Summary: Neurons as numbers, layers as structure

Last lesson named the goal: a function that takes 784 brightness numbers in and gives 10 scores out. This lesson opens that function up, and what is inside is far less mysterious than the word “neural network” suggests. There are no tiny brains and no thinking, just layers of numbers with values flowing forward from one layer to the next. A neuron is a container holding a single number between 0 and 1, the network arranges those containers into layers, and reading a guess is as simple as finding the tallest number in the last layer. This is the scan-it-in-five-minutes version.

Core ideas

• A neuron is just a number. Forget biology: a neuron is a container holding one value between 0 and 1, called its activation. Near 1 it is “lit up,” near 0 it is quiet. The whole network is many of these boxes with numbers flowing between them.
• The input layer is one neuron per pixel. A 28 by 28 image has 28 times 28 = 784 pixels, so the input layer has 784 neurons, each holding its pixel’s brightness (0 = black, 1 = white). The thing your eye reads instantly as “a 3” enters as 784 numbers in 784 boxes.
• The output layer is one neuron per answer. Ten neurons, one per digit (0 through 9). Each holds a confidence score, and the network’s guess is simply the digit whose neuron is tallest. A near-tie between two of them means the network is hesitating.
• Hidden layers sit in the middle and do the work. “Hidden” only means “not input, not output.” The example network is 784, 16, 16, 10, which is 826 neurons in all. The two hidden layers of 16 are a clean teaching choice, not a rule; real networks vary enormously.
• Feedforward means one-directional flow. Numbers move input to output, each layer feeding the next, with no loops and no later layer feeding back. That one-way flow is the simplest neural network architecture there is.
• The edges-to-loops story is a hope, not a fact. It is tempting to say the first hidden layer detects small edges and the second combines them into loops and strokes. That is a useful first picture, but a real trained network organizes itself more messily and less readably than that clean story suggests. Hold it as motivation, with a mental asterisk.

What changes for you

The next time a headline talks about an AI’s “brain” or a model’s “billions of neurons,” you have a deflating translation ready: that means billions of little numbers, nothing spookier. This reframing does real work when you use AI tools. It explains why these systems have no awareness of what they are doing, why their “confidence” is literally just which output number came out tallest, and why the same layered structure can read digits, faces, or audio without caring which, because it is always numbers in, numbers out. The lesson left one question open on purpose: what actually makes one neuron light up more than another? That is weights, biases, and the squish, and it is where the next lesson goes.