Commentary: Wonders - The Silicon Gourmet; April 1997; Scientific American Magazine; by Morrison; 2 Page(s)

Strange but enlightening news is at hand to help explain our human sense of smell. Worldwide, the industry of flavors and fragrances is beginning to employ workable, if still naive, PC-based simulators capable of the subtle discriminations made by the human nose. There are many sensor types that show reproducible electrical responses to small amounts of odorant vapors. (One early instance uses semiconductors that run quite hot. They vary in resistance when sample molecules oxidize on the conducting surface.) Yet these are still no match for the discerning power of the host of olfactory cells hidden in the nose.

The aggregate reports from a small collection of such assorted sensors, each with its own response to a great many odors, can be extracted by a number of mathematical weighting procedures, made verifiable mainly by their agreement. The most general means is through "neural network" computation. This wide class of unusual computer programs has a long history, dating back 30 years to a kind of computer called the perceptron. Such programs march to no firm clockbeat, carry out no precise logical chains, tolerate substantial error and indecision, and learn best from patterns shown to them as examples. (To be sure, they are usually realized on a PC program, designed to simulate the intertwined multiple feedback loops of the neural-network process.)