| Faception Can Spot Terrorists And Intelligent People Using How They Look |
| Written by Mike James |
| Monday, 30 May 2016 |
|
This one is a can of hornets. A new company, Faception, claims to have trained an AI system to recognize various categories of people from just looking at their face. In short, it is an instant profiler working with just a photo of the human in question. A miracle or a nightmare? You can train a neural network, or any other sort of machine learning algorithm, using any data and hope to get good results. In this case the idea is that you take a lot of photos of human faces that represent one category and photos of another - say intelligent people v unintelligent people and train the network to tell them apart based only on their face. Why is this even possible in the wildest of imaginings?
The argument is that if a characteristic like intelligence is controlled by a particular DNA type then this DNA type will result in a correlation between the way people look and the characteristic. The web site contains references to scientific research on identical twins and on how genes control facial appearance. However all of these are indirect and even disputed to some extent. There is no direct evidence that how you look is correlated with personality traits. The situation is so complicated that the only way to debunk such an idea is to show that there is no correlation. This is a hard thing to do because of the vagueness inherent in measuring "what someone looks like". However, a neural network, and many learning methods, don't need definitions of what might be different - they learn what is different from the data. If the machine learning can discriminate the different categories with sufficient accuracy then there is something going on between the way something looks and the characteristic in question.
The company claims to have a set of classifiers for: HIGH IQ Endowed with a reasoning skills, like logic, spatial skills. Self-made people, free-thinkers and entrepreneurs. Exceptionally gifted, tend to be less socially oriented, value truth, facts and logic more than emotional relations. Creative and independent minded, with exceptional concentration abilities, a high intellect and mental capacity. Academic Researcher Endowed with sequential thinking, high analytical abilities, a multiplicity of ideas, deep thoughts and seriousness. Creative, with a high concentration ability, high mental capacity, and interest in data and information. Professional Poker Player Endowed with a high concentration ability, perseverance and patience. Goal-oriented, analytical, with a dry sense of humor. Silent, devoid of emotion and emotional expression, strict and sharp minded, with high critical perception. Bingo Player Endowed with a high mental ceiling, high concentration, adventurousness, and strong analytical abilities. Tends to be creative, with a high originality and imagination, high conservation and sharp senses. Brand Promoter Endowed with a high self-confidence, authoritative, charismatic and magnetic personality, with high intellect and high verbal ability. Tends to be kind, sociable and direct, and very practical. White-Collar Offender Tends to have a low self-esteem, a high IQ and charisma. Anxious, tensed and frustrated, competitive, ambitious and dominant. Usually loves to take risks and have a dry sense of humor. Terrorist Tends to be aggressive, active, thrill seeking, cruel and psychologically unbalanced. Usually suffers from mood swings, a sense of inferiority and unsettled self-confidence. Pedophile Suffers from a high level of anxiety and depression. Introverted, lacks emotion, calculated, tends to pessimism, with low self-esteem, low self image and mood swings. The AI can provide a score for each of these classes after a few seconds looking at a photograph of the person. Of course, the important fact to know is the accuracy and, unfortunately, no precise figures are given. The Washington Post reports that Faception chief executive Shai Gilboa says that some of the classifiers are 80% accurate. This may sound good, but notice that means some of them are less than 80% accurate. In addition 80% accuracy isn't really that good and may not be good enough for many tasks. For example, suppose terrorists are 1% of your sample, they are likely to be much less than 1%. Then a 20% error rate means that in 1000 people you will accuse approximately 200 innocents and let though 2 terrorists. You may be thinking "but there is a signal". Not necessarily. The history of AI is littered with strange demonstrations that really didn't demonstrate what they claimed to. There is the well-known, very early, example of a single perceptron trained to tell males from females from photos. It hit the headlines when it classified all of the Beatles as girls. In practice what it was doing was measuring the amount of hair in each photo. My only experience of the effect didn't hit the headlines but might better illustrate the problem. I had a classifier working well, 99% accuracy, on discriminating a set of photos of planet Mars, trying to pick out possible signs of flowing water. Then I noticed that the photos of one class looked slightly different to the other. It turned out that the photos of the two groups had been printed on different grades of photographic paper (this was some time ago) and the majority of group one had a higher contrast than group two. A human could only really see it when they were told it was so; the learning rule picked up on it at once. When a learning rule discriminates it might not be using the sort of features that you expect. They may be entirely incidental to the desired classification. For example, photos of the high IQ group might simply have a better skin condition or be fatter because the subjects earn more. This is the AI equivalent of the "correlation does not imply causation" rule. Whatever is going on, this isn't going to be an uncontroversial use of AI. To quote Alexander Todorov, a Princeton psychology professor whose research includes facial perception (Washington Post): “Just when we thought that physiognomy ended 100 years ago. Oh, well.” More InformationRelated ArticlesOpenFace - Face Recognition For All How Old - Fun, Wrong, Potentially Risky?
To be informed about new articles on I Programmer, sign up for our weekly newsletter, subscribe to the RSS feed and follow us on, Twitter, Facebook, Google+ or Linkedin.
Comments
or email your comment to: comments@i-programmer.info |
| Last Updated ( Monday, 30 May 2016 ) |
