Computer Science I / Systems of Linear Equations
Practice question · Multiple choice

A neural network layer is a matrix multiply followed by a non-linear function. Why is the non-linearity essential rather than decorative?

Hints
  1. Multiply two matrices. Is the result a matrix?
  2. Ask what a network of pure matrix multiplies could represent that a single matrix could not.
Show the answer

A. Because a composition of linear maps is itself linear

Why

AB is a matrix, so composing linear layers collapses to one, depth without non-linearity is a very expensive way to compute a single linear map. ReLU or similar breaks the collapse, and that is the whole reason deep networks can represent what shallow ones cannot.

Read the lesson: Systems of Linear Equations →

Practise Systems of Linear Equations

The app has 5 more questions on this lesson, and keeps your place in the course. Computer Science I is free to start.

More questions on Systems of Linear Equations