A neural network is a type of machine learning model composed of layers of simple computational units, called neurons or nodes, that are connected to one another with adjustable strengths called weights. Each neuron takes in numbers, multiplies them by weights, adds a bias, and passes the result through a nonlinear function. By stacking many such layers, a neural network can learn to map complex inputs, like pixels, words, or audio waveforms, to outputs such as class labels, translated sentences, or generated images.
How a Neural Network works
During training, the network is fed examples (for instance, thousands of photos labeled "cat" or "dog") and produces a prediction. A loss function measures how wrong that prediction is, and an algorithm called backpropagation calculates how each weight in the network contributed to the error. An optimizer, typically a variant of gradient descent, then nudges every weight slightly to reduce the error. Repeating this process across many examples causes the network's weights to settle into values that capture useful statistical regularities in the data.
The depth of a network matters: the first layers tend to learn simple features such as edges or letter strokes, while deeper layers combine those features into richer concepts like shapes, words, or objects. This hierarchy of representations is what makes deep neural networks so effective on perception-style tasks. A widely cited overview of the architecture and learning algorithm is available in LeCun, Bengio, and Hinton's 2015 Nature review of deep learning.
Why it matters
Neural networks underpin most of the AI capabilities that have become mainstream in the 2020s, including image classification, speech recognition, machine translation, recommendation systems, and large language models such as the models behind conversational assistants. They excel at problems where hand-written rules are brittle but large amounts of labeled or unlabeled data exist, because the same architecture can be retrained for new domains with relatively little code change.
Key types
- Feedforward neural network (FNN): The simplest form; signals move in one direction from input to output. A multilayer perceptron is the canonical example.
- Convolutional neural network (CNN): Uses shared-weight filters, ideal for images and video.
- Recurrent neural network (RNN): Has loops that retain a memory of prior steps, suited to sequences such as text or sensor data; largely superseded by transformers for language.
- Transformer: A modern architecture based on attention rather than recurrence. It is the backbone of today's large language models and many vision systems.
- Generative adversarial network (GAN): Pairs a generator with a discriminator that learns to tell real from fake samples, used for image synthesis.
Each variant rearranges or specializes the basic neuron-and-weight recipe to suit a particular kind of data, but the underlying principle — learning weights by gradient descent on a loss — remains the same.