Neural Network

A neural network is a computational model composed of interconnected layers of numerical units (neurons) that learn to map inputs to outputs by adjusting connection weights through exposure to training data.

A neural network is a function approximator built from successive layers of simple processing units called neurons. Each neuron computes a weighted sum of its inputs and passes the result through a nonlinear activation function such as ReLU or sigmoid. Data flows forward through the layers — from an input layer, through one or more hidden layers, to an output layer — producing a prediction or learned representation at the end.

Training adjusts the weights through backpropagation combined with gradient descent: an error (loss) is computed by comparing the network's output to the correct target, and that error signal is propagated backward through the layers to update each weight incrementally. Modern frameworks such as PyTorch and TensorFlow automate this process entirely, enabling networks ranging from a few thousand to hundreds of billions of parameters.

Neural networks underpin virtually every major AI capability of the past decade: image recognition (ResNet, EfficientNet), speech synthesis (WaveNet), protein structure prediction (AlphaFold2), and natural language generation (GPT family). Their core advantage is the ability to learn useful feature representations directly from raw data — pixels, audio samples, or text tokens — eliminating the hand-crafted feature engineering required by earlier machine learning approaches.

As of 2026, neural networks span an enormous scale range, from tiny models running on microcontrollers in IoT devices to multimodal systems distributed across thousands of specialized AI accelerators. Research priorities include improving energy efficiency, interpretability of learned representations, and robustness to inputs outside the training distribution.