Holes in the Neural Network

A powerful data-analysis tool arrives on the Windows desktop but needs improving to be ready for prime time

Ben Smith

If computers were only as smart as their creators, they would all have neural-network software and could solve the world's most complex problems. However, even the fastest computers aren't as powerful as a mere fruit fly. To believe that software running on even the most recent Intel processor can do more than a rough analysis of complex data sets is like believing that the tooth fairy will bring peace in Bosnia. Yet, there's an important place for SPSS's Neural Connection 1.0 in the toolbox of statisticians.

The problems that Neural Connection addresses are old ones: data segmentation/classification, categorization, prediction, and time- series analysis. When the data falls into simple curves and clusters, more traditional statistical methods will do fine for dete rmining curves and equations; but when data scatters wildly or the curves defy mathematical description, you need neural-network algorithms.

Neural Connection provides you with three neural-network models: a multilayer perceptron, the radial basis function, and the Kohonen network. Because the order and content of the training data are so important to the success of neural networks, Neural Connection provides tools for viewing, filtering, combining, and generating your data. Additionally, it gives you tools for looking at, formatting, and graphically viewing your results. All this is wrapped in a Windows-based GUI (see the screen). You can design your data flow by arranging and connecting the various tools in the Neural Connection work space.

We evaluated Neural Connection by feeding it 2000 random points on the curve for the declinatio n of the moon. This curve is roughly predictable with just a few trigonometric factors, but in fact it's highly complex--the equation has more than 50 elements in it. A 486/50 PC cranked away, training the neural network for more than 4 hours. Neither the software nor we were happy with the results.

We then gave it a simpler problem, asking it to predict points on a sine wave. Again, Neural Connection took more than 4 hours. This time, however, the resulting neural network had some value, though we wouldn't use it for drawing a circle.

This is software for statisticians, not engineers, for discovering loose order in apparent chaos, not for developing precise empirical predictions. In an engineering context, it would be more useful in fluid dynamics than in mechanics. In theory, Neural Connection should be helpful with market analysis and research, financial research, and their associated predictive needs. If any technique is able to improve a direct-marketing campaign's cost-effectiveness by eve n a few percentage points, it's valuable.

Even with a real need for Neural Connection, however, it may not be the product to fill that need--not, at least, the present version. We found the GUI difficult to navigate and nonintuitive to configure. The training software, which uses the scripting language NetAgent, was bug-ridden and offered no simple escape.

The depth of neural network required for valuable work exceeds what it's reasonable to do on a PC. Serious data analysis on this scale needs a stream-oriented data flow. Engineers and statisticians with the technical expertise to build appropriate data sets for training a neural network aren't likely to be attracted by a cutesy, icon-based interface for building relatively simple data-processing paths. Finally, the software is unreasonably expensive for learning and experimentation.

---

PRODUCT INFORMATION

Neural Connection 1.0.........$995

SPSS, Inc.
Chic
ago, IL
Phone:    (800) 543-2185 or (312) 329-3500
Fax:      (312) 329-3668
Internet: 
http://www.spss.com

Circle 1111 on Inquiry Card.

HotBYTEs
 - information on products covered or advertised in BYTE

---

Graphic Network Representation

screen_link (66 Kbytes)

These two screens show the visual programming style of Neural Connection, along with two of the graphical ways it can repre sent the resulting neural network.

---

Ben Smith is a computer consultant, a former BYTE Lab testing editor, and the author of Unix Step-by-Step (Hayden Books, 1990). You can contact him on the Internet at ben@ronin.com .