Network
Technology
|
Neural Network Technology |
NeuralystTM Run-time Library |
The NRL provides a straight-forward method of implementing the backpropagation neural network, the most popular and generally useful neural network algorithm, within an application. Routines for data and neural network set-up, neural network training and neural network prediction are provided.
The NRL supports the same optimized neural network computational engine used in Neuralyst. NRL can be set to use 16-bit fixed point arithmetic for the highest performance, without the need for a floating point co-processor, with no loss of neural network capability. In fact, Cheshire Engineering benchmarks show that fixed point mode runs 3-4 times faster than the corresponding floating point mode. A full 64-bit floating point mode, which requires a floating point co-processor, is also available when full precision is required.
Prepare your data as arrays of numeric or symbolic input types and output types within your application. Then call NRL routines to set-up the neural network and other routines to train it. Once a neural network has been trained or a pre-trained neural network has been loaded, additional NRL routines can be called to evaluate new data and make predictions from those.
If you have an IBM-PC or compatible, then a compiler and linker consistent with Microsoft C and Windows/DLL calling conventions (assembly language programmers can simply adhere to such conventions) and object formats is required.
If you have a Macintosh, then a compiler consistent with Think/Symantec C calling conventions is required.
The combination of Neuralyst and the NRL makes an ideal neural network development system. Use Excel and Neuralyst to train your neural network, using the full scope of Excel's powerful capabilities to prepare, analyze and automate the process.
When you are satisfied with the performance of the neural network, you can embed the trained network into your application with the NRL. Once embedded, the extensive training process and training data set will be impossible to reverse engineer.
Neural Network Engine - is a highly optimized subroutine package which implements the back-propagation neural network algorithm with selectable activation functions. Computation can be selected to occur in 16-bit fixed point arithmetic or 64-bit floating point arithmetic. The engine also includes the Genetic Supervisor for neural network optimization.
Data Types - may be any numeric magnitude represented by a C floating point number. Conversion to fixed point, scaling and normalization is automatically performed by the neural network engine interface. Symbolic data can be entered as text labels.
Input and Output Limits - The total amount of space available for inputs or outputs is determined by the combination of physical memory, memory model, segment size limits and processor capabilities for the target system.
Neural Network Capacity Limits - the neural network may be configured with as many as six layers. The neural network engine will support over 130,000 neural connections. The total number of neurons is limited only by the number of neural connections.
Windows Neural Network Performance - For a 60 MHz Pentium: training mode performance can exceed 440,000 connection updates per second in fixed point; predictive mode performance can exceed 1,360,000 connections per second in fixed point.
Macintosh Neural Network Performance - For a 25 MHz Mac Quadra: training mode performance can exceed 200,000 connection updates per second with fixed point; predictive mode performance can exceed 650,000 connections per second in fixed point. A native PowerMac version is available with even greater performance.
License & Royalty Terms - the NRL license includes the non-exclusive, non-transferrable right to distribute software using the NRL, developed by the licensee, without additional royalty. See license agreement for specific details and/or restrictions.