Computers learning to predict smell

Status
Not open for further replies.

parker25mv

Well-known member
Oct 12, 2016
Just thought this might be interesting to some of you.

Some researchers have used an advanced program provided by Google to develop a learning algorithm to be able to predict what a molecule will be likely to smell like based on its molecule structure.

Right now Google has some very advanced artificial intelligence algorithms that can recognize objects in pictures and recognize words in voices, so now some people thought it would be a good idea to turn that technology towards trying to recognize smell.

https://neurohive.io/en/news/learni...ve-trained-a-neural-network-to-predict-smell/
 

Big L

Well-known member
Nov 23, 2019
Yes, this technology is definitely already here. I built one myself a couple of years ago, and the results were already very impressive. I toyed with the idea of starting a business around it but ended up dropping it.
 

giftmischer

Well-known member
Oct 3, 2016
Further reading:
Kowalewski, Joel, and Anandasankar Ray. "Predicting human olfactory perception from activities of odorant receptors." Iscience 23.8 (2020): 101361.
https://www.sciencedirect.com/science/article/pii/S2589004220305484

and finally they succeeded:
Kowalewski, Joel, Brandon Huynh, and Anandasankar Ray. "A System-Wide Understanding of the Human Olfactory Percept Chemical Space." Chemical senses 46 (2021).
https://academic.oup.com/chemse/article-abstract/doi/10.1093/chemse/bjab007/6153471
 

Big L

Well-known member
Nov 23, 2019
This is pretty different (and much more interesting from a scientific standpoint) than what I was working on. Mine was actually very similar to the google one. The deep network was learning a database of molecules described as graphs and annotated with odor descriptions. Basically, the software learned to identify similarities in the structure of molecules with similar odor descriptors. It did the whole thing independently without being fed any previous human-based knowledge (other than odor description). The goal was to let it generate novel molecules on its own, given a detailed odor description. I guess you can imagine what would that be useful for.
 

mnitabach

Well-known member
Nov 13, 2020
This is analogous to similar efforts in medicinal chemistry led by researchers such as Bryan Roth & Brian Shoichet. In no way does any of this provide license or justification for foisting off WAGs based on eyeballing chemical structures onto unsuspecting students of perfumery as if they have any real basis in wisdom or experience with the molecules in question.
 

contrebande

Well-known member
Dec 6, 2019
I received a Basenotes DM from someone enquiring about this so I replied to this person and also in another data-related thread I would write something here. But I didn't have the time until now.

Altough there has been undeniable progress on solving the olfactory QSAR problem and hope is high that it's going to eventually exhibit some above-Turing Test performances someday soon, we're not there yet and the accuracy/precision/recall figures being thrown around by researchers are absolutely bogus and actually nowhere near human performance. Here's why.

I read the Google' seminal 2019 GNN Learning to Smell paper back when it came out, but I only got into it after watching Luca Turin's Secret of smell video lectures and talking to him about the computational approach to constructing an organoleptic olfactive space. But I come from the NLP world way back when it was still called that, and that experience is what's allowing me to see problems where other, greener ML/AI practicionners think everything is a-okay.

First of all, it's not really QSAR. Real QSAR is when you base your regression model on actual receptor activity. And it is not the case here. Not only is there no consensus on how olfactory chemo-reception works, but the studies cited above do not really base their experiments around receptor activity at all: they look for a mapping between SMILES molecular graphs (the "structure" part of QSAR) and "organoleptic tags", "smell words", "percepts" or whatever you want to call them. You can contruct very high quality molecular structure embeddings and fingerprints with GNNs and other methods and they are used elsewhere with great results, so the usage of SMILES representations of molecules is fine. The problem lies with the way receptor activity is approximated with natural language. And I'm not saying we should not use natural language: working with natural language is more difficult than it seems but that's all we have until we know how olfactory receptors and the brain work together to produce the sense of smell. So on the "activity" side of things, there is still a lot of work to do and the techniques currently used are insufficient when not plain wrong.

