Breno

Generative modeling

Let's study generative modeling.

Suppose we have a function f.

We say f is a generative model if the distribution it produces, known as pushforward distribution, matches a data distribution.

But what is a pushforward distribution?

Suppose you have a simple distribution. Now, push every sample through f. The resulting distribution on the other side is the pushforward.

The goal is to learn an f whose pushforward matches real data.

So, generative modeling is a learning problem. Find the right f.

In machine learning, we can have two types of models: generative and discriminative.

Discriminative models focus on mapping individual samples to their corresponding labels, like a classification model that says if an image portrays a rainy day or a sunny day.

Generative models are about mapping from one distribution to another, like transforming an image portraying a rainy day to a sunny day.

There is a novel method termed Drifting Models that realizes that in one step.