机器学习与分布式机器学习
Whether you like it or not, the impact of machine learning on your life is growing very rapidly. Machine learning algorithms determine whether you would get the mortgage for your dream home, or if your resume would be shortlisted for your next job. It is also changing our workforce rapidly. Robots are taking over warehouses and factories, and self driving cars are threatening to disrupt the jobs of millions of professional drivers across the world. Even law enforcement agencies are increasingly using machine learning to screen for potential criminal leads and assess risks.
无论您是否喜欢,机器学习对您生活的影响都在Swift增长。 机器学习算法确定您是否将为梦想中的房子获得抵押贷款,或者您的简历是否会入围下一份工作。 这也正在Swift改变我们的员工队伍。 机器人正在接管仓库和工厂,自动驾驶汽车正威胁着破坏全球数百万专业驾驶员的工作。 甚至执法机构也越来越多地使用机器学习来筛选潜在的犯罪线索并评估风险。
Unfortunately, all these advancements in technology may be perpetuating and exacerbating the biases ailing our society. In one of the early examples of algorithmic bias, 60 women and ethnic minorities were denied entry to St. George’s Hospital Medical School per year from 1982 to 1986, because of a new computer-guidance assessment system that denied entry to women and men with “foreign-sounding names” based on historical trends in admissions. Or more recently, in 2016, TayTweets, a chat bot trained by Microsoft on Twitter data started spouting racist tweets.
不幸的是,所有这些技术进步可能使并加剧了困扰我们社会的偏见。 在算法偏见的早期示例之一中,由于新的计算机指导评估系统拒绝“有“性别歧视”的男女进入,从1982年到1986年,每年有60名妇女和少数民族被拒绝进入圣乔治医院医学院。外来名字”根据录取的历史趋势。 或更最近,在2016年,TayTweets(一种由Microsoft训练而成的聊天机器人,在Twitter数据上)开始喷出种族主义推文。
![Image for post](https://miro.medium.com/max/9999/1*wrKQt8faar7KvGMIj63Bcg.png)
All these advancements are raising very valid questions about how machine learning practitioners can ensure fairness in their algorithms. What is fair is an age old question. Thankfully, a lot of research has been going on in this area. In this post, I am going to talk about the most common set of problems you might run when trying to ensure that your machine learning model is bias free.
所有这些进步都引发了有关机器学习从业人员如何确保算法公平的非常有效的问题。 公平的是一个古老的问题。 值得庆幸的是,该领域已经进行了很多研究。 在这篇文章中,我将讨论在尝试确保您的机器学习模型没有偏差时可能会遇到的最常见问题集。
代表性不足 (Underrepresentation)
One of the most common causes of bias in machine learning algorithms is that the training data is missing samples for underrepresented groups/categories. This is the reason why Siri frequently has a hard time understanding people with accents. This is also what caused the famous Google photos incident where black people were tagged as gorillas. So it is really important to make sure the training data has representation from all the underrepresented groups. Another way to easily detect this early is to deploy a second algorithm which predicts whether the data in production is close to the training data and intervene early if that is not the case.
机器学习算法存在偏见的最常见原因之一是,训练数据缺少代表不足的群体/类别的样本。 这就是Siri经常很难理解带有口音的人的原因。 这也是导致著名的Google照片事件的原因,其中黑人被标记为大猩猩。 因此,确保训练数据具有来自所有代表性不足组的代表性非常重要。 可以很容易地及早发现这种情况的另一种方法是部署第二种算法,该算法可以预测生产中的数据是否接近训练数据,如果不是,则尽早进行干预。
![Image for post](https://miro.medium.com/max/9999/1*yDCzmCHHuFoVR3S7bI6Sxw.jpeg)
![Image for post](https://miro.medium.com/max/9999/1*VR0QitFGUsN6cGntsZhUEw.jpeg)
社会反馈循环 (Social feedback loops)
A recent OpenAI paper about their groundbreaking language model found that “occupations in general have a higher probability of being followed by a male gender identifier than a female one”. There have been numerous cases in the press where biases in the training data are inherited by the models. There have been experiments which conclude that higher paying job ads are only shown to men or ads for houses in white neighborhoods are only shown to white people. Various language models trained on text data from the internet tend to attach negative connotations to women or people of color. This happens, because our existing social biases reflect in the training data. This kind of bias is very hard to correct as in most cases because it’s really hard to get unbiased data. One simple strategy to mitigate this is to find data points which represent positive historical data for the marginalized group and upsample those records in the training data.
OpenAI最近关于其突破性语言模型的一篇论文发现,“一般而言,占领之后,男性性别识别符比女性具有更高的可能性”。 媒体中有很多情况是模型继承了训练数据中的偏差。 已经有实验得出结论,只向男性展示高薪工作广告,或仅向白人展示白人社区房屋广告。 接受来自互联网的文本数据训练的各种语言模型往往给女性或有色人种带来负面含义。 发生这种情况是因为我们现有的社会偏见反映在培训数据中。 像大多数情况下一样,这种偏差很难纠正,因为获取无偏差的数据确实很困难。 减轻这种情况的一种简单策略是找到代表边缘化群体的正历史数据的数据点,并对训练数据中的这些记录进行上采样。
相关领域 (Correlated fields)
One common solution to make sure the data doesn’t train on sensitive attributes is by removing the features from training data altogether. But there might still be correlated attributes which might cause the model to discriminate against underserved communities. For example, zip code might be correlated to race, and name might be correlated to both race and gender, etc. It’s often observed that machine learning models trained on resumes learn to value male names more than female names. An Amazon study found that the name “Jared” is very good to get shortlisted for a job.