For instance, there is something off just with the given figures of 500,000 molecules and ~150 tags (percepts) in the dataset from the latest paper mentionned above. TGSC lists less than 10K molecules that are used in the perfumery industry. To get to half a million, you have to include molecules not for perfumery use from other sources, such as PubChem and such. And even if we assume that all of those molecules are volatile organic compounds that have had their odor properly evaluated and tagged with "smell words" by competent humans, for more than 90% of them there is only one tag: "odorless". So right there, you can get 90% accuracy with a stupid system that "predicts" that every SMILES you feed it doesn't smell anything. So while it's correct to assign the "odorless" tag to a VOC molecule and your system must be able to predict odorlessness, you cannot train a machine properly with such an imbalanced dataset. The thing to do is to keep in the training set only a small sample of the "odorless" molecules so as the "odorless" tag is not disproportionately represented, easy peasy. But nobody seems to be doing this correctly.

A rule of the thumb for me (and many data scientists) is that your training set should be constructed around a relatively equal number of molecules representing every percept (within 5% or within the target margin of error). And that number should be relatively high, around 1,000 molecules associations for every tag (around 1,000 with the tag "rose", for example). I have never seen a training set coming nowhere near that, but I think it is possible/feasible even given only the data we have now and a little bit of scraping and NLP magic.

The second most common tag in databases with that many molecules (usually not used in perfumery) is "characteristic". And again, assuming that the smell evaluators were not just too lazy to assign other existing tags, it means that this molecule produces an odor that has no known natural language referent (other than its own name). So, it's not a "percept" : it doesn't represent a smell at all and predicting a "characteristic" odor doesn't mean anything at all. It is a bad tag and it should be excluded from any training dataset. Yet it is there in plain sight, in the input data of most of the recent experiments and molecule odor prediction systems I've seen. I personally believe that those molecules with only the "characteristic" tag should be re-evaluated and the odor vocabulary should be augmented so as to eliminate the "characteristic" tag completely. And that is not as easy a solution as for the "odorless" tag: olfactometry at scale is a very costly process. And nobody is doing it rigorously enough as far as I know (i.e: publicly). Or again, different data sources could be combined and NLP techniques used to find those "orphan" molecules some smell words.

Without getting into as much details, here are a few other tricky problems with tag-based systems. You will find in those databases a lot of tags that are pure inventions from perfumery marketing departments, intended as market segmentation semantics and for which the olfactory value is very debatable or inconsistent: "fougère", "oriental", "amber", etc. to name a few. Then there is a distinction that is not always clear between tags that are meant to convey olfactory "taxonomy" (the "kind" of scent: "rose", "jasmine", "civet", "musk", "hay", etc.) compared those meant to express olfactory "intensity", strength, impact, detection threshold, etc. (such as "pungent", "powerful", "intense", "transparent", etc.) and longevity (for "top", "heart" and "base" notes: "volatile", "long-lasting", etc.). Provisions for the difference between olfactory activity and trigeminal activity (eg: "hot", "cold" and sometimes "spicy") are also rarely made (and probably should). Finally, there are lots of tags that are applied to a molecule based on where it occurs in nature and do not describe the odor it produces: it's not because a molecule is found in rose oils and absolutes that it should have the "rose" tag.

You could say that ML in general and NNs in particular can get away with noise. It's true, but I feel that researchers in this field could be a lot more dilligent in their data science. If you compare to machine vision, machine olfaction is misssing the mark on data handling best practices by a significant margin and it's not hard to see. Most of the problems illustrated above come from the fact that there is simply not enough data in the training sets for the algorithms to see the signal through the noise. From experience, simply going from a finite and limited tag set to a richer, multi-source, multi-lingual input "corpus" with NLP/ETL dilligently applied (stemming, POS, NER, TF-IDF, ontology/taxonomy extraction from text, better word embeddings, etc.) can give multi-point performance increases (>10%). So, scraping lots of plain text data from a lot of different contexts (like this forum, like Arctander and other books, etc.) along with the more structured databases (TGSC, Leffingwell, Pubchem, EODb, etc.) somehow associating lots of single ingredients to lots of text (not just 4-5 tags) is the way to go at least for me.

