Generative fashions be taught to mannequin the distribution of information — basically, they learn to generate new examples that seem like the coaching information.
As an alternative of simply studying to categorise or predict (discriminative modeling), generative fashions purpose to reply:
“What does the information look like, and the way can we create extra of it?”
- GANs (Generative Adversarial Networks): Generate lifelike photographs from random noise.
- VAEs (Variational Autoencoders): Encode and pattern new information from a latent distribution.
- Diffusion Fashions (e.g., Steady Diffusion): Be taught to denoise random noise into lifelike samples.
- Autoregressive Fashions (e.g., GPT): Predict the following token or pixel in a sequence.
These fashions are skilled utilizing solely the information xx — no labels required — making them inherently unsupervised.
Instance: GANs generate lifelike samples by pitting two networks in opposition to one another.
