A single artificial neuron, the basic building block of a neural network, performs an almost embarrassingly simple operation: it takes several numeric inputs, multiplies each by an adjustable weight, adds them together along with a bias term, and passes the result through a simple function to decide what to output. On its own, one such unit can barely do anything useful. Stack thousands or millions of these simple units together in interconnected layers, though, and the resulting network can learn to recognise faces in photographs, translate between languages, or generate coherent text — capabilities that emerge from the collective arrangement, not from any single neuron being individually clever.
Layers that build up increasingly abstract features
In a typical neural network built for a task like image recognition, information flows through successive layers of these simple artificial neurons, and each layer tends to detect progressively more abstract patterns built from the patterns the previous layer identified. An early layer might respond to simple local features like edges or colour gradients in small patches of an image; a middle layer might combine those edge-detections into recognisable shapes like curves or corners; a later layer might combine those shapes into recognisable parts, like an eye or the outline of a nose. No one explicitly programs any individual neuron to detect "an eye" — that capability emerges from how the weights connecting neurons across all these layers get adjusted during training, not from any hand-designed rule specifying what each layer should look for.
Learning means adjusting weights until errors shrink
Training a neural network means repeatedly showing it examples with known correct answers, checking how far off its actual output was from the correct one, and then adjusting the weights throughout the network, layer by layer, in the direction that would have reduced that specific error — a process called backpropagation combined with gradient descent. Repeated over enormous numbers of training examples, this gradual weight adjustment causes the network to settle into a configuration where its layered pattern-detectors, collectively, reliably produce correct outputs for inputs similar to what it was trained on. The network never receives an explicit rule like "eyes are round and located above a nose" — it discovers useful internal representations purely by having its weights nudged, incrementally, toward configurations that reduce prediction errors across a large training dataset, which is precisely what makes the resulting learned patterns often difficult for a human to fully interpret or explain, even though the network reliably produces correct results.
What we're still unsure about
The basic mechanics of how neural networks are structured and trained through backpropagation and gradient descent are well-understood, extensively documented engineering, not in dispute. What remains a genuinely significant and actively researched open problem is interpretability — understanding precisely what internal representations a large, successfully trained network has actually learned, and why it produces a specific output for a specific input, remains substantially harder than building and training the network in the first place, and this gap between "the network works reliably" and "we can fully explain why it works the way it does" is one of the central unresolved challenges in current machine learning research.
This sits inside Neural Networks & Deep Learning, one of eight topics in AI/ML, one of seven domains in Computer Science, one of seventeen subjects the app can quiz you on.