Lastly, here are two problems that arise in the production of ordor prediction models that are not related to tags or smell words but are worth mentionning anyways. Often there will be mention of an olfactory "space" being constructed through the process. Asssuming that you have been dilligent with the data and that you indeed have created a set upon which there exists a "norm" or a "distance" between two element molecules representing their olfactory "similarity", what you have is pretty useless and most definitely does not constitute an olfactory space. For your model to qualify as a "useful" olfactory space, you would need to have two more things: 1) a commutative addition operator where any two elements A and B could be added together (like in a perfume formula) and the result C of this operation is also an element of the same space representing the odor of the mixture of the two element A and B and; 2) a scalar product where the scalar X represent the dilution or the relative quantity of member molecule A (to represent the dilution or relative quantity of an ingredient in a perfume formula). Those two things make your odor space useful to predict not only the smell of a single ingredient, but also the smell of a mixture of ingredients and thus be useful in composition. Aiming for just the distance/norm/similarity between two molecules is not enough. Another problem is the way isomerism is (not) represented in the input/training data. Some isomers differ in the taxonomy dimensions so much so that they are considered different ingredients altogether like geraniol and nerol, while others will vary greatly in the intensity/scale dimensions, from "odorless" to "intense", like the constituent isomers of Iso E Super and those of Hedione. A lot of isomers have not been properly isolated from each other and the data to distinguish them properly doesn't even exists and since isomerism greatly affects our olfatory sense, then having better isomer data will most definitely improve the construction of an olfactory space. Finally, the experiments described in recent papers are not reproducible and their results are therefore not repeatable, usually because of proprietary data (e.g. captives) or proprietary software. And IMHO, inovation without reproducibility and repeatability has little value, especially with computational olfaction being such a moving target.

That's it. This could go on forever, but I feel that it captures the essence of where I think state of the art machine olfaction is compared to where it seems to be only by looking at the recent literature.
 

mnitabach

Well-known member
Nov 13, 2020
Great post above!!! Thank you for taking the time to explain in tremendous detail multiple reasons why this enterprise is currently completely useless for predicting scents of new molecules, let alone how scents combine, for purposes of perfumery. I have been a neuroscientist since the mid 1980s & paid close attention to the artificial intelligence field since then. In those days neural networks w backpropagation-based learning rules was going to "revolutionize AI". Of course this didn't happen. Every ten years or so, after the last fancy AI buzzword has proved to fail, a new one arises that is going to "revolutionize AI", and then fails too. Machine Learning is the latest, and it will fail.

Genuinely self-driving cars based on ML have been promised any day now for a number of years, yet are not coming any time foreseeably. Your dissection of the odor-molecule mapping problem is identical. The fundamental problem is the same as it's always been: garbage in-garbage out. In the case of self-driving cars, we don't understand in any genuine depth the perceptual cues human beings use to judge the "edge cases" in very confusing & difficult situation to avoid crashes. Self-driving cars work great in lanes on highways; not so much on a roundabout in New Jersey or Belgium. You have done an amply wonderful job explaining the garbage in-garbage out problem for the odor-scent mapping problem. In the case of human vision, at least we have a decent understanding of important "primitive" perceptual features that are relevant: edges. In the case of olfaction, we have ZERO idea what the odorant-receptor mapping is for nearly all of the thousands of odor molecules used in perfumery. And, as you point out, we don't even have a proper ontology or description of the odor quality attributable to each of those molecules, let alone any combinations.
 

Bill Roberts

Well-known member
Mar 1, 2013
One can"t expect to run SMILES and observed organoleptics through any program and produce reliable and accurate predictions, particularly where not generating conformation prediction from the SMILES, which is merely an alphanumeric coding of structure, that is to say connectivity, which is friendly to machines but not to human reading or writing and adds no information beyond a drawn structure or correct chemical name.

Since accurate and reliable prediction cannot be presently done even on having good modeling of exact conformation a molecule will take, even more one cannot do so having only connectivity and trying to make associations from that. What care the receptors of connectivity alone?

If you do perfuming, smell the materials if you can. If you can't, consider observations of those who have rather than structure based guessing.
 
Last edited:

Big L

Well-known member
Nov 23, 2019
Well, Mike, I see this subject is quite emotionally charged, so I will try my best to present some issues without drugging this thread into an argumentative religious-like debate.

