In the year 2018, driving autonomous has accelerated the pace. In the united States, dozens of companies have taken to the roads, a barrage of vehicles in test. One of the consequences has been the increase in accidents where they concur these cars. Some, such as the outrage, starring a car from Uber or the shock of a Tesla in function Autopilot, have produced fatalities.
technology is in a process of refinement, that these tests contribute. One of the aspects that have that polished are the databases. And within these there is a factor of enormous importance. It is the classification or labelling of the images.
The labeling belongs to the bowels techniques of the driving systems employed. You can compare it with the cognitive ability should have any vehicle that you want to be guided by itself.
of course, the labeling of images is one of the keys of the vehicle that can’t fail. The specialist in deep learning Lucas Garcia, of the company MathWorks, which develops analytical software for autonomous cars, stressed the importance of this factor in a conference during the event Big Data Spain. In conversation with THE COUNTRY, this mathematician, who has also been a researcher at the Complutense University of Madrid, summarized the issue: “A wrong labeling of the data can result in an algorithm that is not able to correctly solve the problems.”
A sample of how you perceive a car self-contained the objects of his environment. 2018 The MathWorks Inc
This coupled with a tricky situation on the road makes the vehicle more prone to an accident. If the cameras of an autonomous car are its eyes, the way that knows the reality, the labeling of the database is their cognitive ability. By comparison with the database, the vehicle understands how to is your environment.
in order For the car to make good decisions you need two circumstances basic. “If we want to create an algorithm that detects very well pedestrians, cyclists and other vehicles, previously we have done a process of collecting data and of signals provided by Mariobet the sensors,” stresses Garcia, who adds immediately: “And we also have to label them correctly.”
to Label objects in an image is a work more fine than it may seem. The classification is transmitted to the algorithm of deep learning or neural network, which takes the decisions of the autonomous car. Therefore, the information must be as accurate as possible. “A possibility of labelling is that in a picture where appears a car let’s draw a rectangle on top to identify that it is the car,” says the mathematician. “Another would be to say exactly which pixels of the image corresponds to the car”.
The relationship is clear: the more precise the labeling the better the algorithms that will nurture this. As the whole system based on artificial intelligence, the autonomous car is probabilistic. “The analysis that carry out all of these models of deep learning is based on a probability,” he says. “Despite the fact that the systems try to be very robust, if the model of machine learning expected with a probability of 99.95% that is ahead is a car, obviously there’s a 0,05% chance that it will be another thing.”
All of these models of artificial intelligence have their failure rates. Are design errors that evidently the engineers try to reduce. To do this there is no alternative but to devote time and people to a specialist to do the labeling. There are solutions, as in the works of García from the MathWorks to automate the process, but there must always be a human behind that validates the results.
The neural networks that use autonomous cars require millions of tagged images to work. It is a formidable task, with a manual component inevitable. And this only with respect to the images. But there are more sensors to complement the camera, such as radar, the proximity detector or lidar. This last much more costly to correctly label.
