What Is Machine Bias?

Machines are biased. It's not a secret. Humans program them, and humans are biased, so their machines will be too. But what does that mean? Machine bias comes in two flavors: overfitting and underfitting. Overfitting is when a machine learning algorithm has been trained on a set of data so precisely that it can only make predictions based on that one data set. Underfitting happens when an algorithm needs to be trained better to make accurate predictions on new data or needs more information to make predictions (a common problem in medical research). The root cause of both issues is a human error during the algorithmic tuning process. Algorithmic tuning refers to selecting which hyperparameters will minimize a learning algorithm's loss functions and provide the most accurate outputs. Hyperparameters are machine learning parameters whose value is chosen before the learning algorithm is trained. It's true. The algorithms that power our machines are, in fact, prone to a certain amount of bias and it's not just a matter of human error—it's an intrinsic property of how these algorithms work. When you train a machine learning algorithm using data from the real world, you're basically telling it what to look for. Suppose your training set has a lot of information about race and education, for example. In that case, your algorithm will start to weigh those things more heavily when making predictions about new input. But if your training set doesn't include enough information on race or education, your algorithm won't be able to do its job well—and that could have serious consequences! For example, courts worldwide have started using machine learning software to recommend how long convicted criminals should be incarcerated. Studies have found that when data about a criminal's race and education are weighted too highly in these systems, they can make biased recommendations that lead to unfair sentences for some people who have been convicted of similar crimes.