Machine Learning (ML) is not a contemporary buzzword. This term was coined back in 1959 and is one of the main fields of Artificial Intelligence (AI). It is perhaps more widely used nowadays because it is more accurate (i.e. the machines are learning) and maybe less confusing or misleading than AI.

Backpropagation, actually an algorithm in the field of Machine Learning, did revolutionize the field. Indeed it didn't happen back in the late '80s as some perhaps promised, but it made a comeback at around 2010 and is one of the main forces behind the entire field.

Machine learning is extensively used all around us in many diverse fields and ways. If you ever used google translate, you have used it. As well as every time you talk to Siri, Alexa, or Google Assistant. It is used heavily in image and video processing and generation. It is also, unfortunately, driving missiles fired in conflict zones around the world. The technology for self-driving cars is mostly already here. The issues delaying it are more psychological and legislative.

ML, AI, and Backpropagating Neural Networks are used in the chemical industries to explore chemical spaces, including drug discovery in the pharma industry. Because of the inherent secrecy in the flavor and fragrance industry, I can not say for sure it is used there, but I will be very surprised if it is not.

This is, of course, not to say that computers are developing new drugs or flavor molecules, but they help reduce the number of experiments involved by orders of magnitude. They do it using mathematical models and ML modules that suggest possible molecules of interest.

I agree it might be a bit silly (or at least very ambitious) at the moment to try to apply ML to problems of odorant-receptor mapping. Yet this is not because of shortcomings in the field of ML, but rather since our neurobiological understanding of olfactory is still in its infancy.

Just like the amazing Günther Ohloff used his human intelligence to describe the relationship between molecular structure and odor, so can a neural network undergo a similar process and produce similar results. As mentioned earlier, I build one with my own hands and saw it working with my own eyes.

This is one of the beauties of engineering, the entire biological process doesn't have to be known or well described for us to produce new molecules satisfying specific requirements. This is how the F&F and companies do with their 1 billion dollars a year research budget.

Do you believe the chemists guess new molecules? Or do they use their knowledge of existing "Lily of the Vally like" molecules to develop new molecules smelling like Lilly of the Valley?

Wouldn't that be worthwhile if we can use ML to reduce 10 fold, or even just cut in half the time and money required to develop such new odorants?

As mentioned before, I already had some proof of concept and investors willing to bet their well-earned dollars on it (which arguably means nothing, I know). Still, my (relatively limited) experience in the pharma industry made me reluctant to pick behind the veil of the F&F industry. My decision not to go this path was driven less by mistrust in the technology and much more by my unwillingness to take two things I love – fragrance and machine learning – and ruin them and my health, by fighting bloody business battles in one of the less penetrable industries on the planet.
 

Big L

Well-known member
Nov 23, 2019
Some time ago it was brought out plainly and clearly here that one could not expect to run SMILES and observed organoleptics through any program and produce good predictions. This was not close to plausible.
SMILES in themselves are indeed not very useful for such endeavors. They are ok to use as input but should be seen more as an identifier than anything else. No algorithm is expected to learn directly from them. Some method, mathematical or otherwise (crystallographic data, etc.), is used to predict/determine a 3D structure, and the learning can start from there. Some learning machines use so-called fingerprinting algorithms that generate a sequence based on the presence and structure of active groups in the molecule. Others use graphs to describe the molecular structure directly.
 

Bill Roberts

Well-known member
Mar 1, 2013
For those doing perfuming, as mentioned above, I can see absolutely no reason to place value in a machine prediction over smell vs human observation of smell. Would we agree we can throw that one right out as being a productive use?

For the 0.001% (something like this) that are actually in the business of synthesizing candidate molecules for new AC's, does one really think they are in need for advice outside their industry based on efforts such as proposed?

IMO it is rather like, everything could appear a nail to the man skilled in hardware who is very pleased with the hammer he has just made, but when going to say SpaceX and saying his great hammer will be just the thing to solve whatever problem with their electronics and thinking they will surely buy his hammer. However they will have to tell him, "This is a sophisticated piece of equipment, not a nail." They already have the tools they need and are not in need of his hammer, graphs, learning model that combed through TGSC descriptors or any such alternate effort for productive development. Sad perhaps but true.

