liveBook · Manning

Parameter

As explained in Fight Fraud with Machine Learning

A model parameter is a value learned from training data that determines a machine learning model's behavior.

A parameter is a value learned from training data that determines a machine-learning model’s behavior. In neural networks, parameters are commonly weights and bias terms adjusted during training. section 4.1.2 section 3.1

Counting parameters

A four-input logistic regression model has five parameters: one weight per input plus a bias term. Adding a hidden layer with 25 units creates 4 × 25 additional connections, increasing the total to more than 100. section 4.1.2

Model size can grow from a few parameters to millions, billions, or trillions in deep neural networks. Training larger parameter sets requires substantial computation and generally benefits from larger volumes of training data. section 4.1.2

Distinguish parameters from related model settings:

  • Parameters are learned from data during training. section 3.1
  • Hyperparameters are supplied by the developer, including learning rate, maximum tree depth, and class weights. section 3.1
  • Gradient descent updates parameters during model training. section 4.1.2
Figure 4.7 Logistic regression model expressed here as a neural network. Notice the resemblance with figure 4.3.

Compare logistic regression’s weighted-sum structure with its equivalent single-layer neural-network representation. section 4.1.2

How Fight Fraud with Machine Learning covers it

How deep it goes
Explains parameter counts, learned-versus-supplied settings, and scaling to large models without detailing optimization internals.
Examples
Uses four-input logistic regression, a 25-unit hidden layer, and models ranging from millions to trillions of parameters.
What you can do after
Evaluate/understand parameter counts and distinguish learned parameters from developer-specified hyperparameters.
How it connects
Connects parameters to logistic regression, neural networks, hyperparameters, gradient descent, deep learning, and training data.

Also explained in

Compare how 41 books explain parameter →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Fight Fraud with Machine Learning. The section links point to the parts of the book it draws on.