Link to the University of Pittsburgh Homepage
Link to the University Library System Homepage Link to the Contact Us Form
D-Scholarship @ Pitt Banner and Links to Homepage

Visualizations of 2-D Hidden Unit Space

Munro, Paul W (1991) Visualizations of 2-D Hidden Unit Space. Technical Report. School of Library and Information Science, University of Pittsburgh, Pittsburgh, PA.

[img] Plain Text (licence)
Available under License : See the attached license file.
Download (1kB)

Abstract

Solution of nonlinearly separable Booklean tasks by multilayered feedforward networks is generally accomplished by nonlinearly transforming the space to a representation that permits linear separations of the categories by the output units. For networks with just two hidden units, these representations can be plotted in a plane. The classification boundary imposed by the weights and bias of each output unit can be represented by a line in the same plane. Solutions to a variety of tasks, as determined by the backpropagation algorithm, have been visualized using this technique. The resulting images reveal several properties of the hidden unit representations achieved by backpropagation that have not yet been reported; these include that the normal solution to XOR "collapses" the pattern space to a one dimensional manifold (the mapping is singular) and the high symmetry of the hidden unit patterns achieved in the N-2-N encoder task. The mapping computed by a multilayer feedforward network is thus analyzed not using weight diagrams, but rather, in terms of the relationships between pattern representations in the hidden layers.

Share

Citation/Export:
Social Networking:
Share |

Details

Item Type: Monograph (Technical Report)
Status: Published
Creators/Authors:
CreatorsEmailPitt UsernameORCID
Munro, Paul Wpwm@pitt.eduPWM
Monograph Type: Technical Report
Date: March 1991
Date Type: Publication
Access Restriction: No restriction; Release the ETD for access worldwide immediately.
Publisher: School of Library and Information Science, University of Pittsburgh
Place of Publication: Pittsburgh, PA
Institution: University of Pittsburgh
Department: School of Library and Information Science
Schools and Programs: School of Information Sciences > Library and Information Science
Refereed: No
University of Pittsburgh Series: iSchool Research Report Series
Other ID: LIS035/IS91003
Date Deposited: 10 Apr 2013 18:56
Last Modified: 01 Nov 2017 14:02
URI: http://d-scholarship.pitt.edu/id/eprint/18222

Metrics

Monthly Views for the past 3 years

Plum Analytics

Actions (login required)

View Item View Item