Having fun is great. I hope you enjoy your efforts to guess organoleptics from structure. But it's not going to be of use to perfumers, it just isn't, and I do not think it an expectation based on facts that there is a market of those who produce AC's who would actually need and pay for such. Have any chemists at Firmenich, Givaudan, etc expressed such a need, wishing for someone somewhere to provide a machine learning program off of TGSC data and the like? I am guessing not. If they have then I suppose I am wrong.

You have a great website which is actually useful, very much so, to DIY perfumery so by no means is this intended as overall discouragement or criticism. It's just that running published organoleptics of a large molecule set vs SMILES is simply not rationally expected to give good predictions. It is just hopium, wanting to trust in machine learning, while missing the barriers that are extremely clear on being viewed either from the receptor interaction standpoint or the perceptual standpoint.

Not wishing to beat a dead horse into the ground, but as an example how would one hope from machine learning on other compounds to predict the potency and character differences of say the patchouli ethanone isomers, or the Hedione isomers, other than just blind trust in machine learning? What is the evidence of predictable pattern that could have been found for those? None. Because it's precise details of fit of those precise molecules, fitting differently with different isomers, combined with perceptual factors that are mostly still a mystery to date.

Smell the stuff rather than predict the stuff. We've seen on this forum the result of human pattern recognition and lot of effort by a smart and knowledgeable individual and the success rate to date remains 0.00%. It is a mistaken assumption that everything can be solved in such a manner. For some reason sometimes it has addictive appeal though. I can point only to the 0.00% success rate and the reasons for it.
 

mnitabach

Well-known member
Nov 13, 2020
I'm pretty sure that regardless of the utility of "machine learning" or whatever term one uses for guiding design of new aromachemicals, it is of zero relevance to end user perfumers except to the extent that it leads to useful new molecules. And it certainly cannot be considered license or justification for WAGs about likely odor qualities of a molecule that one hadn't smelled on the basis of eyeballing molecular structures that aren't even accurate 3D representations.
 

Big L

Well-known member
Nov 23, 2019
For the 0.001% (something like this) that are actually in the business of synthesizing candidate molecules for new AC's, does one really think they are in need for advice outside their industry based on efforts such as proposed?
To be honest, we (as in I) are not in the business of synthesizing candidate molecules and not even in the business of perfuming. I am a hobbyist, and as such, I do what I enjoy.

Givaudan, on the other hand, Givaudan has an AI department that is publicly known to develop tools used for formulating perfumes. It is also publicly known that most of their research budget is invested in synthesizing candidate molecules. Should we assume that they even need to hear about my space hammer and not actually already have all the hammers, mallets, and gavels we can imagine?
 

Big L

Well-known member
Nov 23, 2019
Bill, you edited as I was posting and very much in line with what I was writing. My projects are my fun projects. I am assuming the real market players have much better databases and much better AI/ML tools. The main point is that this technology is real and most likely being used, even if we as hobbyists have no access to it.
 

Big L

Well-known member
Nov 23, 2019
I'm pretty sure that regardless of the utility of "machine learning" or whatever term one uses for guiding design of new aromachemicals, it is of zero relevance to end user perfumers except to the extent that it leads to useful new molecules. And it certainly cannot be considered license or justification for WAGs about likely odor qualities of a molecule that one hadn't smelled on the basis of eyeballing molecular structures that aren't even accurate 3D representations.

I totally agree and don't believe this is the goal of anyone developing such tools.
 

mnitabach

Well-known member
Nov 13, 2020
In the cases of in silico screening of enormous libraries of candidate GPCR ligands to narrow down candidates for real-world testing, there exist very high resolution atomic structures of known ligands bound to the GPCR. This is exists for, at most, an exceedingly few odorants & odor receptors.
 

Big L

Well-known member
Nov 23, 2019
You have a great website which is actually useful, very much so, to DIY perfumery so by no means is this intended as overall discouragement or criticism.
To clear any confusion that might have been there unguentarius.com is using no fancy AI, merely high-school level of mathematics. In any case, I thought we are having a nice online discussion and wasn't in any way feeling personally attacked.

t's just that running published organoleptics of a large molecule set vs SMILES is simply not rationally expected to give good predictions. It is just hopium, wanting to trust in machine learning, while missing the barriers that are extremely clear on being viewed either from the receptor interaction standpoint or the perceptual standpoint.

