A weight is an adjustable numerical coefficient that controls how strongly an input contributes to a calculation. In a neural network, each node uses weights to combine incoming values; the node then adds a bias and applies an activation function. section 11.2.1
A weight’s value shapes the influence of its input:
- A large positive weight increases that input’s influence in one direction; a negative weight reverses its influence; a weight near zero gives it little influence. section 11.2.1
- During training, backpropagation computes the loss gradient with respect to each weight, and an optimization algorithm uses the gradients to update the weights. section 11.2
In a linear decision function, the weight vector multiplies the input features; its components set their contributions, while the bias shifts the boundary. section 8.2.1
From image patterns to representations
In an image network, weights help determine how inputs contribute to layer responses. Layers can represent simple visual patterns and combine them into shapes and objects. section 11.2.1







