The extraction feature in the Machine Learning
In classic Machine Learning, when you work on a case such as a prediction of purchase (knowing whether my prospect will buy this product or not), the Data Scientist will make a choice, and extract the data that will influence the prediction: the variables. In this case for our example on the purchase, our variables could be the age of the person, his gender, his income etc..
These are the elements that will allow him to know if a person will buy the product or not. In other words, it is on these variables that the Data Scientist will train his algorithm.
This is called Feature Extraction: selecting the variables that the Data team will work with.
No feature extraction in Deep Learning
In Deep Learning, often it is not possible to do Feature Extraction, and even you don't have to do it.
Why not? In Deep Learning, we often deal with unstructured data: images, sound, text, etc. When you have an image, you won't be able to extract the element that will be the predictor: it is of course impossible for us as humans to manually select all the pixels of each image belonging to the dog or cat (the classic example is to recognize whether a cat or a dog is represented on an image). Therefore, to this extent we do not need Feature Extraction.
The algorithm will be trained to output itself the elements that influence the prediction you want to make.
2 types of algorithms :
--> Convolutional neural networks which are used for example for image processing.
--> We also have recursive neural networks for word processing.
If you want to know more about each type of neural networks, take a look at the Youtube channel "3Blue 1Brown" which will explain in detail these categories of neural networks and how they work.
This is the explanation for the difference between Machine Learning and Deep Learning. On the one hand you will choose the data with which your algorithm will be trained, according to the desired variables. In Deep Learning, you will have to put them in your algorithm in a raw way. But in this case, why not just do Deep Learning?
The limits of Deep Learning
We remember the architecture of a neural network. An input layer, hidden layers and then an output layer. To have a powerful algorithm, you will need a very "deep" algorithm.