Machines can learn. But what do they learn from?
An interesting article in TechCrunch called “Machine learning and human bias: an uneasy pair”, prompted me to look into a number of police initiatives in several US cities involving the use of machine learning tools to help predict certain crimes, put some people on “hot lists” or estimate the frequency of patrols required in different parts of the city. The film “Minority Report” immediately comes to mind, but the truth is that such schemes have been underway for some time already.
A 2013 article in the Chicago Tribune discusses how the city’s police department is using analytical tools to establish which people are more like to be involved in violent crimes, and then sends officers to notify them that they are on a list, hoping that this knowing they are under suspicion will keep them on the straight and narrow. Such an approach would be illegal in many countries, but US law permits such eventualities within certain guidelines.
Oxford, Alabama, is using a machine learning application that divides the city into 150-square-meter blocks and tries to predict which areas are most likely to see crime, supposedly allowing police officers to be more efficiently assigned to trouble spots. The app, called Predictive Policing Software, or PredPol, uses only data that is in the public domain about the type of crime, the place, and the hour it took place, without using any personal data from police files.
What can a machine learn? Until the so-called internet of things fills the world with sensors that give us eyes, ears, and other senses, machines can only learn from the data we put into their algorithms, data that is obviously subject to the ideas of the people who input it.
Let’s not write off technologies that can be very useful, but we should be aware that when we input personal data of any kind, there is a clear risk that incorporating questions into the learning process such as place of residence, racial profile, or religion will be linked to stereotypes based on experience. Transparency when it comes to creating these algorithms or deciding which data should be used suddenly becomes a key factor.
Fairness, Accountability and Transparency (FAT) must be the guiding principles when creating machine learning algorithms, and these principles will become increasingly important as we come to rely on machines that learn.
(En español, aquí)
