Recognize a picture or object
The recognition of images, that is to say, the identification of objects, animals, or any other element of the photo, is the cognitive task that showed the power of a major tool of artificial intelligence (AI) : deep learning. In 2012, during a challenge it using the 150 000 images from the image database ImageNet, a team from the university of Toronto (Canada), led by Geoffrey Hinton’s party (then at Google), commits approximately 15 % of errors, i.e. two times less than its competitors, and two times better than in 2010. At the last contest in June 2017, the error rate fell to 2 %.
These techniques are very efficient algorithms with tens of millions of adjustable parameters. Like a painter mixing several colors to get the shade he wants, the computer system looks for the settings that may help recognize the proper items in a picture. To achieve this, he trained on images annotated previously by humans, indicating for example the presence of a Persian cat, a panda, an orc… the structure of The program looks like a network of neurons, whose connections strengthen or weaken depending on the stimuli received.
Since the victory of the team Hinton in 2012, the concept of deep learning has spread to the point of being confused with the concept of artificial intelligence, which encompasses yet other topics, such as robotics. In the visual domain, it helps in the identification of objects, including videos, which allows for example to correctly caption the subject matter of the photos automatically, or to feed the pilot software of autonomous cars in order to distinguish the nature of the obstacles.
The health sector uses this kind of technique to help and speed up the diagnostics. Several studies have even shown the superiority of the machine on professionals to identify…