Deep Learning
Deep learning is not a fourth answer to How Machines Learn?—it is the same three answers, run through a neural network that is flexible enough to shape itself to the problem. What makes it feel like a different thing is that the network also learns the features, a job that in every earlier section of this article was done by a human.
Take a spam filter. A classical model needs someone to decide in advance what to measure: how many links the email has, whether the sender is known, how many words appear in all caps. The model then learns weights over those hand-chosen numbers. A deep model is handed the raw characters instead, and discovers for itself that certain letter patterns form suspicious words, that certain word patterns form urgency, and that urgency plus an unknown sender means spam. Nobody wrote down "urgency." It emerged because it was useful for lowering the error.
Where each learning signal reappears. The paradigms from section 1 do not disappear inside a neural network; they become the thing that supplies its training signal.
| Learning signal | In classical ML | In deep learning |
|---|---|---|
| Statistical data | Regression, trees, SVMs, Bayesian models fitted to engineered features | Supervised and self-supervised training — the same fitting-to-data, over features the network invents |
| Experience of taking actions | Tabular reinforcement learning; genetic algorithms over explicit parameters | Deep reinforcement learning and RLHF, where the reward replaces the labelled answer; neuroevolution, where the population is a population of networks |
| Rules and logic | Expert systems, search over game rules, formal proof | Neuro-symbolic systems: the network perceives and proposes, a symbolic engine verifies, constrains, or executes |
The overwhelming majority of deep learning is the first row. The second row is how a model is aligned to a goal that has no labelled answer—winning a game, or being judged helpful by a person. The third is the active frontier, and the reason section 6's symbolic methods are still worth learning.
A network, briefly. A single artificial neuron does almost nothing: it multiplies each of its inputs by a weight, adds them up with a bias, and passes the total through a non-linear function that decides how strongly it fires. Stack a few thousand of those into a layer, stack layers on top of each other, and the composition becomes a function capable of expressing relationships no single equation from section 2 could hold.
Training is a loop, and it is pure statistical learning of the kind described in section 1.1. Data flows forward through the layers and the network produces an answer. A loss function measures how wrong that answer was. Backpropagation then works backwards through every layer, computing how much each individual weight contributed to that error, and gradient descent nudges each weight a small step in the direction that would have made the error smaller. Repeat over millions of examples and the weights settle into a configuration that encodes the regularities of the data—the same "for each additional square foot, price increases by about $150" idea from section 1.1, except distributed across billions of parameters that no human wrote or can read individually.
From patterns to meaning. The reason depth matters is that each layer builds on the abstractions of the one below it, so representations become progressively less about the raw input and more about what it means.
In vision, the first layers respond to edges and colour gradients. The next combine those into corners and textures, then into eyes, wheels, and leaves, until the final layers respond to whole objects. In language the same ladder runs from characters to words, from words to phrases, and from phrases to concepts. Along the way each token is represented as a vector—an embedding—positioned so that things used in similar ways sit near each other. Meaning becomes geometry: "Paris" lands near "Tokyo" because both behave like capitals, and the direction that separates "king" from "queen" turns out to be roughly the same direction that separates "actor" from "actress."
Those embeddings start out context-free, which is not enough—"bank" in river bank and savings bank need different representations. Attention mechanisms fix this by letting every word look at every other word in the sentence and pull in the parts that disambiguate it, so a word's vector is rewritten in light of its context. This is the step that turned statistical pattern extraction into something that behaves like semantic understanding, and it is the foundation of the transformer models behind modern language systems.
This is why deep learning has advanced computer vision, natural language processing, speech recognition, and a widening set of scientific fields so sharply: it captures intricate patterns that are real but beyond anyone's ability to write down. The cost is the mirror image of section 3's deterministic models—enormous appetite for data and computation, and weights that resist the kind of step-by-step explanation a decision tree gives you for free.
Follow the link for a full treatment: neural network structure, how networks learn, the architectures that specialise them for images, sequences, and graphs, and the transformers that dominate language today.