Welcome back
Last time, at the novelist’s desk, we watched a network create, choosing one likely word after another until something new appeared on the page.
This is our final lesson. There’s nothing new to memorise today. Find somewhere comfortable, and let’s look up together.
A sky full of lessons
On a clear night, far from the city, the sky is crowded with stars. At first they look scattered, random. Then someone points and says, “There, see? That’s the hunter, and that’s his belt.” Suddenly the scattered lights become a shape you’ll never stop seeing.
Every lesson you’ve learned is a star. Let’s trace the lines.
Tracing the constellation
It began at a chessboard, with one idea: a network learns from examples instead of following rules.
We knelt beside the termites and found that a network is built from tiny, simple neurons, each one listening, combining and passing a signal along. On the soccer pitch we learned that some voices weigh more than others. That’s what weights are. In the music lesson we saw that a neuron only “sounds” when its signal is strong enough. Then aviation lifted those neurons into layers, carrying signals from runway to clouds.
Next, we learned how a network learns. Archery measured how far the arrow strayed. Hiking in fog showed us how to feel our way downhill. The morning shower taught us not to turn the tap too boldly. The detective traced blame backward, and dominoes showed how one push ripples through every layer. Marathon training reminded us that learning is laps, run again and again.
Then we learned to keep a network honest and sturdy: the fair classroom exam, the tailor‘s too-tight coat, the theatre understudy, baking by the tray, the fading whisper, the well-sown garden, the cyclist‘s momentum, the photographer‘s even light.
And finally, the special shapes networks take: the post office sorting its letters, words placed on a map, the birdwatcher scanning patch by patch, the cathedral of features, the campfire tale that remembers, the librarian‘s gates, the dinner party host’s attention, the newsroom where every desk reads every page, the traveller‘s suitcase of skills, and the novelist choosing the next word.
Step back, and it’s all one shape: simple units, connected by weights, adjusted little by little to make fewer mistakes. Everything else is a variation on that theme.
What networks can do
They’re genuinely remarkable at finding patterns in large amounts of data, often patterns too subtle for any person to write down as rules. They recognise faces and voices, translate languages, help doctors read scans, predict how proteins fold, write fluent text and play games beyond any human champion.
What they cannot do, or not yet
It’s just as important to be honest about the limits.
- They only know what their examples showed them. Feed them narrow or biased data, and they’ll learn narrow or biased patterns, often without anyone noticing.
- They can be confidently wrong. As we saw at the novelist’s desk, a fluent answer is not the same as a true one.
- They’re hard to explain. A network’s knowledge is spread across millions of weights, and we often can’t say exactly why it made a particular choice.
- They stumble outside familiar ground. Show them something truly unlike their training, and they may fail in ways a child wouldn’t.
None of this is a reason to fear them, or to worship them. It’s a reason to understand them, and now you do.
What to carry with you
A neural network is many simple parts, connected and patiently adjusted until patterns emerge. It’s powerful, and it’s limited, and you now understand both.
Where to go from here
You’ve done something lovely. Thirty small lessons, one at a time, and now the night sky makes sense. Thank you for walking it with me.
If you’d like to keep going, here are a few gentle next steps:
- Watch a network learn. Look for an interactive “neural network playground” in your browser. You can add neurons and layers and watch the loss fall, live. Many of these lessons will click into place.
- Try a little code. If you’re curious, a beginner’s course in Python followed by a friendly introduction to deep learning will let you train a small network yourself in an afternoon.
- Learn the gentle maths. A short visual series on linear algebra and calculus will turn “downhill steps” and “the chain rule” from metaphors into things you can see.
- Revisit any star. Go back to whichever lesson felt hazy. It will read differently now that you can see the whole constellation.
Clear skies, and happy learning.