Neural Network

A model made of many connected artificial neurons that learns by adjusting the strength of those connections.

A neural network is a class of machine-learning model built from many simple computing units, artificial neurons, connected together: each unit takes a weighted sum of its inputs and passes it through a nonlinear function, the activation function, into the next layer. The strength of a connection is its weight, which is the model's parameter. Training measures the gap between the output and the correct answer with a loss function, then uses backpropagation to compute which direction to nudge each weight, repeating until the error shrinks. A network with at least two hidden layers between its input and output layers is generally called a deep neural network. AlexNet's decisive win over traditional methods at the 2012 ImageNet competition kicked off the deep-learning boom. Today's vision encoders, Transformers, VLAs, and reinforcement-learning policies in embodied AI are all, underneath, neural networks.

ExampleA common reinforcement-learning walking policy for a legged robot is often just a multilayer perceptron a few layers deep: it takes in joint angles, IMU readings, and a velocity command, and outputs a target angle for every joint.

Also called

Deep Neural Network, Artificial Neural Network, ANN, DNN

Related

Multilayer Perceptron · Convolutional Neural Network · Transformer · Backpropagation · Parameter Count (Model Size) · Loss Function

Sources

Neural network (machine learning) (Wikipedia)