Tinker300 dots of pure noise.
Chapter 3 of 7
A learned denoiser
The toy denoiser in chapter 1 needs every training drawing to hand at every step. Real models learn the job once instead.
This chapter: What does a network learn, if it keeps no pictures?
About this chapter
You will learn
- How a network is trained to guess the noise
- Why it learns directions, not a list of points
- How its guesses pull noise onto a shape
Words you’ll meet
- Network
- A small neural network: numbers (weights) that turn an input into a guess, tuned by training.
- Training
- Showing the network noisy points again and again, and nudging its weights toward the right answer each time.
Step 1 of 3
The tray holds 300 dots of pure noise. A small was trained on points from the hinge’s outline.
Predict first.