Why do some websites need you to identify trucks to prove you’re human when machine learning can easily allow computers to do so?
In: 190
The details are kept secret but whilst you are completing those “human tests” they are also looking at your cookies and website history to see if you browse the web as a human would. The image test is more of a safety check.
I think it’s funny that you think you’re not actually teaching the computers machine learning when you select that truck or traffic light.
Because the machine can’t check if its correct to learn as it goes.
Also because your answer helps to teach AI how to think for itself, so that eventually the machine can solve it’s own questions.
Why do websites use them though. I mean, if a ‘robot’ can pass the test.
If I had a security guard at my shop who would let thieves in I would get a different security guard
Best answer is that the captcha is actually monitoring other things to see if you act like a human and your answers are being used to teach image recognition systems. The images are actually not the first implementation of this idea. The inventor of the original captcha (where they gave you a squiggly word you had to type out) hated that he was wasting a lot of peoples time. So he updated the system to have 2 words. 1 the system knew and the other it didn’t. It used this information to learn text recognition. He also created Duolingo and used the same idea to translate webpages on the internet. He spoke about it in a Ted talk.
How do you think the machines learned to do it?
Also, it’s not like there’s some human checking our results. The damn machines are the one’s deciding if we’re human enough to click on each image containing a traffic light.
The key here is that Machine Learning CAN’T easily predict which images are trucks. That is why you’re labeling the blurry image FOR the machine.
So say, the ML model v1 already learned from 100 images of trucks vs. non trucks to get a prediction accuracy of 70% (for example). To make a better ML model, we take these human labeled truck images (ground truth), add them to the old training dataset, so now the ML model
v2 learns from 1000 images of trucks vs. non trucks to get a prediction accuracy of 90% (for example).
The key here is that the accuracy will never reach 100%, meaning the ML can’t EASILY predict these blurry images. It can get pretty close (say 99%) if you use billions of data points to train the model, hence why they’re asking us to label the blurry images as the entry cost for the website.
Extra steps to deter bots. The whole point of using bots is efficiency. If you add extra steps its gonna need more resources to use bots thus deterring them.
They aren’t looking at your answer specifically to tell if you’re a robot. They’re looking at the speed of your mouse, and where you move it.
A machine moves their cursor in a way that no normal person would. So, they record the speed, and direction, to make sure your mouse isn’t teleporting across the screen.
Fun fact tho, these tests are actually why machines are so good at recognizing pictures in the first place! They record your CAPTCHA answers and use them as test answers for machine learning.
Latest Answers