Blog 3 - Signal Detection Theory

Recently in my Human Machine System Design course, we were learning about signal detection theory. In class we ran experiments with 60 short samples to gauge how good we were at detecting motion in a specific direction. Each clip was only shown for a second or less. Needless to say I was not very efficient at detecting the direction correctly.

After class, I began to think about how and it ways could signal detection theory be used in daily life. Once I got home, I did what most people do when they have a question about something, I Googled it.

One way in which this can be used is in a police line-up or trying to identify a criminal from a survelance video.

By using signal detection theory along with machine learning, authorities can more accurately determine the guilty perpetrator in a crime. In recent studies, it was determined that eye-witnesses choosing criminals out of a line-up was only accurate 1/6 or approximately 17 percent of the time. These numbers are terrifying. It really makes you think about how many issues there are with the criminal justice system and how many “innocent” people have been convicted of crimes they didn’t commit. By using Signal Detection Theory and machine learning, one study showed that the accuracy of correctly identifying criminals increased to nearly 80 percent. While 80 percent still not perfect, its a lot better than 17 percent.

The term Signal Detection Theory is slightly misleading. It shouldn’t be thought of as a theory that can be proven or disproven, but more so of a way to think about data. However, because they’re are using four (4) different variables, it becomes difficult to visually represent the data collected. The four (4) different variables include: hits, misses, correct rejections and false alarms.
One data visualization method is performed by using a Receiving Operator Characteristics Curve or ROC. What the ROC does is graph the probability of a hits vs the probability of a false alarm.

Using Signal Detection Theory allows us to get a better look at the data for real life issues like correctly identifying criminals.