Caltech Creates Electronic Nose

eldavojohn writes "Researchers have created an electronic nose that can detect odor and identify which odors are a concern to it. From the article, 'The Lewis Group a division of Chemistry and Chemical Engineering at Caltech have a working model of an electronic nose. The efforts of Caltech scientists has led to an array of simple, readily fabricated chemically sensitive conducted polymer film. An array of broadly-cross reactive sensors respond to a variety of odors. However, the pattern of differential responses across the array produces a unique pattern for each odorant. The electronic nose can identify, classify and quantify when necessary the vapor or odor that poses a concern or threat.'"

Re:Artificial Nose

[cashdot]

It is off-topic, but I could not resist: I don't think that cameras and microphones have surpassed the human capabilities. Show me a microphone that has the same dynamic range as the human ear. Or a vision system that has the same 'postprocessing' capabilities as our visual cortex. Resolution and sensitivity are not the only performance indicators!

Re:Why fix what isn't broken?

[teslar]

But is the chemical the smell, and is the wavelength the colour?

Well, yes, since these represent the necessary and sufficient stimuli for you to perceive the smell as the smell of roses or the colour as the colour red. I don't think there's a need to go philosophical on this point.

What you are really talking about, I think, is the experience of perceiving a smell or odour. Then it's very clear that everything depends on who is doing the smelling/looking and nobody is going to argue that electronicc noses experience these stimuli in the same way we do. So yes, an electronic nose would have a priori problems with qualifying smells in subjective ways (smells good, bad, refreshing, stale) unless you specifically train it with lots of examples from all those categories. But that's not really the point of an electronic nose, it's more about detecting toxins and perhaps reverse-engineering certain odours (e.g. just what did the chef put into that lovely sauce of his?).

I also wonder how it works on things where the 'known' composition can vary. Will it mis-identify them like some other robot did identifying a reporter's hand as bacon (or something similar)?

It depends on what you want to do. If you want to identify a complex odour based on the mix of chemicals you've encountered, then yes, that can happen.
When you design such a system, you take lots of samples from all the inputs that vary so that you get a good idea of the possible variation for every given input (so you train the system with 50 roses, tulips and pieces of bacon instead of just one each). If you were to plot the inputs in a multidimensional space (one dimension per chemical you can detect, and the metric is the concentration of said chemical) you would thus not get e.g. a single point for the odour 'rose', another single point for the odour 'tulip' and a third point for the odour 'bacon'; you would get entire clouds of points.
If you're lucky, there will be plenty of empty space between the clouds and then you can easily train a classifier to discriminate between the odours (sometimes as simply as computing the distance between the odour you detect to the centre of each cloud and going for the closest cloud). If there is some overlap, you can still train a classifier and will be alright most of the time, just sometimes you'll have to qualify your ouput with a probability if your input falls into the overlap region (80% chance it's a rose, 20% it's a tulip). If there is heavy overlap, you're pretty much screwed unless you can think of a funky nonlinear transformation of the input which achieves a better separtion of the the different classes.

So yeah, misclassifications can happen, it depends on your input space and how the classifier deals with it. Especially novel stimuli can be a problem, I'd imagine.