AI Foundations
The secret behind face recognition, language translation, and AI art — explained using layers of invisible specialists who pass notes to each other.
Imagine a factory assembly line where every station does one tiny job: the first station checks whether an object is round, the next checks if it's reddish, the third checks whether the red round thing has a stem. By the time the widget reaches the end of the line, a verdict pops out: "That's probably an apple."
That's essentially how a neural network works. Instead of a factory, it's layers of simple maths. Instead of workers, it's numbers called "weights." And instead of assembly stations, it's layers — sometimes just three, sometimes hundreds. Together they can recognise faces, translate sentences, compose music, and power the AI tools you use every day.
A neural network is a type of computer program designed to find patterns in data. It was loosely inspired by the way biological neurons in your brain connect and fire signals at each other — but the resemblance is mostly in name. A digital neuron is just a number. It takes some numbers in, multiplies them by some other numbers (its "weights"), adds them up, and passes the result forward.
Put thousands or millions of these little calculations into layers, train the whole system on enormous piles of examples, and something remarkable happens: the network starts recognising things it was never explicitly programmed to recognise. It figures out the rules for itself.
Every neural network — from a tiny hobbyist project to a massive commercial AI — has at least three parts:
Raw data goes in here — pixel values for an image, word tokens for text, numbers for financial data. This layer simply receives and passes data forward without doing any clever processing.
This is where the magic happens. Early hidden layers detect simple features (edges in an image, common word pairs in text). Later hidden layers combine those into complex features (faces, sentence meaning). "Deep" learning means many hidden layers.
The final answer comes out here — "this is a cat," "this sentence is positive," "the predicted temperature is 23°C." The output layer turns the network's internal calculations into something human-readable.
Between every layer are millions (or billions) of numbers called weights. They control how strongly each neuron influences the next. Training a neural network means adjusting all these weights until the network produces correct answers consistently.
Here is where the real cleverness lives. You show the network a million labelled photos — "this is a cat, this is a dog" — and at first it guesses randomly. For each guess, you measure how wrong it was (this is called the "loss"). Then you do a mathematical trick called backpropagation: working backwards through every layer, you nudge the weights very slightly in whatever direction would have made the guess less wrong.
Do this billions of times — with a billion example images — and the weights gradually settle into values that produce correct answers. Nobody explicitly coded "a cat has pointy ears." The network figured that out on its own from the data.
The brilliant analogy: Think of it like tuning a million tiny dials at the same time. Each training example tells you which dials were set wrong and by how much. Over enough examples, all the dials end up in just the right position to do the job accurately.
Neural networks excel at tasks where the rules are too complex or too fuzzy to program manually:
Neural networks have genuine limitations that are worth understanding:
A child can learn what a dog looks like from five or ten examples. A neural network often needs tens of thousands. If your training data is scarce, biased, or unrepresentative, the network learns the wrong things confidently.
Even the researchers who build neural networks often cannot explain exactly why the network reached a particular answer. It's not like a rulebook you can audit — it's millions of weights interacting in complex ways. This "explainability" problem is a major area of ongoing research.
Tiny, often invisible changes to an input — a few pixels changed in an image, a slightly rearranged sentence — can cause a neural network to produce a completely wrong answer with high confidence. This is called an "adversarial attack" and is part of why AI safety research matters.
A neural network that reads medical literature can generate plausible-sounding medical advice without actually understanding medicine. It is matching patterns in language, not reasoning from principles. That distinction matters enormously when the stakes are high.
They were inspired by the brain but work very differently. Biological neurons are chemical and electrical; digital neurons are just numbers and maths. Real brains have roughly 86 billion neurons with incredibly complex connections; even the largest AI neural networks have far less going on, and they can only do the specific task they were trained for.
It varies enormously. A simple network might have 3–5 layers; a large language model like GPT-4 has many hundreds of transformer layers stacked together. More layers generally means the network can learn more complex patterns, but also needs vastly more data and computing power to train.
"Deep" simply means the neural network has many layers — deep as in deep stack. A network with just one or two hidden layers is called a shallow network. Deep learning networks have enough layers that each layer can learn progressively more abstract features from the data.
Absolutely, and they can be confidently wrong. Neural networks learn statistical patterns from their training data. If that data was biased, unrepresentative, or simply did not include a particular situation, the network will make mistakes — sometimes spectacularly. That is why human oversight remains important.