A neural network is a computational system of interconnected artificial neurons organized in layers that learns to recognize patterns, make decisions, and generate outputs by adjusting the strength of connections between neurons during training. Inspired by biological brain structure but operating through mathematical operations on matrices of weights, neural networks are the foundational architecture behind all modern artificial intelligence — from image recognition and speech processing to the Large Language Models that power AI assistants. A typical LLM is a neural network with billions of parameters (connection weights) organized in dozens of layers, trained on trillions of tokens of text data. The "deep" in deep learning refers to networks with many layers, which can learn increasingly abstract representations of data.
Why it matters
Neural networks are the technological foundation that makes modern AI possible. Every LLM, every image generator, every voice assistant, and every recommendation system is built on neural network architecture. Understanding neural networks explains why AI systems require massive computing resources for training (billions of weight adjustments across trillions of examples), why larger models tend to perform better (more parameters capture more nuanced patterns), and why AI can fail in unexpected ways (the network learned patterns that don't generalize). For business leaders evaluating AI capabilities, neural networks explain the fundamental trade-offs: model size versus inference cost, training investment versus capability, and specialization versus generalization.
How it works
A neural network processes data through layers of artificial neurons. Each neuron receives inputs, multiplies them by learned weights, adds a bias term, and passes the result through a non-linear activation function (like ReLU or GELU). The network has an input layer that receives raw data, one or more hidden layers that progressively transform the data into useful representations, and an output layer that produces the final result. During training, the network compares its output to the desired result, calculates the error using a loss function, and propagates that error backwards through all layers (backpropagation) to adjust every weight in the direction that reduces the error. This process repeats across billions of training examples. The key insight is that neural networks learn automatically from data rather than being explicitly programmed — the "intelligence" emerges from organizing billions of parameters that capture patterns in training data.
Example
A fintech company builds a fraud detection system using a neural network with 12 hidden layers. The input layer receives 200 features per transaction — amount, merchant category, time of day, device fingerprint, location, spending velocity, and historical patterns. The hidden layers learn progressively abstract representations: early layers detect basic anomalies (unusual amounts, new merchants), middle layers identify complex patterns (geographic impossibilities, rapid category shifts), and deep layers combine everything into a fraud probability score. The network trains on 50 million historical transactions labeled as legitimate or fraudulent. After training, it processes new transactions in under 10 milliseconds, flagging fraud with 98.5% accuracy — catching patterns that rule-based systems miss while generating fewer false positives. Each of the network's 15 million parameters encodes a tiny piece of the fraud detection knowledge that the system learned automatically from data.