Shouting down research

Two researchers at Shanghai Jiao Tong University recently posted a paper which quickly garnered the attention of the online commentariat, some fairly balanced and some like this:

From the department of: I Really Hope This Is Satire, but Given It Is 2016 I Cannot Be Sure: Researchers at Shanghai University published a paper called ‘automated inference on criminality using face images‘. The paper uses deep learning to explore correlations between someone’s appearance and the chance of them being a criminal. It’s modern phrenology – 19th century junk science where people believed you could measure someone’s skull and use it to infer traits about their intelligence (the Nazi’s were influenced by this). I can see no merit to this paper whatsoever and am mystified that the researchers were not warned off of publishing this absurd paper. If someone feels my views are wildly wrong here I’d love to hear from you and will (if you’re comfortable with it) put the correspondence in the next newsletter.

The Nazi-laden denunciation above landed in my mailbox on November 21, courtesy of Import AI, "a weekly newsletter about artificial intelligence, read by thousands of experts", written by Jack Clark, Strategy and Communications Director at OpenAI. Given the considerable cachet of OpenAI, the wide distribution of this newsletter throughout the AI community, the strength of the language used, and its likely impact on the paper's authors (whom I do not know and am in no way associated with), I decided to spend the time needed to respond properly, and did so on November 23. My reply, including my permission to include it in the newsletter, was backquoted and acknowledged in private on November 27. On November 29, after an unusual delay of one day (the newsletter is normally sent out on Mondays), the week's issue of Import AI appeared, having only this to say about the exchange:

Ethical machine learning: should we conduct experiments, even if they seem to be offensive? The answer to that question was 'yes' from a few Import AI readers, who took issue with my characterization of the 'automated inference on criminality using face images' paper from last week. Some readers pointed out that this could be an interesting experiment to run, and I countered by saying I'd need a much larger section of the research paper given over to an evaluation of the ethical and moral context of the experiment.

Since this is obviously not how you "put the correspondence in the next newsletter", I conclude that OpenAI's Strategy and Communications Director has decided to renege on his offer. As I am no more a fan of broken promises and of wasted time than of rants from a bully pulpit associated with an organization which I had previously seen no reason to disrespect, I am posting my response for all to see, and I encourage the other readers of Import AI who've had their replies suppressed to do the same.

My original response, with markup added for readability and titles replacing raw URLs in the reference list but otherwise unchanged, follows below.

Tommy Anderberg

November 29, 2016

"From the department of: I Really Hope This Is Satire, but Given It Is 2016 I Cannot Be Sure"

It's been a while since I last got in trouble for something I wrote. Looks like it's time again.

First, I would like to express my appreciation for OpenAI in general and your newsletter in particular. I would not bother to respond if I considered you easily disregarded.

That said...

Your critique of [1] does not contain a single word about the scientific plausibility of the claimed correlations, nor about the paper's methodology. The Godwin-defying analogies you put forth do not touch upon the validity of the paper's findings; they are all about what you fear that they will lead to, namely facial profiling (that's a fancy term I just came up with by obvious substitution [2]).

I hold the following truth to be self-evident: science is about how the world is, not about how we wish it were. To get the two mixed up is to live in a fantasy, and that does not usually lead to good outcomes. It is an error to reject a finding because we don't like its implications, and it is an error to suppress a finding for fear that it might be misunderstood by others. That would mean letting science be hostage to (presumed) stupid people. Go down that road and the unreflected knee-jerk reactions and jumping to conclusions which you fear from others become your own.

When confronted with a troubling research result, I think it is helpful to slow down and ask the following questions:

Q.1) Is it plausible that things might work this way?
Q.2) Have all possible explanations been considered?
Q.3) If this really is how things work, what can we do about it?

Here are my answers with regard to [1]:

A.1) Yes. Correlations between behavior and genetically determined physical features are not only plausible, they are already well documented.

Let's say I present you with a set of face images, and tell you only this: "half of these people are murderers; identify them". You look at the photos and notice that half of them are men, half are women. Without any additional information, all you have to go on is the known fact that 95% of homicide perpetrators worldwide are men [3]. No matter how much you dislike this fact (as a man, I certainly do) you should pick the men as your primary suspects.

You know this, so you already know that a genetically determined attribute - gender - is strongly correlated with violent crime. It would be surprising if no other correlations were to be found between genetically determined attributes and behavior. Two well known examples of such correlations are digit ratios (the ratio of index finger length to ring finger length) and facial width-to-height ratios (fWHR): low digit ratios and wide faces tend to accompany high testosterone levels and are predictive of stereotypically male (aggressive, dominant, driving, status-striving) personality [4][5][6][7][8][9][10]. Depending on context and other traits, this kind of personality can lead to success in sports and business - or to jail.

