neural network

A neural network is a computational model composed of layered, interconnected units (neurons) that learn input-to-output mappings by adjusting weighted connections and applying nonlinear activation functions.

Training typically consists of a forward pass that produces predictions and a loss/objective, followed by gradient-based optimization: gradients are computed via backpropagation (automatic differentiation using the chain rule) and used to update parameters.

Architectures vary by data and task, including the following:

• Feedforward multilayer perceptrons
• Convolutional networks for images and other grid-structured data
• Recurrent networks for sequences
• Transformer-based models that use self-attention to capture long-range dependencies
• Graph neural networks for graph-structured data

Despite the biological inspiration, modern neural networks are mathematically defined, differentiable function approximators whose parameters (weights and biases) are learned from data by minimizing a loss under supervised, self-supervised, unsupervised, or reinforcement learning setups.