDC, a market-research firm, reckons the world generated 33 zettabytes of data in 2018, enough to fill seven trillion DVDs. But Kathleen Walch of Cognilytica, an AI-focused consultancy, says that, nevertheless, data issues are one of the most common sticking-points in any AI project. As in Amazon’s case, the required data may not exist at all. Or they might be locked up in the vaults of a competitor. Even when relevant data can be dug up, they might not be suitable for feeding to computers.
从理论上讲，世界充斥着数据，这是现代AI的命脉。市场研究公司国际数据公司（IDC）估计，2018年全球生成了33ZB的数据，足以填满七万亿张DVD。但是，专注于AI领域的咨询公司Cognilytica的凯瑟琳·沃尔克（Kathleen Walch）表示，尽管如此，数据问题仍是所有AI项目中最常见的症结之一。和亚马逊Go商店的例子一样，所需要的数据可能根本就不存在。或者数据可能被锁在竞争对手的保险库中。即便相关数据可以被挖出，可能也不适合输送给计算机。
Data-wrangling of various sorts takes up about 80% of the time consumed in a typical AI project, says Cognilytica. Training a machine-learning system requires large numbers of carefully labelled examples, and those labels usually have to be applied by humans. Big tech firms often do the work internally. Companies that lack the required resources or expertise can take advantage of a growing outsourcing industry to do it for them. A Chinese firm called MBH, for instance, employs more than 300,000 people to label endless pictures of faces, street scenes or medical scans so that they can be processed by machines. Mechanical Turk, another subdivision of Amazon, connects firms with an army of casual human workers who are paid a piece rate to perform repetitive tasks.
Cognilytica表示，一个典型AI项目约80%的时间都花在了各种数据整理上。训练机器学习系统需要大量仔细标注的样本，而这些标注通常须由人类添加。大型科技公司通常在内部开展这项工作。那些缺少相关资源或技术知识的公司可以借力一个不断发展的外包产业来完成这个部分。例如，中国公司莫比嗨客雇用了30多万人来标注源源不断的人脸照片、街道场景或医疗扫描影像以便由机器处理。亚马逊的另一个部门土耳其机器人（Mechanical Turk）为企业与一个临时工大军牵线搭桥，向这些工人支付计件工资来执行重复性任务。
Cognilytica reckons that the third-party “data preparation” market was worth more than $1.5bn in 2019 and could grow to $3.5bn by 2024. The data-labelling business is similar, with firms spending at least $1.7bn in 2019, a number that could reach $4.1bn by 2024. Mastery of a topic is not necessary, says Ron Schmelzer, also of Cognilytica. In medical diagnostics, for instance, amateur data-labellers can be trained to become almost as good as doctors at recognising things like fractures and tumours. But some amount of what AI researchers call “domain expertise” is vital.
Cognilytica估计，第三方“数据准备”市场在2019年价值超过15亿美元，到2024年可能增至35亿美元。数据标注业务也差不多：2019年企业在这方面至少支出了17亿美元，到2024年可能达到41亿美元。Cognilytica的罗恩·施梅尔策（Ron Schmelzer）说，掌握某个专业课题并非必要，例如在医学诊断中，业余数据标注员经训练后在识别骨折和肿瘤等方面几乎可以和医生媲美。但掌握一定的AI研究人员口中的“领域知识”至关重要。
The data themselves can contain traps. Machine-learning systems correlate inputs with outputs, but they do it blindly, with no understanding of broader context. In 1968 Donald Knuth, a programming guru, warned that computers “do exactly what they are told, no more and no less”. Machine learning is full of examples of Mr Knuth’s dictum, in which machines have followed the letter of the law precisely, while being oblivious to its spirit.
数据本身可能包含陷阱。机器学习系统将输入与输出相关联，但它们只是盲目地执行，并不理解更广泛的语境。1968年，编程大师高德纳（Donald Knuth）警告说，计算机“完全按你告诉它们的去做，不多也不少”。机器学习中充满了这句话的例证——机器精确遵循规则的字眼，却对其精神一无所知。
In 2018 researchers at Mount Sinai, a hospital network in New York, found that an AI system trained to spot pneumonia on chest x-rays became markedly less competent when used in hospitals other than those it had been trained in. The researchers discovered that the machine had been able to work out which hospital a scan had come from. (One way was to analyse small metal tokens placed in the corner of scans, which differ between hospitals.)
