Neural Networks are getting so popular due to their ability to create any function by feature learning when enough data is provided. Features are the information you are giving to the network, greater the feature size greater is the information you provide. They are primarily used to solve classification problems but research is still being done to make them work for regression problems as well.
Neural networks are analogous to the neurons in the brain of a Human. When asked to solve a problem, humans often break the problem into smaller, easier-to-solve sub-problems at different level of representation. Humans are able to inadvertently exploit intuition and describe concepts in hierarchical ways, based on levels of abstraction. Let us say an individual is asked to recognize a face, he’ll then look over to some specific features over the entire image. The individual sees a human form in the image, notices facial hair, body structure, and clothing and then determined that the image is a face of a man. The problem is broken down on many levels without much conscious thought. Humans look at much smaller features of images, such as lines, curves, and edges to determine higher-level features. These numerous highly-varying, non-linear features organized into layers are what constitute a deep network.
To get a more better intuition of what is happening here let us do a task-
Imagine a human face.
You probably would have conceived a rough figure in your head which is similar to any generic human face. This is exactly what the network has learned after it is trained over face data.
Imagine the face of your best friend.
Now you possibly would have conceived a clean image of a face. Here face of your Best friend was the input and “best friend” acted as a label or target class for your brain network.
Neural network is referred as Deep Artificial Neural Network when the number of hidden layers go beyond three. In this situation new learning techniques will come in the picture, reason mentioned here. The hidden layers are composed of hidden units that can be used to describe underlying features of the data. In facial recognition task the input layer represents the pixels of the image while the output is the corresponding identity(or classification) of the face, while the hidden layers can represent low-level features, such as edges and shapes, to high-level features such as “big-eyes” or “small- nose”.
Learning the structure of a deep architecture aims to automatically discover these abstractions.