What is a neural network?

How neural networks work

A neural network is organized into layers of nodes, also called neurons or units. The input layer receives raw data, one node per feature, such as one node per pixel of an image. One or more hidden layers sit in the middle, and an output layer produces the final result, for example a probability for each possible class. IBM describes a neural network as "a machine learning program, or model, that makes decisions in a manner similar to the human brain, by using processes that mimic the way biological neurons work together."

Every connection between nodes carries a weight, a number that scales the signal passing along it, and each node adds a bias term. A node computes a weighted sum of all its inputs, adds the bias, and passes the result through an activation function that decides how strongly the node "fires." Google's Machine Learning Crash Course notes that without a non-linear activation function, a stack of layers would collapse mathematically into a single linear operation, no better than one layer, which is why activations like ReLU (rectified linear unit) are essential.

Data moves through the network in a forward pass: values enter the input layer, propagate layer by layer through the weighted connections and activations, and emerge as an output. The network's entire behavior is encoded in its weights and biases, often millions or billions of them, which is why training, the process of setting those numbers, is the heart of the technology.

How neural networks learn: training and backpropagation

A freshly built network starts with random weights and produces meaningless output. Learning is the process of adjusting those weights so the output becomes useful. During training, the network is shown labeled examples, it makes a prediction, and a loss function measures how far that prediction is from the correct answer.

The core learning algorithm is backpropagation, short for "backward propagation of errors." After the forward pass computes the loss, backpropagation works backward through the network using calculus (the chain rule) to calculate how much each individual weight contributed to the error. An optimization method called gradient descent then nudges every weight a small step in the direction that reduces the loss. Repeating this over many examples and many passes (epochs) gradually shapes the network into an accurate model.

Backpropagation was popularized by the 1986 Nature paper "Learning representations by back-propagating errors" by David Rumelhart, Geoffrey Hinton, and Ronald Williams, which showed that the method could discover useful internal representations in the hidden layers. It remains the foundation on which essentially all modern neural networks are trained.

Types of neural networks

The simplest form is the feedforward neural network, where information flows in one direction from input to output with no loops. A multilayer version, the multilayer perceptron, is the classic building block and is still used for tabular data and as a component inside larger models.

Convolutional neural networks (CNNs) are specialized for grid-like data such as images. They use filters that slide across the input to detect local patterns, edges, textures, shapes, and build up to recognizing objects. CNNs drove the breakthroughs in computer vision. Recurrent neural networks (RNNs), and their gated variants such as LSTMs, process sequences by maintaining an internal state, which made them useful for text, speech, and time-series data.

The dominant architecture today is the transformer, introduced by Vaswani and colleagues in the 2017 paper "Attention Is All You Need." Transformers use a mechanism called self-attention to weigh the relationships between all elements of an input at once, and they underpin large language models such as GPT and Claude. All of these are neural networks; they differ in how their nodes are wired together.

What neural networks are used for

Neural networks power a large share of the AI systems people use daily. In computer vision they handle facial recognition, medical image analysis, and the perception systems of self-driving cars. In language they drive machine translation, chatbots, search, and text generation. In audio they enable speech-to-text, voice assistants, and music tools.

They also sit inside less visible systems: recommendation engines on streaming and e-commerce platforms, fraud-detection models at banks, demand forecasting in supply chains, and protein-structure prediction in biology. AWS notes that neural networks are used across applications from medical diagnosis and targeted marketing to financial predictions and process control.

The common thread is pattern recognition on complex, high-dimensional data, images, audio, text, sensor streams, where writing explicit rules by hand is impractical. Instead of a programmer specifying the logic, the network learns the mapping from examples.

Limitations and risks

Neural networks are data-hungry. They typically need large, well-labeled datasets to reach high accuracy, and their performance degrades on inputs that differ from what they saw in training. They are also compute-intensive: training large models requires specialized hardware such as GPUs and consumes significant energy.

Interpretability is a persistent problem. Because knowledge is distributed across millions or billions of numeric weights, it is hard to explain why a network made a particular decision, the "black box" critique. In high-stakes domains such as healthcare, credit, and law, that opacity is a serious concern, and it has spurred a whole research field of explainable AI.

Neural networks can also absorb and amplify biases present in their training data, and they can be brittle, small, deliberately crafted changes to an input (adversarial examples) can fool an otherwise accurate model. None of this negates their value, but it explains why responsible deployment pairs them with testing, monitoring, and human oversight.

The bigger picture

Neural networks are the structural foundation of modern artificial intelligence. When people talk about deep learning, they mean neural networks with many hidden layers; when they talk about large language models, they mean very large transformer neural networks. Understanding the network, nodes, weights, layers, and backpropagation, is understanding the machinery underneath almost every recent AI advance.

The idea is old, artificial neurons date to the 1940s and 1950s, but three things made it work at scale in the 2010s: far larger datasets, vastly more computing power (especially GPUs), and refined training techniques. That combination turned a decades-old concept into the engine of today's AI boom.

For a reader trying to make sense of the field, the useful mental model is simple: a neural network is a flexible function that learns its own rules from data by repeatedly adjusting weights to reduce error. Everything else, the different architectures, the scale, the applications, is elaboration on that single idea.

Frequently asked questions

Is a neural network the same as artificial intelligence?

No. Artificial intelligence is the broad field of building systems that perform tasks associated with human intelligence. A neural network is one specific type of machine learning model within that field. Most, but not all, of today's headline AI systems are built on neural networks; other approaches, such as rule-based systems and classical statistical models, also fall under AI.

How is a neural network different from deep learning?

Deep learning is the use of neural networks with many hidden layers, deep networks, to learn hierarchical features from raw data. Every deep learning model is a neural network, but the term "deep learning" specifically emphasizes depth: multiple layers that each build more abstract representations. A shallow network with one hidden layer is a neural network but is not usually called deep learning.

What is a node or neuron in a neural network?

A node (or neuron, or unit) is the basic computing element of a neural network. It receives inputs from nodes in the previous layer, multiplies each by a weight, sums them, adds a bias, and passes the result through an activation function to produce an output. That output then feeds nodes in the next layer. A network's ability comes from how thousands or millions of these simple units combine.

Do neural networks actually work like the human brain?

Only loosely. The metaphor of interconnected neurons inspired the design, and both involve weighted signals between simple units. But artificial neural networks are mathematical functions trained by backpropagation and gradient descent, mechanisms with no clear biological equivalent. Neuroscientists caution that real neurons are far more complex, so the brain analogy is a helpful intuition, not a literal model.

How many layers does a neural network need?

There is no fixed number; it depends on the task. Simple problems can be solved with one hidden layer, while complex tasks such as image recognition or language modeling use dozens or hundreds. Adding layers increases a network's capacity to learn intricate patterns but also raises data, compute, and overfitting risks, so architecture is chosen empirically for each problem.

What is backpropagation in simple terms?

Backpropagation is the algorithm that tells a neural network how to correct itself during training. After the network makes a prediction and its error is measured, backpropagation works backward through the layers to calculate how much each weight contributed to that error, then those weights are adjusted slightly to reduce it. Repeated over many examples, this process is how the network learns.

Related

Explainers: Deep learning, LLM, Transformer

Glossary: Deep learning, Machine learning, Backpropagation, Transformer (architecture)

Sources

• IBM Think — What is a neural network?
• AWS — What is a neural network?
• Google — Machine Learning Crash Course: Neural networks
• Rumelhart, Hinton & Williams (1986) — Learning representations by back-propagating errors
• Vaswani et al. (2017) — Attention Is All You Need