Flow-based generative model

Can you imagine creating a perfect digital copy of anything?

Imagine you want to design a perfect digital replica of a famous painting for an online gallery.

Think of a generative model as a digital artist that learns to mimic the painting's style. It starts with a simple pattern and gradually transforms it into the complex style of the original painting, learning the intricate details along the way.

Example

Starting with a plain square, the model gradually adds strokes and colors until it resembles the famous painting.

Remember this

The model learns the painting's style by transforming a simple pattern into its complex form, capturing the essence of the original artwork.

Text adapted from Wikipedia, licensed under CC BY-SA 4.0.

Related concepts

the reverse process learns: p_θ(x_{t-1}|x_t)

Can we trace back the roots of life?

Generative adversarial network

Ever wondered how computers can create new images that look real?

Langevin dynamics does: adds noise to gradient descent to sample from a distribution

How can we simulate randomness in physics?

Stable Diffusion

Stable Diffusion generates images from text descriptions

Likelihood function

Why can't we just guess probabilities correctly all the time?

Pre-LN transformers are easier to train

Pre-LN transformers use residual connections, allowing gradients to flow more smoothly during backpropagation

Swipe through more Machine Learning concepts

Open Pocket Polymath