Machine Learning from Zero
by Sergey Okinchuk
Models AgeNeurons and Layers

Chapter Eight · Scaling Up

Neural Networks & Deep Learning

Everything so far learned from tables. Dish photos, the meaning of a review, a voice note — those stayed out of reach, because nobody could hand-craft features for a million pixels. This chapter opens the machinery that cracked them: networks of laughably simple units that learn their own features, why decades-old ideas suddenly took over around 2012, and — the honest ending — where all this firepower is exactly the wrong tool.

5 topics

Back in Topic 02 you filed "deep learning" away as a box inside machine learning — the family behind face unlock, photo search, and the chatbots of Chapter 9. This chapter opens the box. The surprise inside is how little is new: the same training loop from Topic 11, the same overfitting disease from Topic 14, the same file of numbers from Topic 04 — just stacked, scaled, and pointed at the kinds of data that defeated everything else.

Five topics make the case. First, the unit itself — a neuron is a weighted sum, and the power lives in the stacking. Then the history question every newcomer secretly has: the ideas are from the 1980s, so why did the takeoff wait until around 2012? Then the two payoffs in person — how a network turns pixel numbers into "pad thai, 0.93", and how it places words on a map of meaning so that "cold food" and "arrived freezing" land side by side. And finally the chapter's epilogue, in this book's honest key: when deep learning is a helicopter sent on a bakery run.

The chapter in one picture — a dish photo climbing the layers
"Pad thai, 0.93"
the final layers just classify — Chapter 4 machinery, unchanged
Late layers — dish parts
noodle tangle, lime wedge, plate rim
Middle layers — shapes & textures
curves, gloss, repeating patterns
Early layers — edges & color patches
the simplest marks a picture is made of
The photo — a grid of pixel numbers
numbers in, as promised in Topic 07

Topics in This Chapter

Topic 35
Neurons and Layers
Despite the brain-borrowed name, one neuron is a weighted sum — dumber than a calculator. The power is in the stacking, and "deep" turns out to mean nothing more cosmic than "many layers".
Concept
Topic 36
Why Deep Learning Took Off
Neural networks existed in the 1980s and spent decades as a backwater. Around 2012 they started winning everything — and the ideas barely changed. Three ingredients explain the entire AI moment.
Concept
Topic 37
How Machines See
From a grid of pixel numbers to "pad thai, 0.93" — layer by layer, with nobody writing the features. The single biggest thing deep learning changed, and why vision models are still easy to fool.
Concept
Topic 38
How Machines Read
"Terrible" and "awful" share no letters and mean the same thing. Embeddings give every word an address in a space of meaning, where similarity is distance — the idea under every language model you have ever talked to.
Concept
Topic 39
When Deep Learning Is Overkill
After the fireworks, the confession: Plateful's most valuable model is still the churn random forest. Where depth earns its cost, where tables keep choosing trees, and why "we used deep learning" is a cost report, not a quality claim.
Concept