Hyper Dimensions in Olfactory Space

Artwork by Alex Grey

The title of this post is named after an article about categorizing smells, although it would work just as well as the title of a work of science fiction. It's from last August 2018, so it's old news by now; but that title isn't getting old anytime soon.

Probing the interconnected-ness of odors, and sketching a map of an omnicategorical odor network, the article starts out with a basic premise.

Let's say the olfactory system is designed to warn us of poisons in the environment. But a poison could be many chemicals, or a chemical we've never encountered before. So it would be necessary for the odors of those chemicals to be classified not by features intrinsic to the chemicals themselves, but by the likelihood of their co-occurrence with other chemicals. You can't be born with a database of chemicals to recognize and avoid. So instead, the hypothesis here is that an odor is only identified in its relationship to the other odors it's with.

This idea, at least to my ears, sounds really similar to the way statistical correlation text analysis can determine whether a piece of writing was written by a robot or not. (Also called visual forensics.)

It's way easier to visualize, so I'm taking these three images from the paper itself:

In the above 3 images, the first is a chunk of text written by a robot (most of the words are green, with a few yellows sprinkled in), the second is a real New York Times article (only half is green, the rest is yellow, with some red, and a sprinkle of purple) and the third picture is a clip from "the most unpredictable human text ever written", James Joyce's Finnegan's Wake (the colors green, yellow, red, purple are all evenly distributed about the page).

Green words are very predictably the next word. Yellow words are less likely to show up after the word they show up after. And red and purple are for when the next word is something you absolutely did not expect.

Because text-writing algorithms today use a statistical correlation program based on a compendium of written language (so they know what words typically occur together, and can therefore sound more like a normal person) the output of such algos will tend to look like the topmost image with all green words. Very predictable. The algos can't think for themselves, they can't "come up with" new stuff, and they can't be unpredictable. The whole point of writing an algorithm to do this is to prescribe what it's going to do in advance, i.e., it's predictable.

Bringing this back to olfaction, unfortunately there isn't a compendium of odor associations such as a Bible for smells or an encyclopedia for volatile organic compounds in nature. Furthermore, even if there were, we would need to augment it with a companion encyclopedia of the odors in the anthroposphere, because your supermarket isn't "nature" and yet it organizes a whole lot of our daily scentscape. One day though.

Notes:

Hyperbolic geometry of the olfactory space.

Yuansheng Zhou, Brian H. Smith, Tatyana O. Sharpee

Science Advances  29 Aug 2018: Vol. 4, no. 8, eaaq1458

DOI: 10.1126/sciadv.aaq1458

http://advances.sciencemag.org/content/4/8/eaaq1458

Catching a Unicorn with GLTR: A tool to detect automatically generated text.

Hendrik Strobelt and Sebastian Gehrmann. Association for Computational Linguistics: Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: System Demonstrations. Florence, Italy. July 2019. DOI:10.18653/v1/P19-3019

http://gltr.io

Cracking the Black Box

New technique elucidates the inner workings of neural networks trained on visual data

Jul 2017, phys.org

"Deep learning" and "neural networks" are terms that have become firmly planted in our popular lexicon. They all refer to the same thing, which is an artificial brain-like thing that teaches itself via feedback loops. I talked about this in Hidden Scents because the way our brain decodes olfactory information is a lot like the way these deep learning networks process their own big data.

This deep learning approach is way more effective than traditional computing for lots of problems like facial recognition or natural language translation. They're also really good at handling Big Data, you  know, like all that stuff cybercriminals keep stealing for ransom? Thing is, once these networks 'figure out' how to do whatever it is that they do, we have no idea how they did it.

Usually, with traditional programming, we write the code, so we know what it does and how. With this, the network essentially writes its own program, and since it seems to know what it's doing, we don't ask how. We just take the results.

