Comment Bayesian statistics (Score 5, Interesting) 375
I work for a company that develops neural network software which is used for face recognition on a number of airports. The problem we've had over and over again is that government officials and airport security personnel have great difficulty understanding some elementary statistics.
Let me give you an example. One version of the software offers 99.99% accuracy (symmetrical true positive and true negative), a number that always seems very impressive to various officials.
What they don't understand and what we have to remind them all the time is that they need to take into account the large number of faces that are scanned by the software and that a 0.01% false positive rate isn't something you can ignore.
For instance in a large airport that has say a million people getting scanned yearly it means that 100 people will be incorrectly flagged by the system. The prior probability that a traveler is a 'person of interest' is less than 1/100,000. Plugging the number into Bayes' theorem you get that when the system flags a passenger, the probability that the passenger was actually a person of interest is around 9%.
The officials typically only listen to the 99.99% figure and ignore the reality of the relatively large numbers of false positives when dealing with huge numbers of people. Subsequently they treat the people the systems flag much worse than they would if they realized that the probability of a 'catch' being correct was less than 10%. We've done our best to try to educate them but usually they don't want to listen as it's an uncomfortable truth and it's much more convenient to say that the system has an accuracy of 99.99%.