Not wishing to beat a dead horse into the ground, but just an example how would you hope from machine learning on other compounds to predict the potency and character differences of say the patchouli ethanone isomers, or the Hedione isomers, other than just blind trust "Computer smart!" What is the evidence of predictable pattern that could have been found for those? None. Because it's precise details of fit of those precise molecules, fitting differently with different isomers, combined with perceptual factors that are mostly still a mystery to date.

Smell the stuff rather than predict the stuff. We've seen on this forum the result of human pattern recognition and lot of effort by a smart and knowledgeable individual and the success rate is 0.00%. It is a mistaken assumption that everything can be solved in such a manner. For some reason sometimes it has addictive appeal though. I can point only to the 0.00% success rate and the reasons for it.

I would not expect any computer (or human) to know the potency and character differences of patchouli ethanone isomers, not without meticulously synthesizing and evaluating them first. After that has been done, I would hope it is possible to invent a new neroli ethanone :)wink:) and after a few of these efforts to be able with some level of accuracy (at least better than the former 0) to tell which isomeres are more likely to be the desirable ones.

I believe the frustration in this thread arises from the false assumption someone wants to "predict the stuff rather than smell the stuff". What do we imagine? Some companies generating products that were never evaluated and presenting them to the public?

The goal of all such in silico methods is not to replace the physical evaluation by the noses of field experts. Rather it is to reduce the time and effort taken by these experts by using worse-than-human computers to do as much as possible of the preliminary "dirty work".
 

Bill Roberts

Well-known member
Mar 1, 2013
It's not a false assumption -- on this forum anyway!

We get more than a few cases where one could wait for someone who actually has smelled something, but instead we get guesses of smell based on structure. Wrong every time so far.

Further, if software were put out available to DIY'ers that made such guesses, or output from such software, then in DIY perfumery one would have exactly that choice, so it would not be a false assumption that this choice would exist and it would be exactly the issue at hand. We'd just have another source of useless guesses, hardly an improvement on the current situation! My advice would be, go by smell, preferably one's own but failing that seeing what others say, and pay no mind to software guesses should anyone offer them to you. None.

So yes it's an entirely necessary point when talking about software predictions of smell and how that would relate to DIY perfumery, which after all is our topic.

Moving then to those actually in the business of making new AC's, rather a different topic and with not many here really having background closely enough related to give some grasp of rational development programs to achieve goals related to olfactory receptor activation and perception, again, I highly doubt they need outside software to help them, so it seems a moot point what such outside software might do for them.

If the discussion is supposed instead to be not of outside efforts based off of TGSC's database etc but the inhouse efforts of exactly such companies, then there is really nothing to discuss, or I can't see what, as all that is quite proprietary and provides nothing to us other than final product. We can discuss that they are doing it -- not much to say there, about one sentence I guess -- and that at best it can produce sets of plausible candidates with as yet quite low likelihood of any given one smelling really as predicted or aimed for, though that can be acceptable when producing large numbers of candidates and having a big budget, but that is about it!
 

mnitabach

Well-known member
Nov 13, 2020
It is not a false assumption, sadly, that WAGs of scent character on the basis of eyeballing molecular structures are promulgated on this forum.
 

mnitabach

Well-known member
Nov 13, 2020
Seems like Bill gave up on his promise to never reply to a thread I participated in. He was put on time out by the moderators because of his toxicity replying to the thread I started about SMILES and aldehydes linked above where Bill's first reply was to admit he didn't know anything about SMILES. I do not know what he wrote here. I blocked him long ago. But let me guess: he says that he was right all along and that what I wrote is an admission of defeat on machine olfaction? If so, allow me to be clear: if you think that's what my post means, then don't get into machine olfaction. You don't have what it takes and you will hurt yourself. Stay away from any discussion about it as well if you can: there are plenty of people making this work and they are very enthusiastic about what they do. You will only look like a fool should one of them try to reason with you. I wish you kept your promise not to interact with me and any thread I participated in, Bill. So this advice goes to you too. Also, if you put words in one's mouth, one might just defend oneself.

