I’ve been filling them out now for nearly 10 years, and they’re always the same! If someone can explain to me how this can be, I’d be very grateful <3
In: 321
Now that you talk about it, I remember very différents phases.
Early captchas asked to type words,
Then I remember a lot of stop sign or trafic light
Then it was motorcycle,
Bus,
Trucks
Still about self driving car probably but different items
You can’t make them any harder than human beings can detect.
Imagine using Where’s Waldo? type images. There are certainly people who could quickly find Waldo in every image. But most people couldn’t. Most people would get the answers wrong – and trying to train your AI based on mostly wrong answers isn’t going to result in a particularly good AI.
They have changed over the years.
The first batch was things like which is a bus, and they would have you pick between bus and bike and cows.
Then the images where the entire photo and you had to do regions that had the crosswalk but they covered multiple cells.
Now there are captchas that are trying to tell the difference between similar looking objects like bus and train.
And some captchas are to check the work of the ai. So you have two images that are confirmed bikes, and somewhere in the set is the one the ai thinks it’s a bike. These tend to be the ones that replace one cell with a couple more one after the other till it is no longer a “bike” before you verify.
I think a lot of the word ones were dumped or shifted to less well known sites due to places like 4chan figuring out what parts were confirmed and replacing the unknown words with racial slurs.
I’d just like to understand the mechanics of this. I pick out all the crosswalks or whatever. The captcha has to know which squares are the “correct” answers already, to evaluate if I am a robot. So that data could already be in the AI database, without me ever being included. So how does my input help the AI in any way?
They have changed plenty.
Dog, Banana, Sedan please pick the car.
Then it was identify every picture with a car.
Now many of them are like pick every square in this picture that contains a traffic light.
We have advanced from teaching them pictures of things to teaching them how much of a picture is that thing.
Also don’t forget, the teaching AI is a side benefit. The primary purpose is actually keeping out bots. So even simpler ones are fine for that.
it really only seems like it hasn’t changed to you because a stop sign at almost any angle you see it is still a recognizable stop sign… to a computer that’s not the same. a stop sign at a 45* angle to the right is not the same as a stop sign at a 30* angle to the right.
it also takes massive amounts of data for machine learning to be accurate. there are obviously some percentage of identifications that are wrong and it requires enough detections to know for sure. also, ideally machine learning is in constant refinement increasing its accuracy with every more data.
Autonomous vehicle related data acquisition is so full of middle men, it already turned into selffullfilling job where having it completed isn’t a target. And as in every middleman infested enterprise, there is no way that news of nonsense and mistakes found on lowest levels ever reach the right ear. Parameters are copy pasted from one project to another without a hint of common sense. For example you are supposed to ignore lidar data that show the height above 180cm even though the car used in a project is originally 182cm tall and there is a million dollar lidar toy mounted on its roof.
When training AI’s, quantity *is* complexity. The more different photos of traffic lights, cars, crosswalks, etc can be identified, the better the AI is. You want the machine to be correct in every time of day, every street corner, every weather situation, and every angle. Even when the object is partially obstructed, or in shadow, or with glare, which just requires more and more samples to be identified.
That and, I believe they have gotten more sophisticated, and ask more specific questions now than ten years ago. But then, I don’t have a comprehensive list of every captcha for the last decade.
I have a very very vague memory of a comp sci lecturer saying that it wasn’t actually so much to do with getting it right but more to do with how you came to getting it right. It doesn’t matter if the robot gets it right, it looks at the way your mouse moves, speed of answer and the pattern of selection. If you’re too rigid/consistent, you’re a robot. Pretty sure that’s how the “tick box if you’re not a robot” ones work too, just mouse tracking to see how imprecise/inconsistent you are.
Would be great if someone could confirm or deny if this is correct? If it is, that’s why it doesn’t matter how hard the “puzzles” are. I suppose it’s really hard for a robot to act organically messy? Maybe we’re not training the ai to find cars but more to identify them in an innately human way
Latest Answers