算法偏见是什么_算法有何偏见
算法偏见是什么
After the end of the Second World War, the Nuremberg trials laid bare the atrocities conducted in medical rearch by the Nazis. In the aftermath of the trials, the medical sciences established a t of rules — The Nuremberg Code — to control future experiments involving human subjects. The Nuremberg Code has influenced medical codes of ethics around the world, as has the exposure of experiments that had failed to follow it even three decades later, such as the infamous .
第⼆次世界⼤战结束后，纽伦堡的审判揭露了纳粹在医学研究中进⾏的暴⾏。 在试验之后，医学界建⽴了⼀套规则-《纽伦堡守则》，以控制未来涉及⼈类受试者的实验。 纽伦堡守则(Nuremberg Code)影响了世界各地的医学道德守则，甚⾄在三⼗年后仍未遵循该守则的实验的曝光，例如臭名昭著的 。
The direct negative impact of AI experiments and applications on urs isn’t quite as inhumane as that of the Tuskegee and Nazi experimentations, but in the face of an overwhelming and growing body of evidence of algorithms being biad against certain demographic cohorts, it is important that a dialogue takes place sooner or later. AI systems can be biad bad on who builds them, the way they are developed, and how they’re eventually deployed. This is known as algorithmic bias.
AI实验和应⽤程序对⽤户的直接负⾯影响并不像Tuskegee和Nazi实验那样不⼈道，但是⾯对越来越多的越来越多的证据表明算法偏向某些⼈⼝统计群体，这⼀点很重要对话迟早会发⽣。 ⼈⼯智能系统可能会根据谁构建它们，它们的开发⽅式以及最终部署⽅式⽽产⽣偏差。 这称为算法偏差。
While the data sciences have not developed a Nuremberg Code of their own yet, the social implications of rearch in artificial intelligence are starting to be addresd in some curricula. But even as the debates are starting to sprout up, what is still lacking is a discipline-wide discussion to grapple with questions of how to tackle societal and historical inequities that are reinforced by AI algorithms.
尽管数据科学尚未制定⾃⼰的《纽伦堡守则》，但在某些课程中已开始研究⼈⼯智能研究的社会意义。 但是，即使辩论开始兴起，仍然缺乏针对整个学科的讨论，以解决如何解决由AI算法强化的社会和历史不平等问题。飞向蓝天
We are flawed creatures. Every single decision we make involves a certain kind of bias. However, algorithms haven’t proven to be much better. Ideally, we would want our algorithms to make better-informed decisions devoid of bias so as to ensure better social justice, i.e., equal opportunities for individuals and groups (such as minorities) within society to access resources, have their voices heard, and be reprented in society.
武汉社保局我们是有缺陷的⽣物。 我们做出的每个决定都带有某种偏见。 但是，尚未证明算法会更好。 理想情况下，我们希望我们的算法做出⽆偏见的明智决策，以确保更好的社会公正性，即社会中的个⼈和群体(例如少数民族)获得资源，听到⾃⼰的声⾳并保持沉默的平等机会。代表社会。
When the algorithms do the job of amplifying racial, social and gender inequality, instead of alleviating it; it becomes necessary to take stock of the ethical ramifications and potential malevolence of the technology.
当这些算法完成放⼤种族，社会和性别不平等的⼯作，⽽不是减轻它时； 有必要盘点该技术的道德后果和潜在的恶意。
This essay was motivated by two flashpoints : the racial inequality discussion that is now raging on worldwide, and Yann LeCun’s altercation with Timnit Gebru on Twitter which was caud due to a disagreement over a downsampled image of Barack Obama (left) that was depixelated to a picture of a white man (right) by a face upsampling machine learning (ML) model.
技术经济与管理本⽂的出发点有两个：种族不平等的讨论现在在全球范围内进⾏，以及Yann LeCun与Twitter上的Timnit Gebru发⽣争执，这是由于对Barack Obama(左)的降采样图像的不同意见所致，该图像被去像素化为像素。脸部向上采样机器学习(ML)模型拍摄的⽩⼈(右)图⽚。
The (rather explosive) argument was sparked by this tweet by LeCun where he says that the resulting face was that of a white man becau of a bias in data that trained the algorithm. Gebru responded sharply that the harms of ML systems cannot be reduced to biad data.
