A parametric machine-learning model separates its fixed functional form from a collection of adjustable numbers — the weights that scale each input or intermediate signal and the biases that shift it — which are collected into a single parameter vector. The core principle is that learning changes only these numbers and never the model's structure, so the entire space of models the learner can reach is coordinatized by that vector. This framing is what allows a model to be treated as a point in a continuous high-dimensional space that an optimization procedure can move, as distinct from hyperparameters, which are set outside training and are not adjusted by it.
In machine learning theory, weights and biases constitute the parameters defining a linear transformation within models such as linear regression or neural networks. Weights function as scalar multip…