# 2026-06-21 Noel Handwritten notes (reMarkable) on **neural nets, as explained by Noel**. Two pages: page 1 is diagrams + formulas in blue ink; page 2 is a red-ink caption at the bottom. ## Transcribed text ### Page 1 (blue ink) A single neuron / perceptron sketch with these annotations: - `xᵢ ∈ [0, 1]` — inputs - `wᵢ` — weights on the incoming edges - `f(Σ wᵢxᵢ)` — activation function applied to the weighted sum (inside the neuron body) - `→ y ∈ [0, 1]` — output Below the neuron: - `Error = Σ yᵢ − expected_i` - `ΔE → Δwᵢ` (change in error drives change in weights) At the bottom, in handwriting: - `x ≠ y` ### Page 2 (red ink) > neural nets, as explained by Noel > > Get diagrams from WhatsApp ## Diagram descriptions **Top of page 1 — single neuron (perceptron):** A drawing of one neuron drawn as an oval/cell body. Several arrows feed in from the left labelled with inputs `xᵢ` and weights `wᵢ`. Inside the body is the function `f(Σ wᵢxᵢ)`. A single arrow exits to the right producing output `y ∈ [0, 1]`. This depicts the basic perceptron: weighted sum of bounded inputs passed through an activation function to produce a bounded output. **Middle band of page 1 — three small sketches (left to right):** 1. A hand-drawn grid/mesh (looks like a warped square grid) with some scribbled digits inside — suggesting a pixel grid / input image being represented. 2. A column of 0s and 1s on the left with lines fanning out to a small quadrilateral on the right — suggesting a binary input vector feeding into a layer/region (flattening an image into a vector of bits). 3. A tall vertical strip of cells filled with 0s and 1s, beside a numbered list `0 1 2 3 4 5 6 7 8 9` — i.e. a one-hot / output vector mapping to the ten digit classes (classic MNIST-style digit classification). **Bottom of page 1 — network graphs:** - Left: a small node-and-edge graph (a few neurons connected by edges converging toward a single node at the bottom), captioned `x ≠ y`. This illustrates a tiny network / the XOR-style case where output differs from input. - Middle: two parallel columns of small circles (0/8-like shapes) — two layers of nodes drawn as stacked neurons. - Right: a denser multi-layer network with many nodes and crossing edges between layers (a fully-connected feed-forward net), with an arrow/marker at the top. ## Notes The red caption on page 2 indicates these are notes from an explanation of neural networks given by **Noel**, with a reminder to **get the (cleaner) diagrams from WhatsApp**. The blue sketches are the working diagrams: a single perceptron, the input-encoding pipeline (image → binary vector → digit classes), and progressively larger network graphs, alongside the core training relationship `Error = Σ(yᵢ − expectedᵢ)` and `ΔE → Δwᵢ` (gradient-descent style weight updates driven by error).