Tutorials · Chapter D (4/4) · ~10 min

Neural nets by building

Play → read → next

Steer a tiny neuron with weights until the output matches your target — a demo you can narrate.

Playground

Neural nets by building

Tweak weights. A neuron: weighted sum → sigmoid → decision.

x1 (input): 1x2 (input): 0w1: 0.80w2: -0.40bias: 0.10

z = 0.90 · sigmoid(z) = 0.71

Recap

What you just did

You controlled a mini neural unit by hand. When you raised a weight, the output moved. That’s intuition you can’t get from a diagram alone: networks learn by changing numbers that change behavior. You can say, “Watch — I nudge this weight, the prediction flips.”

Teach

How it works

One neuron, stripped down:

def neuron(x1, x2, w1, w2, b):
    z = w1 * x1 + w2 * x2 + b
    # step activation for the toy
    return 1 if z >= 0 else 0

# Want AND-ish: both inputs 1 → 1
print(neuron(1, 1, 1.0, 1.0, -1.5))  # 1
print(neuron(1, 0, 1.0, 1.0, -1.5))  # 0

Inputs → signals
Weights + bias → importance / threshold
Activation → turn the sum into a decision

Stack many of these and you get deep nets — same idea, more knobs.

Use it

When you'd use this

Understanding why training “moves gradients” (weights change on purpose)
Debugging “dead” units (weights so small the neuron never fires)
Explaining AI to someone: “It’s weighted combinations, not vibes”

Watch out

Hand-tuning feels like understanding, then you meet thousands of weights. Scaling needs automatic training (the previous labs). Don’t confuse a toy neuron with “I built ChatGPT.”

Try next

Try this next

Flip the sign of one weight and try the same inputs. Explain aloud why the decision changed.