Neural networks created for the field of information technology are modeling the nerve cell system of the human beings. With this idea, parallel data processing and adaptive learning becomes possible, so earlier learning algorithms in artificial intelligence cannot compete with neural networks in certain tasks.
From nerve cell models, like neurons, we can build as huge and complex networks as we want, and these networks can solve different problems effectively, such as recognition tasks.
Besides recognition tasks, neural networks can provide a solution within a reasonable time to such difficult problems with large number of operations, which other algorithms cannot solve, even with coputers reaching high performances.
Convolutional neural networks can recognise different objects, formations and shapes on pictures. Learning algorithms of neural networks probably do not highlight exactly the same details during classification as the nerve system of the human beings do.
We can understand better the behavior of the neural networks and we can specialize the database of a given task by learning about the details of the mechanical neural network’s learning algorithm in each layer.
In my thesis I will reveal and display the deep neural network’s recognition features layer by layer. The pictures I use as a database contains street views mostly, and I would like to show the steps and details of the recognition of pedestrians, as detecting pedestrians on a given picture is an important research area in the projects of the vision of automatic cars.
I will visualise the results in my thesis with some of the pictures from this database of street views, using MATLAB, some functions from MATCONVNET, and VGG-16 pretrained neural network.