Turing Tests and the Problem of Artificial Olfaction
Unlike sight and sound, smells have never been accurately reproduced. Now one researcher has begun to think seriously about how artificial olfaction might work.
Here’s an interesting problem. When it comes to human senses, we’ve found ways to reproduce the look and sound of the real world reasonably accurately. There are even technologies for reproducing the feel of certain experiences, such as flight and car simulators.
But the problem of reproducing smell is much more intractable. The 1960 SmelloVision experiment is a case in point. This involved some 30 odors that were released into the cinema at certain times during a movie. Only one film—Scent of Mystery—ever used the system, which rapidly failed.
The truth is that nobody has found a way to accurately reproduce odors from the real world. And consequently, artificial olfaction is a technology that sits stubbornly beyond our reach.
The problem is essentially to measure an odor at one point in space and then reproduce it at another. And it is a task of surprising complexity and subtlety.
But even if it were possible, how would we test such a system? How would we know that the artificial odor was an accurate reproduction of the original?
That may sound like a trivial problem, but today we get an insight into its surprising complexity thanks to the work of David Harel at the Weizmann Institute of Science in Israel. Harel has developed a kind of Turing test for artificial olfaction that helps to explore the issues this problem raises.
To begin with, Harel explains why the problem of olfactory reproduction is fundamentally different to that of reproducing visual or sound stimuli. Reconstructing a visual stimulus is simply a matter of reproducing the spatial distribution of its wavelength and luminance. And for sound, pitch, loudness, and timbre define a tone.
Harel gives the example of the first photograph to include people, taken by Louis Daguerre in 1838, and the first telephone call, made by Alexander Graham Bell, in which he successfully summoned his assistant from the next room. “In both cases, the generated artefacts were immediately recognized as being true renditions of the originals. Not perfect, of course, but unmistakably recognizable,” says Harel.
For that reason, it is reasonable to think of photography and telephony as methods that produce faithful reproductions.
Smell presents a different challenge, however. Odors are made up of molecules that our olfactory system detects. It then transmits appropriate signals to the brain that result in the perception of a smell. “Hence analyzing and synthesizing smell is not just a question of using an appropriate set of mathematical functions to emit outputs involving accurately computed wavelengths,” says Harel.
Instead, Harel outlines the more complex process that is required to reproduce smells. It consists of three parts. The first he calls a “sniffer”—a device that transforms an input odor into a digital signature. The second is the “whiffer”—a device containing a range of fixed odors that can be mixed and released in carefully measured quantities and concentrations.
The third part of this system is perhaps the most important. This is the interface between the sniffer and whiffer. “[This] analyzes the signature coming from the sniffer and instructs the whiffer as to how it should mix its pallet odorants to produce an output odor that is perceived by a human to be as close as possible to the original input,” explains Harel.
Having determined how the device would work in principle, Harel turns to the important question of how well it can perform. He asks whether the goal should be to exactly imitate the smell or just to reproduce it well enough to allow humans to recognize it.
There’s an important distinction here. When we see a photograph or listen to the radio, we recognize the image or sound while knowing that they are not the real thing. “In both cases one becomes immediately convinced that the artificially produced output is an adequate reproduction of the original, albeit artificial,” says Harel.
Not only that, it is straightforward to be convinced that this process of reproduction is entirely general and will work just as well regardless of the input. So it’s easy to think that the process of photography should reproduce a cityscape as well as it reproduces a landscape or a portrait.
But the sense of smell is different. “Such an approach would be totally inappropriate for olfaction,” says Harel.
Here, the technology comes up against some important limitations. Perhaps the most significant of these is the inability of human language to describe smells. “No methods exist for verbally describing the essence of arbitrary odors,” says Harel.
Some methods attempt to borrow words from other senses to describe smells as cool or green, for example. Others have devised words relating only to smell, such as musky, putrid, and floral. But none of these are able to span the entire spectrum of human-discernable odors.
That makes odor identification a tricky business. A human might be able to recognize the scent of coffee or an orange but would surely founder when asked to recognize smells associated with more general scenes, such as moss in a dark cave, the smell of screeching tires, or the odor of some unknown animal in a faraway forest.
So Harel has an alternative approach, loosely based on the Turing test for artificial intelligence. In this test, a human has to distinguish an artificial intelligence from a human intelligence. Harel’s idea is to ask a human to distinguish smells produced by the artificial olfactory machine from real ones.
The method is straightforward and cleverly designed to avoid any verbal characterizations. Harel says audio and video can give the tester a sense of immersion. So the method would involve a tester watching a video of the place where the smell had been gathered and then deciding whether the associated smell is real or artificial. (Harel also suggests a couple of variations of this method.)
Repeating this process with many different samples and testers would soon give a sense of how well the artificial olfactory system performs.
Of course, care must be taken not to make the task too onerous. For example, starting with 10 bottles of wine and asking testers to determine whether the smells are real or artificial would be too hard for most people. But in principle, such a testing system could work.
That’s an interesting thought experiment that focuses on the nature of experience and how we make sense of it. It’s not hard to imagine more immersive challenges that involve more than one sense—sight sound, smell, touch, and so on—and ask whether it is possible to distinguish reality from some kind of artificial experience. That’s a topic many science fiction films have explored.
In that case, we may get an idea of how important olfaction is in contributing to realistic experiences. For food-related experience, it is clearly important. But the terrible truth may be that for the rest of the time, smell may be a tiny or insignificant component of experience. Perhaps that’s why we lack the vocabulary to describe it accurately.
In any case, for the moment at least, artificial olfaction remains beyond reach. It’s now 50 years since the SmelloVision experiment. Perhaps Harel’s ideas will help trigger some innovative thinking about how to achieve true artificial olfaction in the near future.
Ref: arxiv.org/abs/1603.08666 : Niépce-Bell or Turing: How to Test Odor Reproduction?