Producing evidence for the hypotheses of large neural networks

This paper presents results allowing large networks to be analysed using a fast approximate system. This provides an understanding of their hypotheses and actions, permitting networks to be used with more confidence in complex applications. The technique used has been to calculate the best input pat...

Full description

Saved in:
Bibliographic Details
Published inNeurocomputing (Amsterdam) Vol. 10; no. 4; pp. 359 - 373
Main Authors Fletcher, G.P., Hinde, C.J.
Format Journal Article
LanguageEnglish
Published Elsevier B.V 24.04.1996
Subjects
Feedbackward
Feedforward networks
Hypothesis
Interpretation
Feedforward networks
Model
Interpretation
Feedbackward
Hypothesis
Online AccessGet full text

Cover

More Information
Summary:This paper presents results allowing large networks to be analysed using a fast approximate system. This provides an understanding of their hypotheses and actions, permitting networks to be used with more confidence in complex applications. The technique used has been to calculate the best input pattern that given the internal hypothesis gives the desired output. If the calculated pattern correctly displays the features of the input space then this is evidence that the network has learnt the correct hypothesis. The converse is also true: if the shape differs from the target then this should provide some indications for the best training examples to improve the hypothesis.
Bibliography:ObjectType-Article-2
SourceType-Scholarly Journals-1
ObjectType-Feature-1
content type line 23
ISSN:0925-2312
1872-8286
DOI:10.1016/0925-2312(94)00064-6