LU TP 93-29
CERN-TH.7135/94
December 1993

JETNET 3.0 -- A Versatile Artificial
Neural Network Package

Carsten Peterson and Thorsteinn Rögnvaldsson
Department of Theoretical Physics, University of Lund,
Sölvegatan 14 A, S-223 62 Lund, Sweden
Leif Lönnblad
Theory Division, CERN, CH 1211 Geneva 23, Switzerland
Submitted to Computer Physics Communications

• PROGRAM SUMMARY
• LONG WRITE-UP
• Introduction
• Learning in Feed-Forward Artificial Neural Networks
  • The Back-Propagation Family
  • Second-Order Algorithms
    • Heuristic Methods
    • Conjugate Gradients
  • Interpretation of the Results
• Guidelines and Rules of Thumb
  • Choosing the Model and its Parameters
    • ANN versus other methods
    • Choice of ANN Model
    • Number of hidden units
    • Number of hidden layers
    • The activation function
    • Exploiting symmetries
    • Modular methods
  • Choosing the Learning Algorithm
    • Back-propagation
    • Quickprop
    • Rprop
    • Conjugate Gradients
  • Preprocessing the data
  • When to Stop Training
  • Regularization and Pruning
• Practical Implementation Issues
  • Initialization
    • Initial weight values
    • Backpropagation
    • Manhattan
    • Langevin
    • Quickprop
    • Conjugate Gradient
    • Rprop
    • Batch training
  • Training the Network
  • Dynamic Learning Parameters
  • Conjugate Gradient Learning
    • The line search
  • Output
  • Pruning
  • Computing the Hessian
  • Receptive Fields and Shared Weights
  • The Saturation Measure
  • Limited Precision - Hardware Implementations
• Program Components
  • Feed-forward Network (JN)
    • Main subroutines
    • Other user invoked subroutines
    • Internal subroutines
    • Internal functions
  • Common blocks
    • Interface common blocks
    • Internal common blocks
• Restrictions and Technical Information
• Sample Program
• Acknowledgements
• References
• About this document ...