3 ms·

If you think about it a bit, it should be obvious that nothing is being trained on your results. You're not being exploited for training data, because computers

by jsnell 6y ago

If you think about it a bit, it should be obvious that nothing is being trained on your results. You're not being exploited for training data, because computers are already better at identifying traffic lights from street view images than you are. The image recognition task has not changed for a very long time. How many captchas do you think have been solved by users in the last 5 years, all the same kind of recognition problems on the same kind of data? Trillions? It would be way past any kind of diminishing returns for this task to get more labels. If there was value there, they would have at least switched to a different kind of task. In addition to that they're trying to get sites to move to Recaptcha v3, which does not have the user solve any kind of. Why would that be if the answers had any kind of value? The fundamental problem with using captchas for training is that it devalues their value as a tool for security, since it sets up conflicting incentives. And the latter is what Google is selling. (Yes, selling. Recaptcha Enterprise costs money to use at scale, and the list price seems to be $1 / 1000 calls). The people paying for that service are using it to block attackers and let good users through, not to challenge the good users just to get some more labeling data from them.

grundoon 6y ago
"computers are already better at identifying traffic lights from street view images than you are" -- I'm still better than they are at identifying traffic lights IRL though :-)
tmottabr 6y ago
first, computers are not better at identifying those things than us, our brain evolved for that and it is a massively complex problem for a computer. sure we have models that can do it running on server on static images, but the problem they are trying to solve is not street view, the problem is training ML models capable of doing this in real time in an autonomous car. taking a step back, identifying things in images is a problem very complex for a computer, best way in this day and age is to use ML but you need to train it to do it, for that you use a bank of know images and that is what they use reCaptcha for, to built those image banks that they will later use to train and evaluate ML models. Second, they can use reCaptcha for that and still provide all those things that you mentioned. another thing, they are not continuously using the same set trillions of times, they are increasing the image set over time, and this is what they use reCaptcha for to classify those new images, they dump new images in the set and every time you see a reCaptcha it will show you a few know images and a new unknown one that need to be classified. Your input alone say nothing about that image, but after hundreds of peoples saying that, for example, there is a traffic light in a image while getting all the other know traffic light images right indicate there is a big likelihood of that picture containing a traffic light, once they have a big confidence on the image classification after hundreds or even thousands of captcha they move it to the know images set.

by jsnell 6y ago

grundoon 6y ago

"computers are already better at identifying traffic lights from street view images than you are" -- I'm still better than they are at identifying traffic lights IRL though :-)

tmottabr 6y ago

first, computers are not better at identifying those things than us, our brain evolved for that and it is a massively complex problem for a computer. sure we have models that can do it running on server on static images, but the problem they are trying to solve is not street view, the problem is training ML models capable of doing this in real time in an autonomous car. taking a step back, identifying things in images is a problem very complex for a computer, best way in this day and age is to use ML but you need to train it to do it, for that you use a bank of know images and that is what they use reCaptcha for, to built those image banks that they will later use to train and evaluate ML models. Second, they can use reCaptcha for that and still provide all those things that you mentioned. another thing, they are not continuously using the same set trillions of times, they are increasing the image set over time, and this is what they use reCaptcha for to classify those new images, they dump new images in the set and every time you see a reCaptcha it will show you a few know images and a new unknown one that need to be classified. Your input alone say nothing about that image, but after hundreds of peoples saying that, for example, there is a traffic light in a image while getting all the other know traffic light images right indicate there is a big likelihood of that picture containing a traffic light, once they have a big confidence on the image classification after hundreds or even thousands of captcha they move it to the know images set.