Discriminative model & Generative Model

Machine learning can be classified into supervised learning and unsupervised learning. In a simple way, the former can be seemed as a classifying task, whose object is to train the classifier according to the data set, while the latter has no information about the categories of the whole data set, which means that the task is to cluster.
Generally, in the supervised learning, we often divide the whole data set into 2 parts: the training data and test data. The key point is to train a model on the basis of the training data, and the test data will be used to test the accuracy of the model.
In a general way, there are two functions used as the model：
***decision function : Y=f(X)
conditional probability distribution: P(Y|X)*
In the decision function, there is a threshold value, and the classification of X is based on the comparison of X and the threshold. While, in the conditional probability distribution, the classification is based on the value of probability.
Then, we call the two kinds of methods to obtain the model as the discriminative approach and generative approach, and the output is discriminative model and Generative Model, separately. Generative model implicitly or explicitly represent a data distribution, often allowing for new data to be sampled or “generated” through a process, hence their name. Discriminative model, on the other hand, are less ambitious. Rather than modeling distributions, they attempt to model class boundaries.
In the supervised learning setup, where we have data X and
labels Y, discriminative model is used for primarily prediction tasks, and it learn the conditional P(Y|X) or the decision function Y=f(X). In contrast,generative model learn the joint probability P(X ,Y) first, except that, it also need to learn P(X). After that , it can get the conditional P(Y|X) to classify.
In the last, we need to talk about the strengths and weaknesses of these two models. The discriminative approach is simplified, which deals with the data directly, and it can abstract the data or define the feature as casually as possible. However, it can not reveal the inherent characteristics of the data, and it aims to find the boundary of different kinds.
While, generative approach aims to express the distribution of data from a statistical point of view, and it can reveal the similarity of data itself. In addition, we can get the discriminative model from the generative model, vice versa. And if the data is enough, the generative model is faster to converge. The most important, it can be applied into the problems consisting hidden variables.

Reference:

[1]: D Ramachandram, GW Taylor. Deep Multimodal Learning: A Survey on Recent Advances and Trends. IEEE Signal Processing Magazine [J] , 2017 , 34 (6) :96-108.