Perceptron Learning Revisited: The Sonar Targets Problem

Authors:
Martina Hasenjäger;Helge Ritter
Affiliations:
Universität Bielefeld, Technische Fakultät, Postfach 10 01 31, D–33501 Bielefeld, Germany, e-mail: {pmhasenj,helge}@techfak.uni-bielefeld.de;Universität Bielefeld, Technische Fakultät, Postfach 10 01 31, D–33501 Bielefeld, Germany, e-mail: {pmhasenj,helge}@techfak.uni-bielefeld.de
Venue:
Neural Processing Letters
Year:
1999

Citing 4
Cited 3

Accelerated learning in layered neural networks

Complex Systems
Support-Vector Networks

Machine Learning
Playing billiards in version space

Neural Computation
Characterization of the Sonar Signals Benchmark

Neural Processing Letters

The Minimum Number of Errors in the N-Parity and its Solution with an Incremental Neural Network

Neural Processing Letters
Learning Perceptrons and Piecewise Linear Classifiers Sensitive to Example Dependent Costs

Applied Intelligence
Boosting interval based literals

Intelligent Data Analysis

Quantified Score

Hi-index	0.00

Visualization

Abstract

Recently it was pointed out that a well-known benchmark data set, thesonar target data, indeed is linearly separable. This fact comessomewhat surprising, since earlier studies involving delta ruletrained perceptrons did not achieve the separation of the trainingdata. These results immediately raise the question of why a perceptronwith a continuous activation function may fail to recognize linearseparability and how to remedy this failure. The study of these issuesdirectly leads to a performance comparison of a wide variety ofdifferent perceptron training procedures on real world data.