L19 | English F3 Student Learning Log

Persona: I'm a Form 3 (Grade 9) student and I only read English. I know basic Python (variables, for loops, lists) and school algebra up to quadratics. I learn from the official pages, the videos with captions, and the book. Kaggle notebook steps are written as a practice log that follows the official notebooks.

What I understood (in my own words)

Diffusion training is clever: deliberately make an image noisy, then teach the model to guess what noise was added.
Adding noise needs no learning, just a formula. Removing it is the hard part.
The model predicts the noise, not the image directly.
Keeping dropout on at test time and predicting several times shows how unsure the model is.

Where I got stuck, and how I got unstuck

Where I got stuck	How I solved it
Why feed the timestep t into the model?	Different steps have different amounts of noise; the model needs to know how noisy the image is to know how much to remove.
Lots of α and β symbols	I noted that they control how much noise is added at each step and skipped the details.

My answers to the exercises

Q1. What does a DDPM model predict during training, and what is the loss?

The model takes a noisy image (and the timestep) and predicts the noise that was added. The loss is the mean squared error between predicted and actual noise.

Q2. Which needs a trained network, the forward or reverse process? Why?

The reverse process. Forward noising is done by us with a formula, so the answer is known. Going from noise back to an image has no formula, so a network must learn it from many examples.

Q3. With dropout on at test time, you predict 10 times and the answers vary a lot. What does that mean?

The model is unsure about that input; different combinations of neurons give different answers. Its prediction isn't reliable, so it may need a human check or more training data.

My Zettelkasten card

Card｜To learn to restore something, first learn to spot the noise

DDPM training is clever. You add noise to clean images yourself, so you know the right answer, and train the model to guess what noise you added. Once it can do that, you can start from pure noise and remove predicted noise step by step to create a brand-new image.

It reminds me of listening practice in French class. Our teacher played conversations with café noise in the background, starting quiet and getting louder. Over time we got better at separating voices from noise. The model also practises at every noise level. There's one big difference, though: the model can make an image from pure noise, but no person could hear a conversation in pure static. That's the truly surprising part of generative models.

My next step: add three strengths of random noise to an 8×8 greyscale image in Python and plot them side by side.

🔗 Related notes to grow next: → Learning: 'Make practice questions where you know the answer' → AI: 'Generation is removing noise with a direction' → Uncertainty: 'If repeated answers disagree, you're not sure yet'

Back to lesson note：L19_DDPM與Dropout_202611041526