Neurons as numbers, layers as structure

What you’ll learn

The opening lesson named what we are after, a function that takes 784 brightness numbers in and gives 10 scores out, and then deliberately left it as a sealed box. This lesson opens the box. You will see that the inside is far less mysterious than the phrase “neural network” suggests: there are no tiny brains and no thinking, only layers of numbers with values flowing forward from one layer to the next.

You will learn that a neuron is simply a container holding one number between 0 and 1 (its activation), trace a real pixel into the input layer (one neuron per pixel, 784 for a 28 by 28 image), and read a guess off the output layer (one neuron per digit, the tallest activation wins). You will meet the hidden layers that sit in between and do the work, see why the example network’s 784-16-16-10 shape adds up to 826 neurons, and learn why this one-directional design is called feedforward. The lesson closes with the appealing edges-to-loops story of what hidden layers might be doing, framed honestly as a hope to hold loosely rather than a proven fact.

Where this fits

This is lesson 2 of the track and the first of three lessons that crack open the function from lesson 1. The opener named the goal; this lesson reveals the layered structure; lesson 3 shows what makes a single neuron light up (weights, biases, and a squishing function); and lesson 4 zooms back out to see the whole thing as one big function with thousands of adjustable knobs. Together these four lessons answer “what is a neural network, structurally?” before the track turns, in Phase 2, to how such a network actually learns.

Before you start

Prerequisite (within this track): lesson 1, The handwritten-digit problem, so that “a function from 784 numbers to 10” is already familiar and you know why we are reaching for a learned function rather than hand-written rules. No math, coding, or installation is needed. If you can picture an image as a grid of brightness numbers, you are ready.

By the end, you’ll be able to

Explain that a neuron is a container holding one number (its activation) between 0 and 1, with the brain-cell analogy set aside
Describe the input layer as one neuron per pixel (784 for a 28 by 28 image) and the output layer as one neuron per answer, read by its tallest activation
Explain what hidden layers are and that the example network’s 784-16-16-10 shape (826 neurons) reflects design choices, not rules
Explain why this kind of network is called feedforward (numbers flow one direction, no loops)
Hold the edges-to-loops account of hidden layers as a useful hope rather than a proven description

Time and difficulty

Read time: about 10 minutes
Practice time: about 14 minutes (a load-a-pixel-and-read-a-guess exercise, a count-the-neurons drill, and flashcards)
Difficulty: standard