denoising score matching does: learns to denoise, which equals learning the score

Can you clean up a noisy picture?

Imagine you're looking at a blurry photo of your friend, but it's so noisy (unwanted random stuff) that you can't see their face clearly. You want to fix it so you can recognize them.

Denoising score matching is like cleaning up that noisy photo by learning what the clean picture should look like. It's a smart way to guess the missing parts by looking at the patterns in the noisy parts.

Example

If the noisy photo has a pattern of random dots, denoising score matching learns to guess what your friend's face should look like by noticing the patterns in those dots.

Remember this

It's like teaching a computer to guess the missing parts of a noisy photo by learning from the patterns in the noise.

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

Related concepts

score matching does: learns the gradient of the log-density without normalizing

Ever wonder how we can compare apples and oranges fairly in studies?

Stable Diffusion

Stable Diffusion generates images from text descriptions

Convolutional neural network

Can a neural network learn too well?

DDPM stands for: Denoising Diffusion Probabilistic Model

Can you imagine creating perfect photos from scratch?

classifier-free guidance does: interpolates between conditional and unconditional generation

"Classifies samples as either conditioned or unconditioned, guiding generation towards desired outcomes."

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

Can we trace back the roots of life?

Swipe through 100 ML concepts daily

Open Pocket Polymath