The idea was to study image classification through the notion of social norms and normality: for instance, what can be considered as ugly ? Also, what are the implications of subconscious physiognomy ? This program is supposed to show the limits of this approach by having the computer to decide for us who is good, bad or (and?) ugly.
My idea was to first train the classifier with images from various source : Supreme Court Justices as Good people, Most wanted people by the FBI as Bad people, and face pictures of people considered as Ugly (from a website I found on the internet, their work is morally questionable, so i won’t display these images). Then I used a database of face images as testing images : for each one, the classifier output a triplet of probability to belong in each of these categories. Their image is then mapped into a triangle.
The result is not that great, mainly because I think that the features where not that relevant (pixels of the images), and the 3 categories are not well-balanced (a 2 dimensional grid should have been more appropriate, with beautiful/ugly and good/bad as axes), but it is not that important, since he purpose here is to show the limit of classification.