Neural Networks

Learning Objectives

This lesson sits in World 3: How AI Learns, and it's the one where the "brain" metaphor kids hear everywhere about AI finally gets opened up. By the end of the lesson, across all four age bands, a student should be able to:

• Explain that a neural network is a structure made of layers of simple math units, only loosely inspired by how brain cells connect to each other — it is not a tiny computer brain and does not think or feel anything.
• Describe what the three layers of a network do: an input layer takes in raw information (like the numbers behind a photo's pixels), hidden layers look for patterns in that information, and an output layer produces the final answer or guess.
• Explain that a weight is a number attached to each connection between units, that it controls how much influence one unit has on the next, and that "learning" means slowly adjusting those numbers rather than storing memories the way a person does.
• Describe the basic guess-then-fix cycle at their level: the network makes a guess (a "forward pass"), that guess is compared with the right answer, and the error is sent backward through the network to nudge the weights — repeated an enormous number of times until the guesses get good.
• For Hacker and Architect classes: connect at least one real architectural idea — convolutional networks, vanishing gradients, skip connections, or scaling laws — to why today's AI systems are actually built the way they are, not just "brain but bigger."

Teacher Background

You don't need a computer science degree to teach this lesson well, but the "neural network" name causes real confusion, so it's worth being precise about what it actually is before you stand in front of a class.

A neural network is a piece of software built from many small, identical units, usually called artificial neurons, arranged in layers. Each unit does something very unglamorous: it takes in some numbers, multiplies each one by an importance value called a weight, adds them up, and passes the result to the next layer through a simple mathematical filter. That's the whole trick, repeated millions or billions of times across a network. There is no memory, no imagination, and no awareness anywhere in that process — just arithmetic, organized in a particular shape.

The name comes from a loose analogy to the brain. Biological neurons really do connect to each other and pass signals, and the pioneers of this field in the 1940s–1950s borrowed that picture as inspiration for a mathematical model. But the resemblance mostly stops at the picture. A real neuron is a living cell with complex chemistry; an artificial "neuron" is a single line of arithmetic. Your brain has roughly 86 billion neurons wired in ways scientists still don't fully understand; even a large modern AI model, while it may have billions of connections, is organized in tidy, repeating layers that look nothing like brain tissue. The analogy is useful for building intuition and that's exactly why AIQ uses it — but it is an analogy, not a description, and it's worth saying that out loud to students.

The layers themselves do have a genuinely useful structure to explain. Take a network that identifies photos of animals. The input layer receives the raw pixel numbers of the photo. The hidden layers in between are where pattern-finding happens — commonly, early hidden layers pick up on very simple things like edges and color patches, and later hidden layers combine those into more complex shapes like "ears" or "whiskers." The output layer turns all of that into a final answer, often as a confidence score: "95% cat, 5% dog." No single layer "knows" it's looking at a cat — the recognition emerges from the combination.

Learning is the process of adjusting all those weights so the guesses get better. The network first makes a guess with whatever weights it currently has — this is called a forward pass. That guess is compared against the correct answer, producing an error. The key step, called backpropagation, sends that error backward through the network, computing exactly how much each individual weight contributed to the mistake and nudging it a tiny amount in the direction that would have reduced the error. Do this across millions of labeled examples and the weights gradually settle into values that make good predictions — that settling process is called convergence. "Deep learning" simply means a network with many hidden layers stacked between the input and output; it does not mean the network thinks more deeply about anything.

It's worth naming why the "nudge the weight a tiny amount" step matters so much: it's tiny on purpose. If a weight jumped a large amount every time it was slightly wrong, the network would overcorrect wildly and never settle down. The size of each nudge is controlled by a setting called the learning rate — small enough that thousands or millions of gentle corrections, each based on real examples, gradually steer the weights toward good values instead of bouncing around. You don't need to teach "learning rate" as a term to any age band in this lesson, but it's useful background if a student asks why learning takes "millions of examples" instead of getting it right after a handful of tries: each individual correction is deliberately small, so it takes many of them.

It's also worth connecting this lesson back to AIQ's broader premise: AI is AI-native, not AI-dependent. The point of opening up the "brain" metaphor isn't to make neural networks feel magical — it's closer to the opposite. Once a student sees that a network is repeated arithmetic guided by loss and correction, "the AI understood my essay" or "the app knows what I like" starts to sound like what it actually is: a very good pattern-matcher trained on a lot of examples, not a mind that shares their understanding. That's the seed of healthy skepticism this lesson is trying to plant at every age.

The Hacker (11–14) and Architect (15–18) scripts go further into real, current concepts: convolutional networks that share weights across an image, the vanishing-gradient problem that makes very deep networks hard to train, dropout as a way to prevent over-reliance on single units, skip connections that let gradients flow through extremely deep networks, and the scaling laws that describe how model performance improves with size and data. These are not simplified myths invented for the app — they are the real vocabulary of the field, explained at a level a strong teenager can follow. Read the Hacker and Architect scripts through once before class even if you're teaching a younger band; the "why" behind the terms will sharpen how you explain the simpler version too.

Materials & Prep

This lesson needs almost nothing beyond what AIQ already runs on:

• One device per student (or per pair) with a browser and internet connection, loaded to the AIQ app at lesson 3.2.2 — no install, account, or login required to learn.
• A projector or shared screen is helpful for walking through the hook and quiz as a group before or after individual/paired play, but is not required.
• For Hacker and Architect classes, scratch paper or a whiteboard is genuinely useful for sketching a layer diagram (input → hidden → output, with arrows for forward and backward passes) — it's a two-minute drawing that makes the backpropagation idea click.
• For Explorer and Builder classes, open floor space is useful if you run the "human neural network" extension activity described at the end of each script.

Common Misconceptions

Pick your grade's script