What is a generative model in deep learning?

A generative model in deep learning is a model that learns the underlying patterns or probability distribution of training data so it can create new data that resembles what it has learned.

Instead of only assigning labels or making predictions, a generative model can produce new examples.

What can generative models create?

Depending on the training data, generative models can produce:

• Images
• Text
• Audio
• Video
• Synthetic data
• Other structured outputs

How is a generative model different from a discriminative model?

A discriminative model mainly learns how to predict a target from an input.

For example:

Image → Cat or Dog

A generative model learns enough about the data distribution to generate new samples:

Learn image distribution → Generate a new image

Common types of generative models

Deep learning includes several important generative-model families, such as:

• Generative Adversarial Networks (GANs)
• Variational Autoencoders (VAEs)
• Diffusion models
• Autoregressive models

Each family generates data differently.

How are generative models trained?

The exact training objective depends on the architecture, but the general goal is to learn the structure and statistical patterns of the training data well enough to produce realistic new examples.

For example, a model trained on many face images may learn patterns related to facial structure, lighting, and texture, then use those learned patterns to generate new faces.

The next articles cover major generative approaches in more detail, including “What is a generative adversarial network (GAN)?”, “What is a variational autoencoder (VAE)?”, and “What is a diffusion model?”.

In short, a generative model learns the structure of data so it can create new samples that resemble the data it was trained on.