去年11月我和我的博士生张熙在arxiv上贴出了一篇题为“Automated Inference on Criminality usingFace Images”的论文。该论文在各国学术界，尤其是互联网上引起了广泛的关注和争议。国外一些媒体哗众取宠，歪曲我们的研究是搞机器看脸识罪犯，国内不少网络媒体也纷纷跟进。本来我想，有英语原文在，清者自清，我们研究的纯学术性一目了然。现在发现由于互联网上不少不准的翻译，加上标题党的渲染，不少善良正直的民众对我们的科研有了本质性，原则性的误解，对他们（可能不能读我们的英文原文）我们有责任澄清如下。

首先，英语词criminality，类似personality（个性），是一个很泛的中性词，指可能促成犯罪的个人或群体的属性，大概可以译成犯罪性吧。正如我们说人都有兽性的一面，好人也有一定的犯罪可能性，所以好的制度是如此重要。我们的结论绝对不是机器可以看脸识罪犯，而是机器发现了脸部特征与犯罪性有一定的相关性。我们文中反复声明，我们没兴趣，也没社会学、犯罪学、心理学的专业知识去解释这种相关性，更没有暗示我们的方法可以实用。

很多人对我们科研的误解显然是源自一个叫“基础概率谬误”（base rate fallacy）的常见判断误区。人脑往往被一个特定事件的高条件概率锁住，而忘记了该事件在大环境里发生的极低概率。我们文章中最强的基于深度学习的面相分类器有89%的识别率（注意：这是基于我们目前的训练数据，还有待用更大的数据核实），很多人就认为，这么高，这还不一试一个准！（国外有文章报道我们时就惊呼“correct 9  out 10 times”）。有人在网上调侃“教授，把你的脸放进去试试”。好吧，这倒是蛮好玩的。假设我的脸被测阳性，我有多高概率有犯罪倾向呢？计算这个概率需要用贝叶斯定理：

P(罪|+)  =  P(+|罪)*P(罪) / [ P(+|罪)*P(罪) + P(+|非)*(1-P(罪)) ]

上式中P(+|罪)=0.89 是罪犯的脸被我们深度学习测试方法判阳性的概率，P(罪)=0.003是中国的犯罪率，

P(+|非)=0.07是我们方法假阳性的概率。将这些数值代入贝叶斯公式，结果是武筱林有3.68%的概率犯罪：-）。我想，这一路从89%到3.68%走下来，原来不少骂我的人就释怀了吧。那些叫着要纪委用我们的方法的网友也该歇歇了。不过，我这里再次郑重声明，我们坚决反对在执法司法中使用我们的方法，原因不仅仅是上面演算的结果。

上周，新智元公众号登出了谷歌三个研究员讨论人工智能伦理的长文，文中多处提到我们的工作。新智元的翻译有很多错误，还玩标题党：“谷歌研究员两万字批驳上交大用深度学习推断犯罪分子”。译者显然既没有读我们论文的全文，也没有向我们求证，用“上交大用深度学习推断犯罪分子”恶意歪曲庸俗化我们的研究。照单全收国外网站上的东西，对我们居高临下进行道德审判，靠触发网上铺天盖地的谩骂来赚取点击率。

我们欢迎严肃认真地就我们的论文开展学术讨论。这半年来各国的教授、学者、研究生（如来自Harvard，Cornell、CalTech、Univ. of Washington等大学从事遗传学、心理学、计算机科学研究的）纷纷来emails交换意见。《科学美国人》的自由撰稿人数次来电话，讨论乃至争论数小时。他们关心的问题与谷歌文中的那些高度重合，一并解答如下：

Garbage in?

As much baffled by the intellectually chauvinistic tone of theGoogle authors, we agree with them on their progressivesocial values. There is really no need to parade infamous racists in chronicorder with us inserted at the terminal node. But the objectivity does exist, at least in theory, independent ofwhatever prevailing social norms.  

One of us has a Ph.D incomputer science; we know all too well “garbage in and garbage out”.  However, the Google authors seemed to suggestthat machine learning tools cannot be used in social computing simply becauseno one can prevent the garbage of human biases from creeping in.  We do not share their pessimism.  Like most technologies, machine learning is aneutral tool.  If it can be used toreinforce human biases in social computing problems as the Google authors argued,then it can also be used to detect and correct human biases (prejudice).  They worry about the feedback loop butconveniently do not see that the feedback can be either positive or negative. Granted,the criminality is a highly delicate and complex matter; however, well-trained humanexperts can strive to ensure the objectivity of the training data, i.e.,rendering correct legal decisions independent of facial appearance of theaccused.   If the labeling of training face images isfree of human biases, then the advantages of automated inference over humanjudgment in objectivity cannot be denied.

Even in the presence of labelnoises, regardless they are random or systematic, scientific methods do existto launder and restore/enhance credence to the results of statisticalinferences.  We should not short changescientific knowledge for any shade of populism.

Risk of Overfitting

Our critics are quick topoint out the relatively small sample set used in our experiments and the riskof data overfitting. We are sorely aware of this weakness but cannot get moreID images of convicted Chinese males for obvious reasons (this Google article mighthave dashed all our hopes to enrich our data set).  However, we did make our best efforts tovalidate our findings in Section 3.3 of our paper, which opened as follows butcompletely ignored by the Google authors.    

“Given the high socialsensitivities and repercussions of our topic and skeptics on physiognomy [19],we try to excise maximum caution before publishing our results. In playingdevil’s advocate, we design and conduct the following experiments to challengethe validity of the tested classifiers …”

We randomly label the faces of our training set as negative andpositive instances with equal probability, and run all four classifiers to testif any of them can separate the randomly labeled face images with a chancebetter than flipping a coin.  All faceclassifiers fail the above test and other similar, more challenging tests(refer to our paper for details).  Theseempirical findings suggest that the good classification performances reportedin our paper are not due to data overfitting; otherwise, given the same sizeand type of sample set, the classifiers would also be able to separate randomlylabeled data.

White Collar

Regarding to the question onthe white-collared shirts in and not in the ID photos, we forgot to clarifythat in our experiments of machine learning, we segment the face portion out ofall ID images; the face-only images are used in training and testing.  The complete ID photos are presented in ourpaper only for illustration purposes.  

Nevertheless, the cue of white collar exposes an importantdetail that we owe the readers an apology. That is, we could not control forsocioeconomic status of the gentlemen whose ID photos were used in ourexperiments.  Not because we did not wantto, but we did not have access to the metadata due to confidentialityissues.  Now reflecting on this nuance,we speculate that the performance of our face classifiers would drop if theimage data were controlled for socioeconomic status.  Immediately a corollary of social injusticemight follow, we suppose.  In fact, thisis precisely why we think our results have significance to social sciences.  

In our paper, we have also taken steps to prevent the machinelearning methods, CNN in particular, from picking up superficial differencesbetween images, such as compression noises and different cameras (Section 3.3).

谷歌的长文详细回顾了关于这个问题的历史渊源，讲到从古希腊、达尔文、希特勒、直到今天美国的种族主义者都热衷面相学，含沙射影我们搞“scientific racism”。他们完全不顾我们论文中的声明及整章的技术细节，也不像其他学者那样向我们求证，而是将他们断章取义的臆断强加给我们，非常的不公正，不专业，用的是文革中打棍子盖帽子的手法。