Thursday, July 5, 2007

Artiicial neural networks

Artificial Neural Networks

It was often assumed in the early years of neural network research that implementation in special hardware would be required to take advantage of their capabilities. Such hardware, in particular, would probably be analog and involve multiple parallel processing elements and connections between them. However, the tremendous growth in the digital computing power of conventional von Neuman machines has allowed NNW simulations in software to achieve great success in a number of applications. Meanwhile, the development of hardware especially designed for NNWs has been slow and with only modest commercial success. This overview looks at some possible reasons for this slow development and some of the areas where hardware NNWs in fact have been very useful and where future growth will occur.

NNW Applications in General

NNW's, despite all appearances to the contrary, are appearing in ever increasing numbers of real world applications and are making real money:
OCR (Optical Character Recognition)
· Caere Inc ($3M profit on $55M revenue in 1997) "OmniPage Pro 6.0 significantly
increases accuracy with its exclusive Quadratic Neural Network(TM) (QNN)
technology, an enhancement to its industry-leading OCR engine..."
Data Mining
· HNC ($23M profit on $110M revenue in 1997). Their flagship product is Falcon.
"Falcon is a neural network-based system that examines transaction, cardholder, and
merchant data to detect a wide range of credit card fraud...".

These days a purchase of a new scanner typically includes a commercial OCR program. The algorithms used are proprietary but most OCR programs are believed to use NNWs. (Calera, started in 1986, did not admit to using NNW in its OCR programs until 1992 when Caere began advertising the use of them in its OCR products). Designers of OCR programs may choose NNWs to accomplish one or more of these steps with NNWs while using for other steps other techniques such as conventional AI (If-Then rules), statistical models, hidden Markov models, etc. The point is that NNWs are becoming commonly used tools but, just like other math techniques such as FFT and least squares fit, they are still only tools, not the whole solution. Few real problems of interest can be totally solved by a single NNW.