LeCun的这条推⽂引发了(颇具爆炸性的)争论，他说，由于训练该算法的数据存在偏差，最终的⾯Kong是⽩⼈。 Gebru敏锐地回答说，机器学习系统的危害⽆法减少到有偏见的数据上。
In most baline ML algorithms, the model fits better to the attributes that that occur most frequently across various data points. For example, if you were to design an AI recruiting tool to review the résumés of applicants for a software engineering position, you would first need to train it with a datat of past candidates which contains details like “experience”, “qualifications”, “degree(s) held”, “past projects” etc. For every datapoint, the algorithm of the hiring tool would need a decision or a “label”, so as to “learn” how to make a decision for a given applicant by obrving patterns in their résumé.
在⼤多数基线ML算法中，模型更适合各种数据点上最频繁出现的属性。 例如，如果您要设计⼀个AI招聘⼯具来审查软件⼯程职位申请⼈的履历，则⾸先需要使⽤过去候选⼈的数据集对其进⾏培训，其中包含“经验”，“资格”，“对于每个数据点，招聘⼯具的算法将需要⼀个决定或⼀个“标签”，以便通过观察来“学习”如何为给定的申请⼈做出决定简历中的样式。
For an industry where the gender disparity in reprentation is large, it is reasonable to assume that a large majority of the data points will be male applicants. And this collective imbalance in the data ends up being interpreted by the algorithm as a uful pattern in the data rather than undesirable noi which is to be ignored. Conquently, it teaches itlf that male candidates are more preferable than female candidates.
对于代表性别上的巨⼤差异的⾏业，可以合理假设⼤多数数据点是男性申请⼈。 并且，数据中的这种集体失衡最终被算法解释为数据中的有⽤模式，⽽不是被忽略的不希望有的噪声。 因此，它⾃称男性候选⼈⽐⼥性候选⼈更可取。
I wish that this was merely an imaginary, exaggerated example that I ud to prove my point.
我希望这只是我⽤来证明我观点的⼀个虚构的，夸张的例⼦。
LeCun wasn’t wrong in his asssment becau in the ca of that specific model, training the model on a datat that contains faces of black people (as oppod to one that contains mainly white faces) would not have given ri to an output as absurd as that. But the upside of the godfather of modern AI getting dragged into a spat (albeit unfairly) has meant that more rearchers will now be aware of the implications of their rearch.
LeCun的评估没有错，因为在特定模型的情况下，在包含⿊⼈⾯Kong(⽽不是主要包含⽩⼈⾯Kong)的数据集上训练模型不会产⽣荒谬的结果这样。 但是现代AI教⽗的优势被拖进了争吵中(尽管这是不公平的)，这意味着更多的研究⼈员现在将意识到他们研究的意义。
The misunderstanding clearly ems to emanate from the interpretation of the word “bias” — which in any discussion about the social impact of ML/AI ems to get crushed under the burden of its own weight.
这种误解显然源于对“偏见”⼀词的解释，在对ML / AI的社会影响的任何讨论中，它似乎都在⾃⾝的负担下被压倒了。
As Sebastian Raschka puts it, “the term bias in ML is heavily overloaded”. It has multiple ns that can all be mistaken for each other.
正如塞巴斯蒂安·拉施卡(Sebastian Raschka)所说，“机器学习中的术语偏差严重超载”。 它具有多种感觉，⽽所有这些感觉都可能会相互误解。
(1) bias (as in mathematical bias unit) (2) “Fairness” bias (also called societal bias) (3) ML bias (also known as inductive bias, which is dependent on decisions taken to build the model.) (4) bias-variance decomposition of a loss function (5) Datat bias (usually causing 2)
赤梨美来
(1) 偏差 (以数学偏差为单位)(2)“公平” 偏差 (也称为社会偏差 )(3)ML 偏差 (也称为归纳偏差) ，取决于建⽴模型的决策。 (4)损失函数的偏差-⽅差分解(5)数据集偏差 (通常导致2)
I imagine that a lot of gaps in communication could be covered by just being a little more preci when we u the terms.
我认为，使⽤这些术语时，只要稍微精确⼀点就可以弥补沟通中的许多空⽩。
On a lighter note, never mind Obama, the model even depixelized a dog’s face to a caucasian man’s. It sure loves the white man.
