Tinker300 dots of noise; the rings are the training points.
Chapter 4 of 7
Copy or create
Larkspur’s lawyers ask a fair question: will the model draw new pictures, or hand back copies of the ones it was trained on?
This chapter: Does a denoiser copy its training pictures or make new ones?
About this chapter
You will learn
- Why the ideal denoiser can only copy its training points
- Why a learned network draws new points along the shape
- Why real models can still memorise
Words you’ll meet
- Ideal denoiser
- The best possible guess for exactly these training points: an average of them, weighted by fit.
- Memorise
- Hand back a training example, or something very close to it, instead of a new one.
- Generalise
- Draw new examples that fit the shape of the training data without copying any of them.
Step 1 of 4
The red rings are 24 training points on the hinge’s outline. The is the best possible guess for them.
Predict first.