Home » Technology » Scientific Computing » If It Smells Like a Rose ... Maybe You're a Neural Network

If It Smells Like a Rose ... Maybe You're a Neural Network

© Adam Hughes

In the last couple of articles, we've skimmed the surface of learning what artificial neural networks (ANN's) are and what types of problems they might be used to solved. And while our survey hasn't been particularly deep in nature, several key characteristics of neural networks have come to light. Specifically, neural networks are software or hardware constructs designed to roughly model the workings of the human brain. This is accomplished by employing a set of computer "neurons" that pass messages via a complex set of weighted connections. The importance (the weights) of individual connections can be dynamically altered to allow for a "learning" process. This adaptation is what leads to the attraction to neural networks for approaching "fuzzy" problems that don't necessarily have straightforward numerical solutions.

In order to gain a better understanding of just what can be done with artificial neural networks, particularly in the area of scientific computing, we'll spend the next few weeks looking at some example applications from researchers across the country. Because we're all human and have a pretty good feel for the thought process, and because neural networks are modeled to some degree on the human brain, sticking close to what we know is probably a good way to start out : thus, the Electronic Nose.

In 1993, researchers at the Pacific Northwest National Laboratory (PNNL) began working on the development of an artificial neural network for use in chemical vapor identification problems. The initial setup consisted of a bevy of chemical sensors integrated with an artificial neural network. Because of the interaction of vapor identification and neural networks, this type of system has been dubbed an "electronic nose".

Electronic noses are (roughly) composed of a sensing system and a pattern recognition device. Given what we've learned about neural networks, and because they can't "smell", it probably won't come as a great shock that ANN's serve as the pattern recognition engines in these devices. And it is the pattern recognition that holds great promise for increasing the power of these constructs.

Traditionally, an electronic nose is built of several sensors, each of which can recognize one chemical. It's obvious, then, that any mixture of chemicals must have now more species than the number of sensors available to detect them if we are to correctly identify everything present. When an ANN is introduced into the system, though, it can learn the characteristics of many chemicals. Then, sensor data can be propagated through the ANN, even if the sensor doesn't detect the specific chemical it is set up for. If the ANN has learned about the species in the past, the electronic

The copyright of the article If It Smells Like a Rose ... Maybe You're a Neural Network in Scientific Computing is owned by . Permission to republish If It Smells Like a Rose ... Maybe You're a Neural Network in print or online must be granted by the author in writing.