Neural networks learn patterns from examples by adjusting billions of weighted connections across many stacked layers, a process called training. This hierarchical approach powers modern AI systems like chatbots and image generators, though experts still debate fundamental questions about how these systems truly work.
Every time you ask a chatbot a question and get a surprisingly coherent answer, or watch a translation app render speech into another language in real time, or see an image-generator conjure a picture from a sentence, you are witnessing the output of a technology called deep learning. This course is about pulling back the curtain on that technology — not by teaching you to build it yourself, but by giving you a solid, honest, intuitive understanding of what it is, how it works, why it has become so central to modern artificial intelligence (AI), and where even the experts who build it still disagree about what's really going on inside it.
You don't need a background in math, statistics, or programming to follow this course. You just need curiosity and a willingness to think in analogies. Wherever technical terms show up, we'll pause and explain them in plain language the first time they appear.
At its core, deep learning is a way of building computer systems that learn from examples rather than being explicitly programmed with step-by-step instructions.
Think about how traditional software works: a programmer writes precise rules — "if the customer's balance is below zero, charge a fee" — and the computer simply follows them. That approach works well for tasks you can fully specify in advance. But it breaks down for tasks like "recognize a cat in a photo" or "write a coherent paragraph," because nobody can write down an explicit rulebook detailed enough to cover every possible cat photo or every possible sentence.
Deep learning solves this differently. Instead of being told the rules, a deep learning system is shown enormous numbers of examples — millions of labeled photos, or vast amounts of text — and it gradually adjusts itself to detect the patterns in that data on its own. This self-adjustment process is called training, and it's the heart of how these systems come to do what they do.
The tool that does this learning is called an artificial neural network, so named because it is loosely inspired by the idea of neurons in a biological brain — though the resemblance is loose, more a source of inspiration than a faithful copy.
A neural network is built from simple units sometimes called neurons or nodes. Each one does something quite modest: it receives some numbers as input, combines them, and passes along a resulting number as output. Individually, a single neuron is almost trivially simple. The power comes from connecting huge numbers of them together.
Neurons are organized into layers:
The connections between neurons each carry a weight — a number that determines how strongly a signal from one neuron influences the next. Think of weights as the "settings" or "dials" of the network. Initially, these dials are set randomly, so the untrained network's answers are essentially nonsense. Training is the process of nudging all these dials, bit by bit, across millions of examples, until the network's outputs start matching reality.
Here is a simplified view of how information flows through such a network:
A simplified neural network: data flows from the input layer, through one or more hidden layers where patterns are detected and combined, to the output layer that produces the final answer. Every connecting line carries a weight that training adjusts.
The word "deep" in deep learning refers specifically to the number of hidden layers stacked between input and output. A network with only one or two hidden layers is technically a neural network, but practitioners don't usually call it "deep learning" — the term is reserved for networks with many stacked layers.
Why does depth matter so much? Because of an idea called hierarchical feature learning. Early layers in a deep network tend to pick up on very simple, low-level patterns. In an image-recognition network, for instance, the earliest layers might detect simple edges and color contrasts. The next layers combine those edges into shapes — curves, corners, textures. Later layers combine shapes into recognizable parts — an eye, an ear, a wheel. Still later layers combine those parts into whole concepts — a face, a car. Each layer builds on the abstractions discovered by the layer before it, which is why stacking many layers lets a network capture far richer, more abstract patterns than a shallow network ever could.
This same layering principle, applied to language instead of images, is part of what allows modern systems to go from recognizing letters and words to grasping grammar, meaning, and even something that looks like reasoning.
How do all those weights actually get adjusted? At a conceptual level: the network is shown an example, it makes a prediction, and that prediction is compared to the correct answer. The difference between the prediction and the correct answer is a kind of error signal. The network then adjusts its weights — the strength of connections between neurons — to make that error a little smaller next time. Repeat this process across millions or billions of examples, and the network's weights gradually settle into values that produce genuinely useful behavior.
The precise mechanics of how errors get distributed backward through all those layers (a procedure called backpropagation) are mathematical and outside the scope of this course. What matters for your understanding is the outcome: training is how a neural network extracts patterns from data, entirely through repeated exposure and adjustment, without anyone writing explicit rules for what those patterns should be.
One of the defining discoveries behind the recent explosion of AI capability is that performance tends to improve in fairly predictable ways as you increase three things together: the size of the network (number of neurons and weights), the amount of training data, and the amount of computing power used to train it. These predictable improvement patterns are often called scaling laws, and they are a large part of why AI labs have raced to build ever-larger models.
But scale is not a guaranteed path to unlimited progress, and this is one of the genuine open debates in the field today. Some researchers argue there are walls ahead: for instance, concerns that scaling laws may limit how well models can express appropriate uncertainty about their own answers, and worries that the supply of high-quality human-generated text needed to keep scaling up may eventually run out. There is also real disagreement about whether capabilities sometimes appear suddenly and unpredictably as models grow — so-called "emergent abilities" — or whether such apparent jumps are partly an illusion created by how researchers choose to measure performance. You'll explore these debates in more depth later in the course; for now, the key point is that bigger has generally meant better, but nobody is certain how much further that trend will hold.
Most of today's headline AI systems — the large language model (LLM) chatbots you may have already used — are built on a particular neural network design called the transformer. You'll learn about the transformer in detail later, but the throughline to hold onto now is this: a transformer is still, fundamentally, a deep neural network made of layers, neurons, and weights, trained on enormous amounts of text through the same basic learning-from-mistakes process described above. What makes transformers distinctive is how they handle context — the ability to weigh the relevance of every word in a passage against every other word — which turns out to be extraordinarily effective for language. Scale this architecture up with huge amounts of data and computing power, and you get the conversational, often strikingly fluent systems that have become part of daily life.
Deep learning is not a settled, fully understood technology — it's an active and sometimes contentious area of science. Later in the course, you'll explore several genuine, unresolved debates among experts, including:
You don't need answers to these questions to get value from this course. Part of understanding deep learning honestly is understanding that some of the smartest people working on it are still arguing about fundamentals — and by the end of this course, you'll be equipped to follow, and even participate in, those conversations.
To summarize the map you'll be filling in across this course:
The conceptual path this course follows: basic building blocks and training combine with depth and scale to form "deep learning"; applied to the transformer architecture, this produces the large language model chatbots in everyday use.
Keep this map in mind as you move through the rest of the course — each concept that follows fills in one piece of this picture in more depth.