L23 | English F3 Student Learning Log
L23 | 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)
- Super-resolution turns a small blurry image into a large sharp one.
- A U-Net shrinks then enlarges, with shortcuts that carry details across.
- If you only compare pixels, the model learns to 'play it safe' by averaging, which makes images blurry.
- Perceptual loss compares whether images look alike, not whether every pixel matches.
Where I got stuck, and how I got unstuck
| Where I got stuck | How I solved it |
|---|---|
| Why MSE blurs | If the model isn't sure whether an edge is a bit left or right, guessing the middle is safest, and you get a blurry edge. |
| What does freezing mean? | Stopping certain layers' weights from updating for a while, like locking parts you've already tuned. |
My answers to the exercises
Q1. Why does using only MSE as the super-resolution loss give blurry outputs?
When unsure about details, outputting the average of all possible answers gives the lowest MSE. Averaging many possible sharp details produces blur, so MSE rewards blurry outputs.
Q2. What do U-Net's cross connections do?
Shrinking loses detail. The shortcuts pass features from each down-path level straight to the matching up-path level, so the reconstruction can reuse the original details and look sharper.
Q3. Why is 'freeze the pretrained layers briefly, then unfreeze everything' better than training everything from the start?
The new up-path starts random. Training everything together lets its messy gradients damage the features the pretrained layers already learned. Freezing first lets the new part reach a sensible state; then unfreezing to fine-tune together keeps the good parts of both.
My Zettelkasten card
Card｜Aiming only for a safe average score produces bland, blurry results
With MSE as the loss, a super-resolution model learns that the safest move is to output the average of all possible answers, which is a blurry image. Switching to perceptual loss compares whether images look alike, and outputs become sharp. What you score is what you get.
It's like creative writing. When our teacher only marked spelling and grammar, everyone wrote the safest, plainest sentences and every story sounded the same. When she started marking vivid detail and original ideas, people's writing finally had personality. Perceptual loss has its own problem, though. The lesson says there's no clear metric to compare models, so in the end people judge by eye. Good scoring rules are hard to design.
My next step: before my next story, list three specific vivid details first instead of aiming for safe sentences.
🔗 Related notes to grow next: → Assessment: 'The marking scheme shapes how students try' → AI: 'MSE rewards averages, and averages are blurry' → Learning: 'Lock in what you've learned, then fine-tune'
Back to lesson note：L23_超解析度_202611041526