Until now. This is one of the researchers, quoted in phys.org:

"We catalogued 1,100 visual concepts—things like the color green, or a swirly texture, or wood material, or a human face, or a bicycle wheel, or a snowy mountaintop," says David Bau, an MIT graduate student in electrical engineering and computer science and one of the paper's two first authors. "We drew on several data sets that other people had developed, and merged them into a broadly and densely labeled data set of visual concepts. It's got many, many labels, and for each label we know which pixels in which image correspond to that label."

Here is the link to their work.

image source

Non Traditional Computing, Complex Problems, and Approximation

Illustration for Death of a Salesman by Brian Stauffer for the Soulpepper Theater Company, Toronto

It may seem like a stretch to write about combinatorial optimization problems (aka the traveling salesman problem) on a blog about the ‘language of smell,’ but Limbic Signal isn’t just about smells, or language, but the connections between olfaction and computation. Our olfactory system is a champion at dealing with very large, very complex datasets.

Olfaction uses our brain in ways the other senses don’t. Some of the ways olfaction diverges from the other senses are akin to novel solutions to very complex problems in computation, such as big-data-sifting, pattern recognition, or the aforementioned traveling salesman problem.

 Also note that, in addition to the magnet network described below, another unconventional solution to the traveling salesman problem is to use mold. In fact, slime mold was used to design Spain's motorways and the Tokyo rail system.

So this article below does a good job of explaining the traveling salesman problem; I straight copied it from the writers at phys.org. And in the second section is an explanation of an interesting solution to the problem.

Researchers create a new type of computer that can solve problems that are a challenge for traditional computers

Oct 2016, phys.org

The traveling salesman problem

There is a special type of problem - called a combinatorial optimization problem - that traditional computers find difficult to solve, even approximately. An example is what's known as the "traveling salesman" problem, wherein a salesman has to visit a specific set of cities, each only once, and return to the first city, and the salesman wants to take the most efficient route possible. This problem may seem simple but the number of possible routes increases extremely rapidly as cities are added, and this underlies why the problem is difficult to solve.

...

It may be tempting to simply give up on the traveling salesman, but solving such hard optimization problems could have enormous impact in a wide range of areas. Examples include finding the optimal path for delivery trucks, minimizing interference in wireless networks, and determining how proteins fold. Even small improvements in some of these areas could result in massive monetary savings, which is why some scientists have spent their careers creating algorithms that produce very good approximate solutions to this type of problem.

An Ising machine

The Stanford team has built what's called an Ising machine, named for a mathematical model of magnetism. The machine acts like a reprogrammable network of artificial magnets where each magnet only points up or down and, like a real magnetic system, it is expected to tend toward operating at low energy.

The theory is that, if the connections among a network of magnets can be programmed to represent the problem at hand, once they settle on the optimal, low-energy directions they should face, the solution can be derived from their final state. In the case of the traveling salesman, each artificial magnet in the Ising machine represents the position of a city in a particular path.

Bacon

 AKA The Food Network

Bacon Flavor (2-Methoxy-4-methylphenol), surrounded by all the other flavors that use that same chemical-flavor; note that it's slightly modified for ease of viewing. 

There is only one chemical used to make the flavor of bacon.

To put it another way, the chemical that is used to make 'bacon flavor' is also used to make 'vanilla', 'jasmine', and 'clove', among others (including whiskey, which for the reason of lexical discrepancies does not show up on the graph).

Take a minute to play with this network graph - a compendium of flavors and fragrances and the relationships between them, using Aldrich's catalog of over 1000 chemicals.

(check out a youtube tutorial on how to use the network graphs)

Aldrich's Organoleptic Properties Index

POST SCRIPT

 Not only does this seaweed taste like bacon, it looks like fine-sliced proscuitto.

Culinologist Jason Ball from Oregon State University’s Food Innovation Center (via University of Copenhagen’s Nordic Food Lab) and aquaculture researcher Chris Langdon from OSU Marine Science Center, July 2015