TE
科技回声
首页24小时热榜最新最佳问答展示工作
GitHubTwitter
首页

科技回声

基于 Next.js 构建的科技新闻平台,提供全球科技新闻和讨论内容。

GitHubTwitter

首页

首页最新最佳问答展示工作

资源链接

HackerNews API原版 HackerNewsNext.js

© 2025 科技回声. 版权所有。

What is Math washing?

21 点作者 mihau大约 7 年前

3 条评论

a_puppy大约 7 年前
The article is correct that machine learning doesn&#x27;t remove bias. If the training data is biased, the output will be biased in the same way. But machine learning doesn&#x27;t add bias, either. If the training data are unbiased, the output will be unbiased. In this sense, algorithms are fairer than humans.<p>So if someone wants to argue that a machine learning model is or is not biased, they should base that argument on how the model is trained. For example: suppose a bank wants to use a machine learning model to predict who to make loans to. Historically, human bank managers made those decisions, and they tended to have a bias against people from the wrong side of the tracks. There are several possibilities:<p>* If the bank trains the model on the bank managers&#x27; decisions, and it uses ZIP code as a feature, then it will discriminate against people from the wrong side of the tracks just like the human bank managers did.<p>* If the bank trains the model on the bank managers&#x27; decisions, but the only features it uses are monthly income and existing debts, then it will probably be unbiased (although it&#x27;s still conceivably possible for there to be a bias).<p>* If the bank runs a controlled experiment by approving loans for 100 people at random, and trains the model on which loans were paid back, then the results of the model will be fair; it will accurately predict how likely people are to pay back loans, regardless of which side of the tracks they live on.<p>* If the bank trains the model on loans made by human bank managers, but it trains the model to predict loan repayment instead of loan approval, then the algorithm will actually _invert_ the bank managers&#x27; biases. If the bank managers never approved loans for people from the wrong side of the tracks unless they were an extraordinarily safe bet, then the algorithm will conclude &quot;people from the wrong side of the tracks always pay back their loans!&quot;<p>Arguments about machine learning bias should be based on these sorts of specific details, rather than assuming &quot;algorithms aren&#x27;t biased&quot; or &quot;algorithms are biased&quot;.
评论 #16615283 未加载
rdtsc大约 7 年前
I like the idea of algorithm transparency. There were a few cases were people have tried to subpoena the source code from red light cameras, or DUI breathalyzers.<p><a href="http:&#x2F;&#x2F;digital.law.washington.edu&#x2F;dspace-law&#x2F;bitstream&#x2F;handle&#x2F;1773.1&#x2F;1069&#x2F;7WJLTA123.pdf?sequence=4" rel="nofollow">http:&#x2F;&#x2F;digital.law.washington.edu&#x2F;dspace-law&#x2F;bitstream&#x2F;handl...</a><p>It&#x27;s not clear cut as it seems is that states can side-step forcing discovery of the source code by arguing that the state prosecutor does not control &#x2F; own the source code.<p>This is of course moot in regards to a private entity like FB, Google or Twitter. They can ban anyone for any reason. Do they even need the &quot;algorithms&quot; excuse? Maybe if they ban someone for whom they get a lot of negative PR, then they can issue an apology &quot;Sorry you feel this way, but algorithms did it, we swear&quot;.
klondike_大约 7 年前
It&#x27;s amazing how big companies like Google can be so oblivious to how biased their algorithms are. See: YouTube &quot;hate speech&quot; detection, Facebook and fake news.<p>Algorithms more often then not inherit the same biases their programmers have. Having those biases determined mathematically makes no difference.