轻松⼀点，不要介意奥巴马，该模型甚⾄将狗的脸去像素化为⾼加索⼈的脸 。 它肯定爱⽩⼈。
Learning algorithms have inductive bias going beyond the bias in data too, sure. But if the data has a little bias, it is amplified by the systems, thereby causing high bias to be learnt by the model. Simply put, creating a 100% non-biad datat is practically impossible. Any datat picked by humans is cherry-picked and non-exhaustive. Our social cognitive bias result in inadvertent cherry-picking of data. This biad data, when fed to a data-variant model (a model who decisions are heavily influenced by the data it es) encodes the societal, racial, gender, cultural and political bias and bakes them into the ML model.
当然，学习算法的归纳偏差也要超出数据偏差。 但是，如果数据的偏差很⼩，则这些系统会对其进⾏放⼤，从⽽导致模型学习到较⾼的偏差。 简⽽⾔之，创建100％⽆偏的数据集实际上是不可能的。 ⼈类选择的任何数据集都是精⼼挑选的，并⾮详尽⽆遗。 我们的社会认知偏见会导致疏忽地选择数据。 当将这种有偏见的数据馈送到数据变量模型(该模型的决策受到其所见数据的严重影响)时，这些偏见数据将对这些社会，种族，性别，⽂化和政治偏见进⾏编码，并将其放⼊ML模型中。
The problems are exacerbated, once they are applied to products. A couple of years ago, Jacky Alciné that the image recognition algorithms in Google apologid for the blunder and assured to resolve the issue. However, instead of coming up with a proper solution, it simply blocked the algorithm from identifying gorillas at all.
⼀旦将这些问题应⽤于产品，这些问题就会加剧。 ⼏年前，杰基·阿尔⾟(JackyAlciné) ， 的图像识别算法 Google为这⼀错误道歉，并保证解决此问题。 但是，它没有提出适当的解决⽅案，只是根本阻⽌了算法识别⼤猩猩。
It might em surprising that a company of Google’s size was unable to come up with a solution to this. But this only goes to show that training an algorithm that is consistent and fair isn’t an easy proposition, not least when it is not trained and tested on a diver t of categories that reprent various demographic cohorts of the population proportionately.修仙女配文
⼀家Google规模的公司⽆法为此提出解决⽅案，这似乎令⼈惊讶。 但这仅表明，训练⼀致且公平的算法并⾮易事，尤其是在未对代表相应⼈⼝统计⼈群的多种类别进⾏训练和测试的情况下。
of facial recognition tech getting it terribly wrong came as recently as last week when a faulty facial recognition match led to a Michigan man’s arrest for a crime he did not commit. Recent studies by and the , or NIST, found that even though face recognition works well on white men, the results are not good enough for other demographics (the misidentification ratio can be more than 10 times wor), in part becau of a lack of diversity in the images ud to develop the underlying databas.
⾯部识别技术的错误是在上周才发⽣的，当时⼀次错误的⾯部识别⽐赛导致密歇根州⼀名男⼦因未犯罪⽽被捕。 和美国 NIST)的最新研究发现，即使⾯部识别在⽩⼈男性上效果很好，但对于其他⼈⼝统计⽽⾔，结果仍然不够好(误识别率可能会⾼出10倍以上)，部分原因是⽤于开发基础数据库的图像缺乏多样性。
Problems of algorithmic bias are not limited to image/video tasks and they manifest themlves in language tasks too.
算法偏差问题不仅限于图像/视频任务，它们也表现在语⾔任务中。
, i.e., it depends on external references for its understanding and the receiver(s) must be in a position to resolve the references. This therefore means that the text ud to train models carries latent information about the author and the situation, albeit to varying degrees.
，即， 理解取决于外部参考，并且接收者必须能够解决这些参考。 因此，这意味着⽤于训练模型的⽂本带有有关作者和情况的潜在信息，尽管程度不同。
Due to the situatedness of language, any language data t inevitably carries with it a demographic bias. For example, some speech to text transcription models tend to have higher error rates for African Americans, Arabs and South Asians as compared to Americans and Europeans. This is becau the corpus that the speech recognition models are trained are dominated by utterances of people from western countries. This caus the system to be good at interpreting European and American accents but subpar at transcribing speech from other parts of the world.
