Criminal Machine Learning UP082


In November of 2016, engineering researchers Xiaolin Wu and Xi Zhang posted an article entitled “Automated Inference on Criminality using Face Images” to a widely used online repository of research papers known as the arXiv. In their article, Wu and Zhang explore the use of machine learning to detect features of the human face that are associated with “criminality”—and they claim to have developed algorithms that can use a simple headshot to distinguish criminals from non-criminals with high accuracy.

In the 19th century, an Italian doctor named Cesare Lombroso studied the anatomy of hundreds of criminals in an effort to develop a scientific theory of criminality. He proposed that criminals were born as such, and that they exhibit both psychological drives and physical features that harken back to what were, in his view, the subhuman beasts of our deep evolutionary past. Lombroso was particularly interested in what could be learned from facial features. In his view, the shape of the jaw, the slope of the forehead, the size of the eyes, and the structure of the ear all contained important clues about a man’s moral composition. None of this turned out to have a sound scientific basis. Lombroso’s theories — many of which wrapped racist ideas of the time in a thin veneer of scientific language — were debunked in the first half of the 20th century and disappeared from the field of criminology which Lombroso helped to found.



In their 2016 paper, Wu and Zhang revisit Lombroso’s program. Essentially, they aim to determine whether advanced machine learning approaches to image processing can reveal subtle cues and patterns that Lombroso and his followers could easily have missed. To test this hypothesis, the authors deploy a variety of machine learning algorithms in a misguided physiognomic effort to determine what features of the human face are associated with “criminality”. Wu and Zhang claim that based on a simple headshot, their programs can distinguish criminal from non-criminal faces with accuracy of 89.5%.