The theme of facial recognition appears to be in the news at the moment. Not only, though, from the introduction of Apple’s latest thousand pound fondle slab, but from news about a research project that has been using machine learning techniques reportedly to identify sexuality.
Now first of all there is a lot of media hype which the researchers have been keen to dampen down. The aim of their research was not to create some sort of sinister gaydar, but rather to test the hypothesis of whether such a technique would work. In their notes they appear to be quite horrified by their findings. And whilst the researchers seemed to identify some traits in the biological makeup of faces pointing to whether someone was gay or straight, they also acknowledged a number of cultural factors (important, because the sample was made up of people who self identified their sexuality). For example,
… consistent with the association between baseball caps and masculinity in American culture, heterosexual men and lesbians tended to wear baseball caps…
So on first inspection that perhaps machine learning has developed a modern day interpretation of the pseudo-science of Phrenology (a topic introduced in day one of my study of Criminology as a laughable load of old bollocks), it turns out that what the researchers have found is that the way that someone looks might indicate their sexuality. That’s a much more nuanced thing.
And here in microcosm is the consistent problem with all of these techniques: correlation without understanding causality. Whilst the researchers might provoke debate and analysis of the reasons why we can ascertain something about sexuality from how people look, the machines do nothing to help explain the importance of why. Is it because of genetic makeup? Is it because of flawed sample data? Is it because of cultural norms? Is it a reflection of all of those things? Bigdatamachinelearningroboticanalytics (TM) will do little or nothing to answer those questions because they can’t.
And relying on correlation without causality is very risky. Especially when someone asks you to justify your decisions.
The debate that is being provoked by the research is interesting, but we shouldn’t obsess about whether you can spot someone who is gay. We should be much more concerned as to whether using machine learning correlations is sensible or morally ethical.
mmitII 