There are basically three components of healthy surfing on the Internet; a healthy connection, a healthy website, and a healthy device. There are many security measures taken to ensure that a site can function properly, not be attacked or damaged by malicious people. Captcha tests are perhaps the most well-known of these.
Recently, we are constantly trying to recognize pedestrian crossings, traffic lights or vehicles in Captcha tests. So why do we constantly struggle with traffic as if we are in a driving school when there are thousands of different questions to ask?
Computers are trying to decipher the traffic because of us.
What we call the captcha test is a simple Turing test used to distinguish between computers and humans. The Turing test is a kind of test that is easy for people but cannot be solved by the computer. The reason why we often see images related to traffic in captcha tests is that these images are difficult to recognize by computers.
Computers do not perceive images the way we do. Computers cannot recognize objects such as traffic lights, pedestrian crossings, bicycles and buses as soon as they see them. The shape, color, location and background of these objects can be complex for computers.
For example, while it may seem easy to tell that a traffic light is green, the brightness, angle of the light and other objects around it can cause the computer to stumble. Therefore, while people are asked to separate these images, computers or software are prevented from entering the sites.
How do computers learn to distinguish things?
Thanks to artificial intelligence, computers can learn to separate things from other things. Although the learning processes are clear, we do not yet know exactly how the head of artificial intelligence works after learning, a complex artificial network is involved.
First of all, unless we make a bot or artificial intelligence for ourselves, the information we can obtain is limited because artificial intelligence and bots are very valuable and tightly guarded trade secrets for companies. They do not give any details other than very general information. Still, the general approach is more or less clear.
Let’s take a look at the neural network systems that have been popular in recent years.
Basically, two bots are produced for the neural network systems that have been popular in recent years. The first of these bots produces new bots, and the second tests new bots. In order to explain the rest of the article without confusing, I will call them producer bots, trainer bots and student bots.
The producer bot produces student bots and sends them to the trainer bot. The instructor bot doesn’t know the distinction we want to make, but he has a quiz of answered questions in his hand. The student applies this test to the bots. It sends the students back to the manufacturer with the test results.
When the children return from school, the productive bot brushes aside those who do well and destroys the others; Instead, another student builds bots based on successful examples. The process of making and testing new bots continues as a while loop.
While the students who were lucky at first survive, at some point a bot emerges that can do the desired job more or less not by chance, but thanks to its structure. In subsequent iterations from that bot, the success rate required to survive gets progressively higher.
In the end, we get an artificial intelligence bot that works successfully, although we do not know exactly how it works. We humans create the answer keys for the tests with millions of questions that will be applied to the billions of bots that will take these tests with the Captcha tests.
Computers, of course, have places to use what they learn.
The information from the captcha tests is used in the development of services such as Google Maps in the first stage. Google tries to keep its maps more accurate and up-to-date by matching the traffic-related images it uses in its Captcha tests with responses from people. Thus, Captcha tests are used both to secure websites and to improve mapping services.
Another point is, of course, autonomous driving systems. The development of these systems requires a huge amount of data. One of the most effective ways to collect such big data is to create control data already by asking billions of people browsing the internet questions about images.
So at the end of the day, people struggle with that many traffic symbols to train bots in Captcha tests. So in the future we will have autonomous cars that follow more accurate maps.
43995
