At the forefront of Artificial Intelligence

Home » Reviews » Software » Neural Networks

NeuroDimensions NeuroSolutions 3.02

Neural networks (NN) are finally getting recognized as viable solutions for a range of commercial applications, and not just researchers play toys. This left the market open for NN development tools - NeuroDimension's NeuroSolutions has been hailed as being at the very forefront of this market. With the release of version 3.0 of NeuroSolutions a range of new features including OLE automation, a macro language and more.

Features

One of the best things about NeuroSolutions (NS) is that all of the really cool features are described in the interactive tutorials that greet you when you start up NS. How to construct a plethora of networks, the macro functionality, C++ code generation, OLE automation and DLL addon capability. The tutorials are well organized, and the interactivity allows the users to play with the tutorial, not just idly sit by and watch diagrams and text. The tutorials introduce you to the mass of probes NS has.

Probes

In NS, probes are drag-and-drop items that you can use to "look into" the network. NeuroDimension has given you just about every kind of representation of a network you could want with the probes. Be it numbers, line graphs, a greyscale graphical representation, Hinton diagrams, even a 3D representation! To the right is a 3DProbe.

The probes are an excellent way of taking the complex results of a neural network and turning them into more usable, or directly visible, results. Whether you need graphical representations, pictures, bar graphs or schematics, NS allows you to place probes at any point in your network.

The interactive demonstrations the NeuroDimensions show how to develop networks as simple as a multilayer perceptron, self-organizing networks, to Jordan-Elman networks, RDF, hybrid and time-lagged networks! The demonstrations use real world examples, ranging from spectral extraction, predicted chaotic series, feature extraction and image recognition. Below is another screenshot of some of the probes during the Linear Associator/Image Recognition demonstration:

The demonstrations (which, incidentally, are powered by NeuroSolutions macro language) build the network in front of you, and provide step-by-step examples of how you would use NeuroSolutions to train such a network to perform your tasks. The demonstration that accompanies the above screen shot shows how the network can map a 48x48 image of a face to a picture of the name (in this instance, Curt). The demonstration allows you to train the network, then step through each exemplar to see the network in full working order!

Interface

It must be stressed that the networks mentioned above are not the only networks that NS can handle. Theoritically, you may build any type of conventional neural network with the interface provided by NeuroSolutions. The interface is a classic drag-and-drop, highly object-orientated development environment. This makes the interface both intuitive and functional. Networks can be created from scratch, or can be created using the NeuralWizard that helps build the network you need quickly - these networks can be used immediately, or you can use it as a 'springboard' to help hone in your specific requirements quicker. Below is a screenshot of a modular neural network:

To give you more of an idea of the interface, here is an extract from the "Logic of the Interface" help topic:

Each neural component is represented by an icon.
Single-clicking on a component's icon will display its Inspector window. A component's inspector is where the user can inspect and alter any variables of the component.
Double-clicking on a component's icon will open its animation window, if such a window exist. Animation windows allow components to display data while a simulation is running.
Single-clicking on a component's icon with the help cursor will display its on-line help quick reference page.
Networks are constructed by placing and interconnecting components on a breadboard.
Components are created by selecting them from palettes and stamping them onto the breadboard. Components can be removed from the topology by means of the cut operation.

Features and Versions

NeuroSolutions has a huge range of additional features - since it allows for OLE automation, you can use it to power Excel spreadsheets or other applications. Custom Solution Wizard will take an existing breadboard (an NS document) and convert it into a dynamic-link library that can then be imported into your own applications! The higher versions can even produce ANSI C++ code that can be copied into your applications.

The complicated aspect of NeuroSolutions is the versions. There are six versions of NeuroSolutions in all: Educater, Users, Consultants, Professional, Developer Lite and Developer. Each level has certain restrictions - again, here is a quote from the help file that best describes the setup:

There are six levels of NeuroSolutions, all of which allow you to implement your own neural models. The Educator, our entry level version, is intended for those who want to learn about neural networks and work with MLPs. The Users version extends the Educator with a variety of neural models for static pattern recognition applications. The Consultants version offers enhanced models that support dynamic pattern recognition, time-series prediction and process control problems.
The Professional version adds ANSI C++ compatible code generation, allowing you to embed NeuroSolutions’ algorithms into your own applications (including learning). Furthermore, this version allows any simulation prototyped within NeuroSolutions to be run on other platforms, e.g. faster computers or embedded real time systems. The Developer versions allow you to extend the functionality of NeuroSolutions by integrating your own neural network, preprocessing, control, and input/output algorithms.

Conclusion

NeuroDimensions has a sure winner with NeuroSolutions - with its large range of features, speed, and unparalleled flexibility NS is a sure solution to any NN developer's requirements. If the price tag can be met, don't look any further.

8.9

Price: Ranges from $195 to $1,995

Liked: Huge range of features, flexibility and interface.

Disliked: Complications that arise with the different versions.

Website: http://www.nd.com/

Submitted: 07/04/2000

Article Toolbar

BibTeX entry