2018年，纽约西奈山医疗系统（Mount Sinai）的研究人员发现，一个经训练从X光胸片识别肺炎的AI系统，在它受训的医院以外的其他医院使用时能力明显降低。研究人员发现，机器能够识别出胸片来自哪家医院。（方法之一是分析片子角上的小块金属标记——各家医院的标记各不相同。）
Since one hospital in its training set had a baseline rate of pneumonia far higher than the others, that information by itself was enough to boost the system’s accuracy substantially. The researchers dubbed that clever wheeze “cheating”, on the grounds that it failed when the system was presented with data from hospitals it did not know.
由于训练集里的一家医院的肺炎基准发生率远高于其他医院，胸片来自哪家医院这个信息本身就足以大幅提高系统的准确性。研究人员把这种巧妙的伎俩叫做“作弊”，因为当向系统出示陌生医院的数据时，它就失灵了。
Different kind of race
另一个族群
Bias is another source of problems. Last year America’s National Institute of Standards and Technology tested nearly 200 facial recognition algorithms and found that many were significantly less accurate at identifying black faces than white ones. The problem may reflect a preponderance of white faces in their training data. A study from IBM, published last year, found that over 80% of faces in three widely used training sets had light skin.
偏见导致了另一种问题。去年，美国国家标准技术研究院（National Institute of Standards and Technology）测试了近200种人脸识别算法，发现许多算法在识别黑人面部时准确性明显低于白人面部。这个问题可能反映出白人面部在机器的训练数据中占了多数。IBM去年发表的一项研究发现，三种被广泛使用的训练集中，超过80%的人脸都是较浅的肤色。
Such deficiencies are, at least in theory, straightforward to fix (IBM offered a more representative dataset for anyone to use). Other sources of bias can be trickier to remove. In 2017 Amazon abandoned a recruitment project designed to hunt through CVs to identify suitable candidates when the system was found to be favouring male applicants. The post mortem revealed a circular, self-reinforcing problem. The system had been trained on the CVs of previous successful applicants to the firm. But since the tech workforce is already mostly male, a system trained on historical data will latch onto maleness as a strong predictor of suitability.
至少从理论上讲，这类缺陷很容易纠正（IBM提供了一个更具代表性的数据集供所有人使用）。其他的偏见来源可能更难消除。2017年，亚马逊喊停了一个通过简历寻找合适人选的招聘项目，因为发现该系统对男性申请人有利。事后检验发现了一个循环的、自我增强的问题。公司用以前成功被录取的申请人的简历训练该系统，但由于技术人员的队伍本身大部分是男性，因此根据历史数据来训练的系统会把男性这个特征作为适合度的强预测指标。
Humans can try to forbid such inferences, says Fabrice Ciais, who runs PwC’s machine-learning team in Britain (and Amazon tried to do exactly that). In many cases they are required to: in most rich countries employers cannot hire on the basis of factors such as sex, age or race. But algorithms can outsmart their human masters by using proxy variables to reconstruct the forbidden information, says Mr Ciais. Everything from hobbies to previous jobs to area codes in telephone numbers could contain hints that an applicant is likely to be female, or young, or from an ethnic minority.
普华永道机器学习英国团队的负责人法布里斯·西亚斯（Fabrice Ciais）说，人类可以尝试禁止机器做这类推导（亚马逊正是这么做的）。在许多情况下他们必须这么做：在大多数富裕国家，雇主不能基于性别、年龄或种族等因素来雇用人员。但算法可以比它的人类主人更聪明，西亚斯说，它们能用替代变量重构出被禁用的信息。从业余爱好到工作经历，再到电话号码中的区号，各种信息都可能暗示申请者很可能是女性、年轻人或少数族裔。
If the difficulties of real-world data are too daunting, one option is to make up some data of your own. That is what Amazon did to fine-tune its Go shops. The company used graphics software to create virtual shoppers. Those ersatz humans were used to train the machines on many hard or unusual situations that had not arisen in the real training data, but might when the system was deployed in the real world.
如果现实世界中的数据难题太过艰巨，那么一种选择是自己创造一些数据。这就是亚马逊改进Go商店时所用的方法。该公司使用图形软件来生成虚拟购物者。这些“人造人”被拿来训练机器处理许多困难或异常的情景，它们在真实训练数据中未曾出现，在实际环境中部署系统时却可能发生。
Amazon is not alone. Self-driving car firms do a lot of training in high-fidelity simulations of reality, where no real damage can be done when something goes wrong. A paper in 2018 from Nvidia, a chipmaker, described a method for quickly creating synthetic training data for self-driving cars, and concluded that the resulting algorithms worked better than those trained on real data alone.
