liveBook · Manning

Linear transformation

As explained in Grokking Bayes

A linear transformation is a mapping whose output is a linear combination of its inputs, possibly including an additive bias in an affine implementation.

A linear transformation maps inputs to outputs by combining input values with learned coefficients. In a neural network, edges between consecutive layers supply those coefficients, called weights. section 10.1

A practical layer also includes a bias: y = mx + b. Weights and biases are network parameters, and stacking transformations with nonlinear activations lets the network represent complex mappings. section 10.1

From edges to a prediction

In a network with hidden layers, the collection of edges from one layer to the next defines a transformation. Each edge weight contributes a coefficient, while the bias shifts the resulting value before the next activation. section 10.1

Figure 10.2 The architecture of an example neural network with two hidden layers of size 6 and 4.

Use the architecture diagram to trace how transformations connect input, hidden, and output layers. section 10.1

How Grokking Bayes covers it

How deep it goes
Defines weights, biases, and layer connections but stops short of deriving the transformation or implementing it in code.
Examples
Uses a y = mx + b transformation and a multilayer architecture with hidden layers of sizes six and four.
What you can do after
Understand how weights and biases parameterize layer-to-layer mappings in a neural network.
How it connects
Connects linear transformations with neural networks, weights, biases, hidden layers, and nonlinear activation functions.

Also explained in

Compare how 3 books explain linear transformation →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Grokking Bayes. The section links point to the parts of the book it draws on.