What Is a Generative Model?

A generative model is a way of analyzing a set of data such that you can make predictions about new entries in that set of data. This article will explain the basics of generative modeling.

Definition of a Generative Model

Statistics is the science of collecting and analyzing data and often comes with the ultimate goal of making assumptions about that data. Statistical modeling is the process by which you can analyze a set of data and make assumptions about that data.

Creating a generative model is a form of statistical modeling. The defining trait of a generative model is its ability to predict, or generate, new entries in a given set of data. Say, for example, you had a generative model built to analyze a series of numbers: 1, 2, 3, 4, 5. You could use your model to predict the next entry in the series, and the model might tell you that 6' would likely be the next entry in the series.

Of course, statistical modeling isn't done on such small, simplistic sets of data, but the core concept here is that a generative model takes in a dataset and outputs predictions about what future entries in the dataset might look like. Generative models can also tell you how likely it is that a given prediction is to be accurate.

Generative vs. Discriminatory Models

Discriminatory models are another type of statistical model, like generative models. But these models work differently from generative models and serve a different purpose.

The defining trait of a discriminatory model is its ability to discern differences, or discriminate, between entries in a set of data. For example, say you created a discriminatory model to check whether or not a student had passed or failed a test, and say you fed that model all your students' tests from previous school years. This model could tell you if new students who took the test passed or failed.

Once again, accurate discriminatory models deal with larger, more complicated data sets. Still, the core concept here is that a discriminatory model takes in a dataset and can then discriminate between data entries. Discriminatory models can also tell you how likely a given label, i.e., pass or fail, is to be accurate.

Generative Models and AI

When it comes to artificial intelligence, generative models and AI together can quickly and efficiently generate new content based on analyzing many examples of similar content. NVIDIA defines generative AI models as follows: "Generative AI models use neural networks to identify the patterns and structures within existing data to generate new and original content."

For example, a generative AI model can analyze images of human faces, identify the patterns and characteristics of human faces, and then use the patterns and characteristics of human faces to generate a totally new example of a human face.

This is the core concept behind AI-generated images. Complicated generative AI models are trained to 'learn' what a thing is by analyzing many examples of that thing and then using that information to generate new examples of the thing it hasn't seen before.

Popular Applications of Generative Models

Generative models, discriminatory models, and AI technology, like neural networks, work together to form Generative Adversarial Networks (GANs). GANs can analyze a given set of data, generate new entries in that set of data, and check how likely it is for a generated data entry to fit in with the original data set.

When you pit generative models against discriminatory models against each other in a GAN, the GAN can 'train' itself to produce better and better results. A generative AI model will produce something new, while a discriminatory AI model will 'check' if what's produced is good enough. If not, the generative AI model will keep trying until it passes the discriminatory AI model's check.

Take any of the sites online you can use to generate an AI image for yourself, like Hotpot, for example. This kind of service is likely the result of a GAN. Say you want to generate an AI image of a man writing an article. Behind the scenes, a neural network can be trained to have an 'idea' of what a man writing an article looks like; a generative AI model can produce an example of a man writing an article. In contrast, a discriminatory AI model can check to see if that image measures up against real images of men writing articles. The result is an AI image of a man writing an article.