What is the basic idea behind neural networks and especially deep learning?
The biological cell in the brain is represented by a mathematical or artificial cell called (node). This cell receives stimuli from the senses and processes them. We simulate that in the artificial cell by inserting features and represent the strength of the links with values called weights that we change during the training process, the center of the biological cell It processes the input information and the same thing in the artificial cell, after calculating the sum of the input product in weights + a fixed number, then in the center of the artificial cell we apply the activation function to calculate the output. What then ? If the essence of the whole process is in the activation function, if we choose a linear activation function, whatever we install from some functions on some and I mean to add internal layers to the network, the ability of this network remains limited to dealing with data that follows a linear pattern, for example: f1 (x) = 5x +3 f2 (x) = 2x +1 G (x) = f2 (f1 (x)) = 2 (5x + 3) +1 = 10x +7 In the event that we use a nonlinear activation function such as (sigmoid) s (x) = 1 / (1+ Exp (-x)) The more we combine it with a new second function, the more we get a more complex and complex complex function, so what do you think when we compound a hundred functions of this type together? 1000 functions of this type? So in the end, the network will be able to deal with data, regardless of its complexity and overlap. What is neural networks and deep learning is in the synthesis of complex and nonlinear functions in order to obtain a very large and very complex final function capable of dealing with complex data and in several axes and able to find a complex hyperplane equation separating the various components of the data Used or categories. This installation is not so simple as it may result in other problems that make the network performance poor, regardless of its size.