As for machine olfaction use cases (finding new molecules vs. finding new perfumes compositions) I must say that finding new molecules or rather guiding molecule synthesis in an industrial context is pretty much a given (but can still be improved). I know Guillaume Godin, the guy leading this group at Firmenich and he's into finding new perfume formulas too, now. And he is critical about the tag-based approach in pretty much the same way I am. Computer aided perfumery on the other hand, what you could say I'm more into, is also called "perfume engineering" and it's a big thing, bigger yet in the academia than it is in the industry, because QSAR in the industry is so well established whereas perfume engineering endeavours are more like pet projects on the rise (but are already used by most professional perfumers in the big five). Finally, I'm glad Mike mentionned Gerkin's Pyrfume effort because he and I are greatly aligned too, especially about ontological engineering and NLP. I talked to him and his students during the pandemic about finding a "universal ontology" for machine olfaction. We're not there yet, and we probably won't announce it here when we are, but it will happen one day soon and it's going to make a difference in the field for sure.

Machine olfaction today is helping make greater perfume faster than was ever possible before and if you don't like the idea of buying, smelling or wearing perfume that comes from the mind of a machine, then stick to making your own. And if you are into DIY perfume engineering and you have the basic skills (organic chemistry, NLP, data science, ML, maths, etc.) to make it work at home then let's talk (I prefer Instagram DM) and you will see how fun and profound a process it is. Chance are even you will find it infinitely more rewarding than coming to this forum for perfume making guidance.

Contrebande, I would definitely be interested in learning more about what you discuss here. I am not on Instagram, but if you are willing to discuss via email, I am at nitabach@yale.edu.
 

Bill Roberts

Well-known member
Mar 1, 2013
It's unfortunate you quoted that, Mike. The content in reference to me is devoid of merit, as a minimalistic summary. I quite understate the matter in using that phrasing.

I had replied to you not to that individual and there is absolutely zero reason for that false content. It is not just out of left field, it is off the planet.

I understand the individual has been instructed to remove it.

If the falsehoods and general personal attack content regarding me remain on the forum, rebuttal will need to remain as well. If they are removed, I am more than glad to in turn remove all reference to that content which has zero place on this forum or in any polite society.

Edit: The above appears continuation of combativeness. I never requested or suggested any deletion. However at this point due to CoC if mods see need to delete this entire matter created by above, fine with me, this is utter ****.
 
Last edited:

giftmischer

Well-known member
Oct 3, 2016
Maybe thats all beyond DIY perfumery, but now this thread is here and I could not withstand post the links.

Joel Kowalewski and Anandasankar Ray dont use any chemical features to run their model.
Their 2020 article is open access, so I thought someone will read that.

Let me cite just two sentences:
"We found that human odorant responses from heterologous assays could be used with comparable and sometimes better predictive success. In part, the result is anticipated by the fact that each OR is presumably selective to very specific physicochemical features themselves."

Although ... very interesting (fascinating?) to DIY perfumery, the wheels of Fig.5 seem to be...
 

xii

Well-known member
Jun 9, 2015
I’d love to have some computational tools while formulating. Like compressed sensing applied to the tower perfume-accord-ingredient, provided we can combine it with semantic mapping or some such. I’ve been doing the thing “manually” for a while already but it’s pain and lots of choices made are arbitrary.
 

contrebande

Well-known member
Dec 6, 2019
Hi Xii, well you know a little bit what I'm up to because I've told you in private in greater details than I ever did to anyone else on this forum. Especially that my primary interest is whether or not perfume compositions really provide a literary opportunity to mean something. The perfumery games you guys play here, the adversarial/competitive aspect of it interests me. Given the chance, they could be a very nice playground to try and see if machines could become or already are better perfumers than humans. But as I've already told you, I've never seen a game topic or goal that worth it me. I'm not interested in "a walk on the beach by myself" or "a walk in the forest by myself" or "what I feel inside when I pass by my favorite flower". The opposite of that is what my friends and I call Parfumerie dangereuse. I for one not only think that machines can help make better, more relevant perfumes, but rather that this new level of relevance is impossible for humans. The most important reason for this is that machines are not afraid of the ideas conveyed by dangerous, relevant perfumery, whereas most if not all humans are. Examples? These:


