Neuro Developer Kit Library (NDK)

NEURO DEVELOPER KIT (NDK)

A LIBRARY FOR NEURAL NETWORKS AND NON-LINEAR MODELS

NEURO DEVELOPER KIT is a library in Windows DLL format, available in free of charge and paid versions, that gathers all the necessary instructions to the design, the training, the selection and the simulation of polynomial models, neural networks, explicit non-linear models and computer codes.

The Neuro Developer Kit library has universal capabilities in regression statistics : it works with all linear and nonlinear regression models as well as all neural networks described by multi-layer perceptrons :

Linear models, polynomial models,
Standard neural networks,
Neural networks and polynomial models for mixture (formulation in chemistry),
Neural networks for classification,
Dynamic neural networks,
Kohonen maps,
Explicit nonlinear models,
Compiled models (dll format dll : proprietary code, differential equations, etc...).

The Neuro Developer Kit library is easy to use and powerful at the same time :

Training algorithm accurate by 10 digits,
Calculation of confidence intervals on the parameters (coefficients) and the predicted outputs,
Manual / automatic selection of the best model,
Simulation in direct mode direct and in reverse mode,
Exclusive NETRAL's script that fully automise the various steps in the calculation, resulting in a saving of time and in an improved efficiency.

The Neuro Developer Kit library can be interfaced with all program languages. NETRAL can assist you if you use one of the following languages : C, C++, Delphi, python, Matlab, Maple.

The Neuro Developer Kit library is distributed in 3 different versions with increasing features and price :

NEURO DEVELOPER KIT FREE : free of charge version without technical support : design of linear, polynomial and neural models, training, display of the results, selection of the best model, simultion in direct mode.
NEURO DEVELOPER KIT SILVER : intermediate version without technical support : NDK FREE + record of the model and the result, loading of previously saved models, export to Excel and C code.
NEURO DEVELOPER KIT GOLD : full features version with technical support : NDK SILVER + dynamic models, confidence intervals, ponderation (weights), complex cost functions, simulation in reverse mode, merge of models, script for one-shot automatic calculation, export of C code with confidence intervals.

We have designed the free version for education purpose and for the validation of the first study of feasability, which means to demonstrate to oneself and to the colleagues that neural networks are the appropriate tools for the given problem. We draw your attention to the fact that, without confidence intervals on the predictions and a measure of the possible degeneration of the model, neural networks can return whimsical or even dangerous results.

We invite you to read the reference manual of the library NEURO DEVELOPER KIT, test the free of charge version which includes several examples in python language, and then, according to the value of your application, to consider the paid versions and the technical support provided by NETRAL.

Download the free of charge and the commercial versions of NEURO DEVELOPER KIT.

Freely use NEURO DEVELOPER KIT FREE.

Buy NEURO DEVELOPER KIT SILVER with a classical purchase order or through our on-line boutique.

Buy NEURO DEVELOPER KIT GOLD with a classical purchase order or through our on-line boutique.

Discover the reference manual related to NEURO DEVELOPER KIT, the instruction list and several examples in python.

Professor, teacher contact us if you wish to use NEURO DEVELOPER KIT in your course and computer exercices.

Please contact us if you are interested in NEURO DEVELOPER KIT and wish a demo.

Features	NDK Free ! (Level 0)	NDK Silver (Level 1)	NDK Gold (Level 2)	Details
Basic Features
First order linear models Second and third order polynomial models Kohonen maps Standard static neural networks Neural networks for classification Mixture models (reduced polynoms, reduced NN) Note : Models with or without normalisation Activation functions : hyperbolic tangeant, sigmoïd
Dataset loading : Ascii files (txt, csv)
Training set, Validation set Through partition in 2 blocks of the dataset Through a random sampling within the dataset
Training algorithm Levenberg-Marquardt (accurate by 10 digits) BFGS Quasi-Newton (accurate by 8 digits) Simple gradient
Training results Cost function, Standard deviation on training set, unbiased standard deviation on training set, standard deviation on validation set, R, R2
Manual selection of the best result
Use of the selected model : simulation, direct calculation Note : as long as the computer is on and the call to the DLL is activated
Intermediate features
Result recording - Model recording
Loading a model recorded by the user
Analytical description of the model in Excel code and C code Export of the transfer function of static models without confidence intervals
Advanced features
Standard dynamic neural networks (loop from the output to the inputs, loop of internal states) Non-standard neural networks (static or dynamic) Knowledge-based models Activation functions : Hyperbolic tangeant, Sigmoïd, Arctangeant, Sine, Cosine, Gaussian, Quadratic, Cubic, Exponential, Logarithm, Inverse, Square root, Hyperbolic arcsine, Difference versus 1 Compiled models
Ponderation (weight) of the examples in training dataset Line by line, External vector of weights, External model of the noise, Through dedicated cost function
Cost functions Ordinary least square, weighted least squares, Crosses entropy, Delta-Log, .... tailored cost function can be used
Training with leverage and confidence intervals (Symetric confidence intervals are estimated at the first order)
Training results Additional results to NDK Free : PRESS (predictive residual error on sum of squares), Rank, Determinant and Conditioning value of the dispersion matrix
Automatic selection of the best result
Merge of single output models into a multi-output model with confidence interval on each output
Generation of C code representing the analytical model Export of the transfer function of static and dynamic models with confidence intervals
Script with a complete automation of the successive steps : design of the various models, training, selection, record, and display of the best result An impressive instruction. A NETRAL exclusivity !

A few applications :

LISA laboratory, from Université Paris XII - Créteil, has developed an application to predict atmospheric pollution above the city of Orléans 24 hours in advance. Daily and hourly measures collected in different points of the city feed the database. The model coefficients are reajusted every three months (further information on this article).
Arcelor has equipped the rolling mill of his factory in Dunkerque with a neural computation system to calculate the steel hardness of the steel roll. The inputs for this model are, among others, chemical composition, width, thickness and temperature. A calculation is made for each steel roll. The application built with the NDK undergoes regular relearning to take into account the new steel grades and the drift of the rolling mill. A specific algorithm based on the k-means has been added on top of neural networks to refresh the database and to keep it at a constant size.