L20 | 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)

Calculating with shorter numbers (16-bit) halves the memory and is faster.
Style transfer blends a photo's content with a painting's style.
A Gram matrix captures which features tend to appear together, and that's the style.
Cellular automata only look at neighbours but grow very complex patterns.

Where I got stuck, and how I got unstuck

Where I got stuck	How I solved it
Why not use 16-bit everywhere?	16-bit has a smaller range and precision; very small gradients can round to 0, so the important parts stay 32-bit.
Computing a Gram matrix	I noted that it multiplies every pair of features and sums them, giving a table of how features relate.

My answers to the exercises

Q1. A 32-bit float takes 4 bytes and a 16-bit float 2 bytes. How much memory do 100 million parameters take in each format (in MB, 1 MB ≈ 1 million bytes)?

32-bit: 100 million × 4 bytes = 400 million bytes ≈ 400 MB. 16-bit: 100 million × 2 bytes ≈ 200 MB. Exactly half.

Q2. In style transfer, what are the content loss and the style loss each responsible for?

Content loss keeps the generated image's middle-layer features close to the photo's, preserving shapes and positions. Style loss keeps its Gram matrices close to the painting's, capturing brushstrokes, colours and texture.

Q3. In Conway's Game of Life, how does each cell decide its next state? How does this relate to Neural Cellular Automata?

Each cell counts how many of its 8 neighbours are alive and follows fixed rules to live or die next step. Neural Cellular Automata also have each cell look only at its neighbours, but the rules are learned by a neural network.

My Zettelkasten card

Card｜Simple local rules can grow a complex whole

In a Neural Cellular Automaton, each cell only looks at its neighbours and follows the same rule, yet together they grow complex, beautiful textures. Nobody directs the whole thing; the complexity emerges from lots of simple local interactions.

It reminds me of leaving a packed concert. Nobody tells each person where to go, but everyone watches the people around them, keeps a gap and follows the person in front, and the crowd naturally forms lanes to the exits. If one person stops suddenly, everyone behind jams up, just as a local change in the automaton affects the whole pattern. The rules have to be just right, though: too simple and nothing happens, too complex and it turns to chaos.

My next step: find a Game of Life simulator online, draw a 'glider' and watch how it moves.

🔗 Related notes to grow next: → Systems: 'Big behaviour comes from small local interactions' → Efficiency: 'Use just enough precision' → Art: 'Style is the relationship between features'

Back to lesson note：L20_混合精度_202611041526