AIQ
AIQ
Part of the Neural Networks lesson guide. Teaching a different grade? 🌈 Explorer (5–7) · 💻 Hacker (11–14) · ⚡ Architect (15–18)
Open with the lesson's real hook line, read it almost word for word: "Inside every modern AI is something inspired by your brain — a neural network! It's layers of math that work surprisingly well."
Ask the class: "Who's used something with AI in it this week?" (Expect answers like a phone camera that finds faces, a game recommendation, a voice assistant, YouTube's "up next.") Pick two or three examples and write them on the board.
Say: "Every single one of those is running something called a neural network. Today we're going to actually build one out of paper and people, and figure out why it's called that."
Quick poll before you start: "Raise your hand if you think a computer has an actual brain inside it." Note roughly how many hands go up — you'll circle back to this exact question after the activity, once they've seen how the "brain" idea is really just an inspiration for a math structure.
Do a one-minute stretch to get bodies moving before the main activity: have everyone stand and play a quick "signal relay." Whisper a two-digit number to the student on one end of a row; that student whispers the number, plus 1, to the next student; each student down the row adds 1 more before whispering it on. The last student says the final number out loud, and the class checks the math together. Say: "That's basically what's about to happen inside our paper neural network — except instead of always adding 1, every connection will use its own number, and we'll figure out where those numbers come from."
Work through the lesson's three scenes as a build-it-yourself demonstration.
Say: "Your brain has about 86 billion neurons, connected by trillions of tiny wires called
synapses. Every thought you have is actually millions of these neurons firing electrical signals
in patterns." Ask: "So if AI copies that idea, what do you think an 'artificial neuron' does?"
Guide toward: it's a stand-in — a small math function. It takes numbers in, multiplies each one
by a weight (a number that says how important that input is), adds everything up, and
sends a new number out. Write on the board: numbers in → multiply by weights → add up →
number out.
Make it concrete with a two-input example: "Say a tiny artificial neuron gets two numbers in
— 4 and 2 — and its weights are 3 and 5. What does it do?" Walk the class to 4×3 + 2×5 =
12 + 10 = 22. Point out: "Notice the weight of 5 mattered more to the final number than
the weight of 3, even though the second input number was smaller. That's exactly what a weight
does — it decides how much influence one input has, no matter how big or small the input itself
is." Have pairs try one more example with numbers you give them, swapping which weight is bigger,
and predict which input will end up mattering more before they calculate.
Draw three columns on the board labeled Input, Hidden, Output. Say: "Let's build a network that recognizes a photo of a dog." Input layer: "This is where the raw picture — really just a giant list of pixel color numbers — enters the network." Hidden layers: "This is where the pattern-finding happens. One layer might notice edges and outlines. The next layer combines those into shapes like ears or a snout." Output layer: "This is the final decision layer — it announces something like 'Dog: 92% sure.'"
Have students work in pairs: hand each pair a picture (printed or on a device) and have them write down, layer by layer, what an input/hidden/output layer might "notice" about it, in the style of the board example. Ask a few pairs to share.
Circulate while pairs work and listen for a common sticking point: students sometimes assume the hidden layers "know" they're looking for ears or a snout ahead of time. Clarify: "Nobody programs the hidden layers to look for ears specifically — during training, the network discovers on its own that noticing ear-shapes helps it guess correctly, and it keeps that pattern because it works. We're skipping ahead and describing what a trained network ends up noticing, not giving it instructions."
Introduce "deep learning": "When people say a network is 'deep,' they just mean it has a LOT of hidden layers stacked up. GPT-4, one of the AI models behind chatbots, has billions of connections spread across dozens of layers." Ask: "Does 'deep' mean it thinks more carefully?" (Expected, after discussion: no — it just means more layers, not more wisdom.)
Before demonstrating, ask the class to predict: "If our network guesses 'cat' but the photo is really a dog, what do you think has to change to fix it?" Take a few guesses — students often say something like "delete the wrong answer" or "start over." Say: "Nobody deletes anything, and it doesn't start over from scratch. Watch what actually happens."
Say: "Data flows forward through every layer and produces a guess — that's called a forward pass. If the guess is wrong, here's the part that surprises most people: the error travels backward through the network, and every single weight gets nudged a tiny bit to make that mistake less likely next time." Demonstrate physically: line up 4–5 students as layers, pass a ball forward for the guess, then have the class shout "wrong!," and have the line pass a different object backward hand to hand while each student says "adjusting my weight a little."
Wrap the science with: "This happens across millions of examples until the weights settle into good values — that's called convergence. Training is done once the guesses stop getting much better."
Check understanding with a quick whiteboard recap: draw an arrow labeled "forward pass" going left to right through three boxes (Input, Hidden, Output), and a second arrow labeled "backward pass" looping right to left underneath it. Ask a volunteer to describe, in their own words, what's traveling along each arrow — a guess going forward, an error going backward — before moving to discussion.
These work well as a think-pair-share: give 30 seconds of silent thinking, then a partner chat, before opening to the whole class.
Close with: "So a neural network isn't a computer brain — it's layers of math, built out of an idea borrowed from the brain, that gets better through millions of tiny corrections. That's genuinely how the AI behind your phone's camera, your game recommendations, and chatbots all work underneath." Revisit the opening poll — ask again who thinks a computer has an actual brain, and see if the show of hands has changed.
It's worth naming the honest limit of today's lesson too: "We built a tiny, three-layer example by hand. Real networks like the ones behind your phone's camera can have dozens of layers and millions or billions of connections, all trained automatically on huge amounts of data — we couldn't do that by hand if we tried all year. But the basic idea, multiply-add-check-and-fix, is exactly the same whether it's our paper version or a giant real one."
Extension activity — Weight Tug-of-War: Give small groups a simple task: sort five photos (or descriptions) into "cat" or "not cat" using only three clue-cards you hand them — "has pointy ears," "has whiskers," "has a tail." Have each group assign a point value (a "weight") from 1–5 to each clue based on how useful they think it is, then score each photo by adding up the weighted clues that apply. Compare groups' weight choices and results — this is literally what a network's weights do, just done by hand instead of by millions of automatic adjustments.
Push the extension one step further if there's time: after groups score their photos, tell them one of their answers came back wrong and ask them to adjust just one weight, by a small amount, to try to fix it without breaking the answers that were already right. This is exactly the tension a real network has to balance during training — fixing one mistake without wrecking everything it already learned — and groups quickly discover why the real fixing has to happen in tiny nudges rather than big jumps.