Ask HN: How does Google get data from ReCAPTCHA yet it knows if I'm wrong?
If it asks me to label all street signs, the common understanding is that I'm labeling data for their ML algorithms.
But if they're counting on me to label it, then how does it already know if I miss a tile? It seems to already have the data labeled, so what's the point?
4 comments
[ 5.4 ms ] story [ 21.6 ms ] threadIf you get a lot of the known data wrong, they throw out your answer and make you answer the CAPTCHA again. If you get the known data right, they assume that your answers on the unknown data are correct and use that to label them, then start adding them in as the known data for other people. They might also show the same images to multiple people and see if they give the same answers - if so, it's probably correct and can be labeled, if not throw them out and make them answer the CAPTCHA again.
They probably require some minimum level of consensus to determine knowns as well.
For instance NYPL’s “building inspector” crowdsourcing program accepts a piece of data as verified if it’s been checked by 3 or more participants and at least 75% of them are in agreement. If 75% hasn’t been reached, it will keep showing it to more participants until it is.
For the images I assume its something similar.