由于语⾔的位置性，任何语⾔数据集都不可避免地会带来⼈⼝统计学偏差。 例如，与美国⼈和欧洲⼈相⽐，某些针对语⾳的⽂本转录模型对⾮裔美国⼈，阿拉伯⼈和南亚⼈来说具有较⾼的错误率。 这是因为训练语⾳识别模型的语料库被来⾃西⽅国家的⼈们的话语所⽀配。 这使得该系统擅长于解释欧洲和美国的⼝⾳，但在抄录来⾃世界其他地区的语⾳时表现不佳。
Another example in this space is the gender bias in existing word embeddings (which are learned through a neural networks) that show females having a higher association with “less-cerebral” occupations while males tend to be associated with purportedly “more-cerebral” or higher paying occupations.
柯基犬好养吗
该领域的另⼀个例⼦是现有单词嵌⼊中的性别偏见(通过神经⽹络学习)，表明⼥性与“较少⼤脑”的职业有较⾼的关联，⽽男性往往与所
谓“较⼤脑”的职业有关联。⾼薪职业。
In the table below, we e the gender bias scores associated with various occupations in the embedding model. The occupations with positive scores are female-biad occupations and ones with negative scores are male-biad occupations.
在下表中，我们看到了嵌⼊模型中与各种职业相关的性别偏见得分。 得分为正的职业是⼥性偏向的职业，得分为负的职业是男性偏向的职业。
Image for post
Occupations with the highest female-biad scores (left) and the highest male-biad scores (right) (
Courtesy : )
⼥性偏见得分最⾼(左)和男性偏见得分最⾼(右)的职业(礼貌：: )
While it is easy for ML Rearchers to hold their hands up and absolve themlves of all responsibility, it is imperative for them to acknowledge that they — knowingly or otherwi — build the ba layer of AI products for a lot of companies that are devoid of AI experti. The companies, without the knowledge of fine-tuning and tweaking models, u pre-trained models, as they are, put out on the internet by ML rearchers (like GloVe, BERT, ResNet, YOLO etc).
虽然ML研究⼈员可以很轻松地举起双⼿并免除所有责任，但他们必须承认，⽆论有意或⽆意，他们都为许多没有AI的公司构建了AI产品的基础层专业知识。 这些公司不了解微调和调整模型，⽽是使⽤经过预先训练的模型，这些模型是由机器学习研究⼈员(例如
GloVe，BERT，ResNet，YOLO等)发布在互联⽹上的。窝沟封闭对牙有什么作用
Deploying the models without explicitly recalibrating them to account for demographic differences is perilous and can lead to issues of exclusion and overgeneralisation of people along the way. The buck stops with the rearchers who must own up responsibility for the other side of the coin.
部署这些模型⽽没有明确地重新校准它们以解决⼈⼝统计学差异是危险的，并且可能会导致⼈们在此过程中被排斥和普遍化的问题。 研究⼈员必须承担起硬币另⼀⾯的责任。
It is also easy to blame the data and not the algorithm. (It reminds me of the Republican stance on the cond amendment debate: “Guns don’t kill people, people kill people.”) Pinning the blame on just the data is irresponsible and akin to saying that the racist child isn’t racist becau he was taught the racism by his racist father.
也很容易指责数据⽽不是算法。 (这使我想起了共和党在第⼆次修正案辩论中的⽴场：“枪不杀⼈，⼈杀⼈。”)将责任归咎于数据是不负责任的，类似于说种族⼉童不是种族主义者，因为他的种族主义者⽗亲教他种族主义。
More than we need to improve the data, it is the algorithms that need to be made more robust, less nsitive and less prone to being biad by the data. This needs to be a responsibility for anyone who does rearch. In the meantime, de-bias the data.
不仅需要改善数据，还需要使算法更健壮，更不敏感并且更不容易受到数据偏差的影响。 这需要对任何从事研究的⼈负责。 同时，对数据进⾏反偏。
完工证明模板
The guiding question for deployment of algorithms in the real world should always be “would a fal answer be wor than no answer?”
在现实世界中部署算法的指导性问题应该始终是“错误的答案会⽐没有答案更糟糕？”
You can visit my page . My Twitter handle is @.
您可以 访问我的页⾯ 。 我的Twitter句柄是@ 。
算法偏见是什么
