The deep learning approach avoids the need for the programmer to define all the knowledge necessary for the computer to solve problems. The hierarchy of solution concepts makes it possible for the computer to form complex solution approaches from simpler approaches. A graph that maps these concepts to one another consists of many layers and is therefore “deep”. This is why this artificial intelligence approach is called “deep learning”. In the following, deep learning techniques and the procedure for image recognition in deep networks are explained.
It is only through the development of convolutional neural networks (CNN) and the use of graphics processors (GPGPU – General purpose computing on graphics processing units) to parallelize the calculations in machine learning that outstanding results can also be achieved in significantly more complex tasks.
CNN are based on the structure of the visual cortex of animals. complex arrangements of neurons react to edge-like patterns. While MLP consists of fully connected layers, the arrangement of the edges in CNN depends on the defined convolutions.
The combination of several convolutional and fully connected layers to form a deep convolutional neural network has proven itself, for example, in the recognition and classification of images.