3 ms·
> As DeepSolar learned to identify the characteristics of panels in the images, the system was eventually able to correctly identify that an image contained sol
by daeken 8y ago
> As DeepSolar learned to identify the characteristics of panels in the images, the system was eventually able to correctly identify that an image contained solar panels 93 percent of the time. About 10 percent of the time, the system missed that an image had solar installations.
How can these two facts both be true? If it fails to identify that an image contains solar panels 10% of the time, then it can't possibly correctly identify that an image contains solar panels greater than 90% of the time, and that would require that it has a 0% false positive rate. I can only assume they mean that it has a 7% false positive rate and a 10% false negative rate, but that is ... not what the first statement says at all.
- desdiv 8y agoGoing by this graph[0], what they meant was: we had a precision of 93% and a recall of 90%. [0] http://web.stanford.edu/group/deepsolar/assets2/img/roc.jpg http://web.stanford.edu/group/deepsolar/assets2/img/roc.jpg
- floatrock 8y agoWikipedia has a good illustration showing the relation between precision & recall vs. false positives & false negatives: https://en.wikipedia.org/wiki/Precision_and_recall https://en.wikipedia.org/wiki/Precision_and_recall Precision is what percent of the identified-positives are actually positives: true-positives / (true-positives + false-positives). Recall is what percent of the actual-positives were identified as positive: true-positives / (true-positives + false-negatives). These are useful summary stats over the true-false/positive-negative matrix because, to quote the wikipedia: "In simple terms, high precision means that an algorithm returned substantially more relevant results than irrelevant ones, while high recall means that an algorithm returned most of the relevant results."