这并非亚马逊独树一帜。无人车公司用高保真模拟现实来做大量训练，在这种模拟中如果出错不会造成真正的破坏。芯片制造商英伟达2018年发表的一篇论文描述了一种为无人车快速创建综合训练数据的方法，并得出结论称由此生成的算法比仅用真实数据训练的算法的效果更好。
Privacy is another attraction of synthetic data. Firms hoping to use AI in medicine or finance must contend with laws such as America’s Health Insurance Portability and Accountability Act, or the European Union’s General Data Protection Regulation. Properly anonymising data can be difficult, a problem that systems trained on made-up people do not need to bother about.隐私关切是“合成数据”的另一个吸引力所在。希望在医学或金融中使用AI的公司必须遵守美国的《健康保险可携性和责任法案》（HIPAA）或欧盟的《通用数据保护条例》（GDPR）等法律。要给数据做恰当的匿名处理可能会很难，而用虚拟人训练的系统根本不用担心这个。
The trick, says Euan Cameron, one of Mr Ciais’s colleagues, is ensuring simulations are close enough to reality that their lessons carry over. For some well-bounded problems such as fraud detection or credit scoring, that is straightforward. Synthetic data can be created by adding statistical noise to the real kind. Although individual transactions are therefore fictitious, it is possible to guarantee that they will have, collectively, the same statistical characteristics as the real data from which they were derived. But the more complicated a problem becomes, the harder it is to ensure that lessons from virtual data will translate smoothly to the real world.
西亚斯的同事尤安·卡梅伦（Euan Cameron）说，诀窍在于确保模拟足够接近现实，以使得经验可以推广。对于像欺诈识别或信用评分这样清晰界定的问题，这很简单。还可以将统计噪声添加到真实数据中来创建合成数据。这样，尽管单个交易是虚拟的，但可以保证它们整体上具有与源数据相同的统计特征。但一个问题越复杂，就越难确保从虚拟数据中汲取的经验能被顺畅地用于现实世界。
The hope is that all this data-related faff will be a one-off, and that, once trained, a machine-learning model will repay the effort over millions of automated decisions. Amazon has opened 26 Go stores, and has offered to license the technology to other retailers. But even here there are reasons for caution. Many AI models are subject to “drift”, in which changes in how the world works mean their decisions become less accurate over time, says Svetlana Sicular of Gartner, a research firm. Customer behaviour changes, language evolves, regulators change what companies can do.
希望在于所有这些与数据相关的折腾都是一次性的，而一旦训练好，机器学习模型将用数百万次自动决策来回报这番努力。亚马逊已经开设了26家Go商店，并已提出将相关技术授权给其他零售商。但即使到了这一步也仍需要谨慎。研究公司高德纳（Gartner）的斯韦特兰娜·希克尔勒（Svetlana Sicular）说，许多AI模型都受到“漂移”（drift）的影响，即随着时间流逝，世界运转方式的变化意味着它们的决策变得不那么准确。顾客的行为在变化，语言在演变，监管机构也会改变公司能做什么的规定。
Sometimes, drift happens overnight. “Buying one-way airline tickets was a good predictor of fraud ,” says Ms Sicular. “And then with the covid-19 lockdowns, suddenly lots of innocent people were doing it.” Some facial-recognition systems, used to seeing uncovered human faces, are struggling now that masks have become the norm. Automated logistics systems have needed help from humans to deal with the sudden demand for toilet roll, flour and other staples. The world’s changeability means more training, which means providing the machines with yet more data, in a never-ending cycle of re-training. “AI is not an install-and-forget system,” warns Mr Cameron.
有时漂移在一夜之间发生。“购买单程机票［在自动检测模型中］曾是一个很好的预测欺诈的指标，”希克尔勒说，“新冠肺炎导致封城后，突然有很多人都在买单程票，却都是清白的。”如今戴口罩已成为常态，一些习惯了识别裸露面部的人脸识别系统碰到了麻烦。自动化物流系统现在需要人员的帮助才能应付对卷筒纸、面粉及其他生活必需品的需求激增。世界的可变性意味着机器需要更多训练，即为它们提供更多数据——这是一个无休止的再培训循环。卡梅伦警告说：“人工智能不是个一劳永逸的系统。”