Internal subroutines

Subroutines are presented in the order they are used when training the network.

• SUBROUTINE JNSEPA
  Sets the internal parameters and switches in common blocks /JNINT2/ and /JNINT3/. If receptive fields are used, the defined geometries are checked for inconcistencies.

• SUBROUTINE JNHEAD
  Writes a header on file number MSTJN(6).

• SUBROUTINE JNFEED
  Feeds the values in OIN through the network and calculates the values of the output nodes, without writing to OUT.

• SUBROUTINE JNDELT
  Calculates the error from the output nodes and current values in OUT.

• SUBROUTINE JNSATM
  Calculates the saturation measure s in eq. () for each layer. JNSATM is only called if MSTJN(22) 0.

• SUBROUTINE JNCHOP(ICHP)
  Switches on ( ICHP>0) or off ( ICHP<0) fixed precision weights, thresholds and sigmoid functions. If ICHP 0 the current values of the weights and thresholds are chopped to the fixed precision. The bit precision is set by switches MSTJN(28-30).

• SUBROUTINE JNERR(IERR)
  If something goes wrong or if any inconcistensies are encountered during execution, JNERR is called and writes out an error message and stops the execution.

• SUBROUTINE JNCOGR
  Controls the Conjugate Gradient training. It is called from JNTRAL if CG learning has been selected. JNCOGR calls the subroutines JNCGBE and JNLINS.

• SUBROUTINE JNCGBE(BETAK,IOP)
  If IOP = 1, JNCGBE computes the CG momentum term in eq. () and returns it in BETAK.
  If IOP = 0, JNCGBE sets BETAK = 0 and computes the scalar product of the current and previous gradients.

• SUBROUTINE JNLINS
  Performs the line search with an algorithm that is a mixture of golden section search and quadratic interpolation. All parameters used in the line search are stored in the common block /JNINT4/.

• SUBROUTINE JNSCGR
  Controls the Scaled Conjugate Gradient training. It calls the subroutine JNCGBE.

• SUBROUTINE JNTRED and SUBROUTINE JNTQLI
  These routines are taken directly from [23] and are used to diagonalize the Hessian matrix and compute its eigenvectors and eigenvalues.