a visual course

From Pixels
to Predictions

Understand neural network foundations in less than an hour.

How does a machine actually see an image? Start with a single pixel and build, animation by animation, all the way to a network that recognises what it's looking at.

Start from the first pixel →

Part 1 — An image is just numbers

1

Pixels & the grid

Zoom into a photo until it dissolves into a grid of numbers.

2

Color & channels

Grayscale 0–255, then RGB as three stacked layers.

3

Images as tensors

Height × width × channels — a picture as a stack of matrices.

Part 2 — Doing things to pixels

4

Pixel math

Brightness, contrast, threshold — functions over every pixel.

5

Convolution & filters

A kernel slides across the grid. The keystone idea.

6

Features: edges & gradients

Why edges carry more signal than raw pixels.

Part 3 — Why this is hard for machines

7

The curse of raw pixels

120,000 numbers, and a 5px shift breaks everything.

8

Hand-designed vs learned

From features we design to features a model discovers.

Part 4 — Teaching a machine to see

9

A neuron & a tiny network

Weights × inputs + bias → a prediction.

10

Convolutional networks

The sliding kernel — but now the numbers are learned.

11

How a CNN sees

Edges → textures → parts → objects, layer by layer.

12

Training: how it learns

Loss, gradient descent, and the ball rolling downhill.