Performances

Home Page

MLPfit
learning methods
performances
copyrights
documentation
applications

Interfaces
text
code
LabVIEW
PAW
Windows

Download

More info
links
talks
release notes
contact us

The main features of the MLPfit neural network package are the following:

• It can be used both for function approximation and classification purposes
• The code is simple and can be installed on any platform
• All computations are performed in double precision
• Powerful learning methods are available
• The code has been optimized for speed. In particular, a fast version of the sigmoid function is used.

The figure above compares the CPU and the memory used by MLPfit and 3 other NN softwares on a classification task with 14 inputs, 1 hidden layer of 40 neurons, trained with 68900 examples.

The table below shows the time (in seconds) needed to compute 35 10⁶ sigmoids on various platforms (not necessarily recent models). Our implementation is up to a factor 2 faster than the standard computation, while being almost as precise (difference smaller than 10^-15).