A.2) In the present case, this question should really be asked of the reader. The authors of [1] only establish a correlation, not what causes it.

A reader may jump to - or fear - the conclusion that people with certain facial features are genetically predisposed to become criminals, but there are other possibilities.

Look at the three discriminative features identified in Fig. 4b: they make the face look more determined or prying (the eyes), less smiling (smaller mouth) and more judgemental or angry (pursed lips). The "normal" face of Fig, 4a is noticeably more relaxed. (The same is true of Fig. 6 a and c versus b and d. The authors claim that they "appear hardly distinguishable"; I disagree.)

Now make a simple thought experiment. Take two twins, identical in everything except for their faces. Twin A has the face of Fig. 4a, twin B has the face of Fig. 4b (or, if you prefer, one of the faces of Fig. 9a). As they go through life, everybody they interact with is going to perceive A as more relaxed and friendly than B, quite independently of their (identical) personality. In most cases it's not going to be a reasoned assessment, just a feeling, maybe not even a conscious one. It takes you less than one tenth of a second to attribute character traits to a new face, and you do it automatically, whether you want to or not [11]. The effect is known to be real and deeply influential: it can determine who gets a loan, goes to jail, becomes CEO or wins a political election [12][13][14][15][16] (see [17] for a nice summary, also featuring twins).

So in a very real sense, facial profiling is something we already do. We have always done it.

Remember the old saying "smile and the world smiles with you"? In our thought experiment, A is going to have a better time than B, even if B never changes. But inevitably, B will change. Being persistently perceived and treated as not so nice will make him just that. And should the resulting coarsening of his personality eventually cause a run-in with the law, his looks alone will make him more likely to be convicted than A would have been.

Incidentally, I didn't notice any outcry when CSAIL presented its large-scale image memorability estimator [18] and MarketWatch used it to show that well known CEOs have highly memorable faces [19]. Here's a hunch: these CEOs would also score highly for "criminality" according to [1], because it is deviations from the norm that makes a face memorable. And I say this as someone whose photo scores even higher than Zuckerberg's for memorability, and therefore, presumably, for crookedness.

A.3) Whether genetic predisposition, social stigma due to unfortunate looks or something else is behind the findings of [1], it would be preposterous to conclude that all people with "criminal" faces should be preemptively incriminated. Note again that the authors suggest no such thing; what I am addressing here is the fear of such suggestions which I expect to be invoked as an argument against pursuing this line of inquiry.

To suggest such things one must of course fail at elementary logic; one must believe that the statement "all idiots are conservatives" implies that all conservatives are idiots, or equivalently that the statement "all suicide bombers are Muslims" (if it were true [20]) would imply that all Muslims are suicide bombers. The resounding condemnation and quick shelving of recently vented political proposals related to the second example was a mild breeze compared to the storm of indignation which would greet the policing of those whose only "crime" is a face deemed deviant by computers.

Could something good come of this? Quite possibly. Remember, the CNN of [1] is doing what you and I have been doing all our lives. Yes, it is terrible that we use facial features to judge each others character, but that's what we are hardwired to do. Denying this reality might make some people feel better about themselves, but it will not help those of us who are most at risk from its consequences. So let's admit it, and let's ponder how an impartial judge of risky looks might be put to good use. Maybe it could gently remind us to be a little nicer to those who would be most helped by it?

Tommy Anderberg

November 23, 2016

P.S. While I am comfortable with this text being put in the newsletter, I would strongly prefer that my email address not be included.

[1] [1611.04135] Automated Inference on Criminality using Face Images

[2] Racial Profiling | American Civil Liberties Union

[3] Global Study on Homicide | UNODC

[4] Second to fourth digit ratio confirms aggressive tendencies in patients with boxers fractures

[5] Digit ratio (2D:4D), aggression, and dominance in the Hadza and the Datoga of Tanzania

[6] In your face: facial metrics predict aggressive behaviour in the laboratory and in varsity and professional hockey players

[7] Facial width-to-height ratio predicts achievement drive in US presidents

[8] Telling facial metrics: facial width is associated with testosterone levels in men

[9] An Examination of the Associations Between Facial Structure, Aggressive Behavior, and Performance in the 2010 World Cup Association Football Players