确保数据不针对敏感属性进行训练的一种常见解决方案是完全从训练数据中删除功能。 但是可能仍然存在相关属性,可能导致模型与服务不足的社区进行区分。 例如,邮政编码可能与种族相关,名字可能与种族和性别相关,等等。人们经常观察到,在履历表上训练的机器学习模型比男性名字更重视男性名字。 亚马逊的一项研究发现,“ Jared”这个名字非常适合入围工作。
忽略敏感属性 (Ignoring sensitive attributes)
Blindly removing the sensitive attributes features from the training data might even cause harm in some cases. For example, women are less likely to cause accidents, hence in some places, it is legal to use gender to determine insurance quotes. In recidivism studies, it’s found that women are less likely to re-offend. In such cases where the disadvantaged category actually does better than the dominant category, it might be better to include this as a feature in the training data.
在某些情况下,一味地从训练数据中删除敏感属性特征可能会造成伤害。 例如,妇女发生事故的可能性较小,因此在某些地方,使用性别确定保险报价是合法的。 在累犯研究中,发现女性不太可能再次冒犯。 在这种情况下,弱势类别实际上比优势类别要好,将其作为特征包括在训练数据中可能会更好。
过度补偿偏见 (Overcompensation for bias)
It’s very unlikely this will be a problem but adding this here to be comprehensive. Let’s say your model decides whether to provide a loan to an individual based on their credit score and if the individual pays back the loan, their score increases, otherwise it decreases. And your long term objective is to increase the average credit score of the population. In this case, to compensate for bias you might want to change the threshold to approve more loans for a certain section of the population. But if you overcompensate so much that the number of defaults are higher than the number of successful paybacks, the average credit score might decrease for this section of the population and ends up hurting them. Google provides a library called ml-fairness-gym which can help with simulating such long run effects of enforcing fairness in machine learning models.
这不太可能是一个问题,但是在此添加它是全面的。 假设您的模型根据他们的信用评分决定是否向个人提供贷款,如果个人还清贷款,则他们的评分会提高,否则会降低。 您的长期目标是提高人口的平均信用评分。 在这种情况下,为了补偿偏差,您可能需要更改阈值,以批准特定部分人口的更多贷款。 但是,如果您过度补偿过多,以致违约次数大于成功回收的次数,则此部分人口的平均信用评分可能会降低,并最终伤害他们。 Google提供了一个名为ml-fairness-gym的库,该库可以帮助模拟在机器学习模型中实施公平性的长期影响。
Allen Turing said that if a machine is expected to be infallible, it cannot also be intelligent. But everyday another critical piece of social infrastructure is adding “artificial intelligence” to its decision making process without any legal checks and balances. It’s very easy for various biases to slip into a machine learning model, unless you are extremely cautious. Also, our legal systems haven’t caught up with the new technology, which continues to deeply affect our lives. Hence right now, it’s the responsibility of people working on these models to ensure that the models are fair and free from biases.
艾伦·图灵(Allen Turing)表示 ,如果希望某台机器无懈可击,那么它也不可能是智能的。 但是每天,社会基础设施的另一个关键部分是在决策过程中增加“人工智能”,而没有任何法律制衡。 除非您非常谨慎,否则各种偏见很容易渗入机器学习模型。 此外,我们的法律体系还没有赶上新技术,而新技术继续对我们的生活产生深远的影响。 因此,现在,致力于这些模型的人们有责任确保这些模型是公平的并且没有偏见。
Fortunately, people in research are paying attention to this issue and adjacent issues like explainability of machine learning models.
幸运的是,研究人员正在关注这个问题以及诸如机器学习模型的可解释性之类的相邻问题。
Google’s Tensorflow has “fairness indicators” to detect biases in your model
Google的Tensorflow具有“ 公平指标 ”,可检测模型中的偏差
Microsoft research launched a FATE (fairness, accountability, transparency and ethics) group to study social implications of artificial intelligence
微软的研究发起了一个FATE (公平,问责,透明度和道德)小组,研究人工智能的社会影响
IBM open sourced the AI fairness 360 tool to help you build bias free models
IBM开源了AI fairness 360工具,可帮助您构建无偏差模型
The paper “Delayed Impact of Fair Machine Learning” by Liu et al was awarded best paper award at ICML 2018.
Liu等人的论文“ 公平机器学习的延迟影响 ”在ICML 2018上获得了最佳论文奖。
But these tools are nowhere close to being enough or perfect. There is a very strong need for much more rigorous tools and research in this area. Technology has been used both for good and bad. It’s the responsibility of the creators of the technology to make sure that the their products make the world a better place for all of us. I really liked this famous talk by Kate Crawford at NIPS 2017, where she calls upon the industry to invest more in understanding the social impact of artificial intelligence.
但是,这些工具还远远不够完善。 在这方面非常需要更严格的工具和研究。 无论好坏,都使用了技术。 技术创造者有责任确保他们的产品为我们所有人创造一个更美好的世界。 我真的很喜欢Kate Crawford在NIPS 2017上的这一著名演讲 ,她呼吁业界投入更多资金来了解人工智能的社会影响。
![Image for post](https://miro.medium.com/max/9999/1*vjSTcnmaiDqw_0ZfMnztdQ.png)
翻译自: https://towardsdatascience.com/biases-in-machine-learning-61186da78591
机器学习与分布式机器学习