Product Units: A Computationally Powerful and Biologically Plausible Extension to Backpropagation Networks

  • Durbin R
  • Rumelhart D
  • 60

    Readers

    Mendeley users who have this article in their library.
  • N/A

    Citations

    Citations of this article.

Abstract

Introduces a computational unit for feedforward learning networks of the backpropagation type, which calculates a weighted product in which each input is raised to a power determined by a variable weight. The unit can learn to represent generalized polynomial terms in the inputs and can form a better representation of data in cases where higher order combinations of inputs are significant without increasing the number of free parameters. Product units can be trained using gradient descent, allow simpler solutions to standard learning problems, have a higher empirical learning capacity than summing units, and act to create a hidden layer representation for an output summing unit. A neurobiological interpretation is presented for this combination of product and summing units in terms of a single neuron.

Get free article suggestions today

Mendeley saves you time finding and organizing research

Sign up here
Already have an account ?Sign in

Find this document

Authors

  • Richard Durbin

  • David E. Rumelhart

Cite this document

Choose a citation style from the tabs below

Save time finding and organizing research with Mendeley

Sign up for free