I am not 57. But I am 43. And it wouldn’t take much time for you to see that, just as it wouldn’t take me much time to see your age with just a look. Your brain is loaded with data about how cartilage wears over time. We carry with us a tacit understanding of skin elasticity. But our inherent understanding of crow’s feet and complexion and smile lines belie the complexity of the immense computation required to make sense of these distinctions. So it’s a wonder that a team from Microsoft called Project Oxford Face Detection got as far as they did when they set out to create how-old.net. Their idea is to teach a computer to recognize not just one face from another, but also the nuances that make up a face. Ultimately, the team aims to have this software used for facial verification, identification, grouping and other practical applications. But, for right now at least, it’s getting most of its press from social-media goers who are either thrilled by its accuracy or insulted by its imprecision. This is how it went for many of my colleagues, who apparently range from 14 to 77.
Or that’s what it might appear as. The reality is I only work with a few dozen people on a daily basis. But what if I worked with hundreds and hundreds of people? Well, then maybe the seeming inaccuracies might dissipate into a number much less significant.
Machine learning is a statistical game. One can talk about precision and recall, and it’s never 100%. In many ways, having it guess the age of an individual is one of the worst ways to show the power of a machine. These are tools that show best over datasets, not individual pieces of data, for trends and averages.
And, as it turns out, it did guess one of my pictures to be 42—not so bad. While my co-workers had inaccurate results when using how-old.net, many more probably found the program to work eerily well for them. It would be most interesting to actually understand what sort of accuracy their model had, but you can’t do that with just one individual’s pictures. You need a whole host of them.
We have to deal with machine learning a lot, we have somewhere north of thirty different models that we’re actively maintaining and re-training periodically, from part-of-speech tagging to entity recognition, across all the different languages we deal with.
It’s a neat little app. It’s gotten great virality (score!), it’s gotten some flack for odd permissions letting them use your images in their advertising (boo!), and it’s a good demonstration of machine learning. I like to know a little bit more – like in what age range is it most accurate? Do different ethnicities have different accuracies? Can it fix those annoying bags under my eyes that I seem to be getting more and more often?