A 99% false positive rate is not the same thing as a 99% failure rate though. You’d need to look at the false negative rate too. Since the goal is to find weapons and prevent school shootings, the false negative rate would be the actual failure rate. Like the TSA when they have undercovers try to bring weapons through, and 95% of them make it through.
A 99% false positive rate isn’t good, but it’s not a huge problem if you also have a false negative rate of 0%. It just means the model is too overzealous. The problem would be that eventually those responsible for reviewing and reporting the hits would start ignoring them.
If it were a 99% false positive rate and a high false negative rate, then yeah the whole thing is useless.
The comment you are replying to is correct in regards to statistics and evaluation. There is more too it that “crying wolf”.
False positive AND false negatives is the true measure of failure. True negative and true positive are the opposite measures.
These four can be optimized for depending on the consequences of being incorrectly positive or negative. The consequences for negatively identifying an object incorrectly can outweigh the consequences of incorrectly identifying an object positively.
Depending on that breakdown, a model might be optimized to falsely identify a bag of chips as a gun “just to be safe”. The consequences of missing a weapon might have outweighed the consequences of being incorrect. Crying wolf might be the desired outcome that the school or company wanted because they deemed the risk acceptable.
That said, this is still stupid and police should not have come out after the alter was cancelled and CERTAINLY should not have behaved that way. 99% is way too high of a false positive rate
ACAB, but I like statistics and wanted to explain it.
A 99% false positive rate is not the same thing as a 99% failure rate though. You’d need to look at the false negative rate too. Since the goal is to find weapons and prevent school shootings, the false negative rate would be the actual failure rate. Like the TSA when they have undercovers try to bring weapons through, and 95% of them make it through.
A 99% false positive rate isn’t good, but it’s not a huge problem if you also have a false negative rate of 0%. It just means the model is too overzealous. The problem would be that eventually those responsible for reviewing and reporting the hits would start ignoring them.
If it were a 99% false positive rate and a high false negative rate, then yeah the whole thing is useless.
You ever heard of this story about a little boy who cried wolf?
I literally addressed that.
Which you prefaced with:
It is a huge problem when it’s a purpose built system to identify firearms in a school and it’s ineffective at its only job.
The comment you are replying to is correct in regards to statistics and evaluation. There is more too it that “crying wolf”.
False positive AND false negatives is the true measure of failure. True negative and true positive are the opposite measures.
These four can be optimized for depending on the consequences of being incorrectly positive or negative. The consequences for negatively identifying an object incorrectly can outweigh the consequences of incorrectly identifying an object positively.
Depending on that breakdown, a model might be optimized to falsely identify a bag of chips as a gun “just to be safe”. The consequences of missing a weapon might have outweighed the consequences of being incorrect. Crying wolf might be the desired outcome that the school or company wanted because they deemed the risk acceptable.
That said, this is still stupid and police should not have come out after the alter was cancelled and CERTAINLY should not have behaved that way. 99% is way too high of a false positive rate
ACAB, but I like statistics and wanted to explain it.
A 99% false positive rate is a broken system. Mitigating a broken use of shoehorned “AI” by saying it’s not good but not that bad is ridiculous.