What is the difference between the term "generative" in classical machine learning and deep learning?

Question

There are lots of explanations on DGM (Deep Generative Model) and generative classifier (most of the explanations on which are about generative classifier vs discriminative classifier)

But, I can hardly find the common parts between the two concepts. In my understanding, 'generative' from DGM is quite straightforward - it almost goes the same with its literal meaning. In the contrary, 'generative' from the comparisons with the discriminative model is a little bit technical but it's the one that took the word earlier than the former one. (Jordan and Ng, 2002).

Is it just that these two concepts are not really unrelated? Were they just used just because they do produce some distributions while learning?

Answer 1

Generative models have in common that they all model the distribution of the training samples. This makes it possible to sample from a distribution that is (hopefully) similar to the training data distribution. Different architecture types such as energy-based models, variational autoencoders, generative adversarial networks, autoregressive models, normalizing flows, and numerous hybrid approaches can all be considered generative. For more information see this review.

The same meaning is true for generative classifiers, they are called this way because they model how a particular class would generate its associated data. Discriminative classifiers learn to detect features to distinguish between the various possible classes. This visualization from here might be helpful in this regard: