Home Page


- overview

- data structures

- preprocessing

- networks

- classification

- graphing

- examples


NetMaker neural engine features:

Neural network types:

  • MLP - feed-forward multi-layer perceptron;
  • RMLP - recurrent multi-layer perceptron (back-propagation through time with teacher forcing);
  • Cascade-Correlation based on S. Fahlman papers (+simplified version).
  • Training algorithms:

  • standard steepest-descend with momentum term (off-line and on-line training);
  • conjugate gradients:
  • update formula: Polak-Ribiere, Fletcher-Reeves, no conjugation;
  • reset condition: standard, Powell-Beale;
  • scaled conjugate gradient;
  • quick-prop;
  • Levenberg-Marquardt; (still developing, but quite efficient already)
  • regularization: weight decay (+option for excluding biases from regularization);
  • Bayesian Framework (MLP networks, function approximation tasks):
  • network output uncertanties;
  • online optimization of the regularization factor;
  • Hessian calculation modes: exact, approximated (around the net error minimum), finite differences;
  • automated training stop;
  • dynamic network size adjustment:
  • smart insertion of new hidden units;
  • pruning of twin, dead and constant hidden units;
  • Optimal Brain Surgeon (OBS) for weights elimination (based on the original paper by B. Hassibi et al.).
  • All structure modifications are safe to the network state (no error increase should be observed).

    Activation functions:

  • standard sigmoid (logistic);
  • softmax (to be used with cross-entropy error function only);
  • 0-centered sigmoid;
  • hyperbolic tangent;
  • Elliott function (+unipolar version);
  • arcus tangent (scaled to unipolar and bipolar);
  • linear.
  • Error functions:

  • standard MSE;
  • cross-entropy (cooperates with softmax output layer);
  • Pow4 and integrated hyperbolic arcus tangent for improved sensitivity on network error distribution tails;
  • integrated hyperbolic tangent for training data polluted with outliers and gross errors;
  • various asymmetric functions for different costs of sig->bkg and bkg->sig misidentification;
  • sample weighted - which allows χ2 minimization in case of MSE and 1/σ2 weight;
  • user defined.
  • Preprocessing:

  • normalization (scaling) to 0-mean and unitary standard deviation;
  • QSVD / ICA transformations - elimination of redundant data dimensions, better data representation;
  • FFT filtering.

    Standard classification algorithms:

  • kNN (k-nearest neighbors);
  • SVM classification and regression based on LIBSVM v2.89 library;
  • probability density estimation.