this

this

this


And then this, this and this. And I can scare you even more that that in private if you want, these are the safest things I could find. Well that is where perfumery still holds any value as an artform, as a media for critical philosophy to me and where I think it's worth investing any machine olfaction effort: language, sex, predation, speciation, intelligence, taboos, etc. And I agree that this forum is an inappropriate platform for this, as it is inappropriate for real perfumery discussions beyond a certain point. I would only feel comfortable and free talking about this in a private club of some kind, with diligently screened individuals. In the meantime, the next best thing is to do it by myself. Sure I can talk about the techniques, but I'll get bored very fast if I don't get some relevant insight out of it in return. And I know we can do it, Xii, we've been there before.
 
Last edited:

contrebande

Well-known member
Dec 6, 2019
Well that's unfortunate. I was interested.

No you were not. I was trying to get the Wayback Machine to archive the problematic second part of the post you deleted, but the archiver cooldown prevented me from doing that. Thank God I kept a full copy of your edited post, including the added second paragraph :

Would you mind actually distilling what you're getting at in something that's digestible in under three hours? I read the parfumerie dangereuse idea, but it sorta reads as hollow philosophically leaning emptiness -- but I'm sure that's because I didn't watch the videos. I could see something such as Eau d' Hermes as a parfumerie dangeruse especially in the late 1940's when it was created I mean the bikini had just barely been invented. If modern Bandit is anything like the vintage being one of these perfumes. Am I going in the right direction? Not necessarily animalic, or stinky, but rather in a time of much higher conservative wear these perfumes were dangerous expressions of human carnality. If I've got the concept down at all, then what would you suggest as perfumes that entail parfumerie dangeruse created since say 2000? I think if I could get an idea of actual perfumes crafted from this mindset it would help.


You may have that Canadian flag next to your name, but you've sure got the French way of speaking. Bit silly to come here, and talk down to everyone. As to machines crafting unique ideas has AI ever been shown to do that? Unique valuable ideas not novel. Will the machines be the resolve from the question of nihilism, or will we fall prey to the technique? If humans are evolutionary inclined, and we craft machines in our image, will a machine crafted perfume be able to do anything other than attempt to appeal to the broadest category of the gender? Isn't art meant to be inefficient? If it is, then how could a machine that's formed around the idea of efficiency ever craft art? In fact, it seems rather than machines becoming more like humans that humans are becoming more like machines. What exists anymore that isn't meant to be as efficient as possible as a means to the end? Perhaps a machine could only craft one philosophical perfume -- the stench of long rotten flesh, putrid death and decay. Perhaps that's what a machine would see as salvation. After all, if I've gotten the idea of parfumerie dangeruse at all down, then all good perfume is by it's nature anti-efficient.

My troll radar is state of the art. And like I said, it's pretty much a DM-only thing at this point. For those interested. And minimally competent. And not racists.

And if you're asking something of me, like explaining things, it's up to me to decide if your interest and competence are genuine and worth the effort.
 

GoldWineMemories

Well-known member
Nov 22, 2019
No you were not. I was trying to get the Wayback Machine to archive the problematic second part of the post you deleted, but the archiver cooldown prevented me from doing that. Thank God I kept a full copy of your edited post, including the added second paragraph :



My troll radar is state of the art. And like I said, it's pretty much a DM-only thing at this point. For those interested. And minimally competent. And not racists.

And if you're asking something of me, like explaining things, it's up to me to decide if your interest and competence are genuine and worth the effort.


If you say so man. I think you're upset about something thing other than fragrance, and I don't want any part of that, so I'll leave it at that, and I hope you have a nice night.
 

xii

Well-known member
Jun 9, 2015
Picked a quote:
I'm not interested in "a walk on the beach by myself" or "a walk in the forest by myself" or "what I feel inside when I pass by my favorite flower". The opposite of that is what my friends and I call Parfumerie dangereuse. I for one not only think that machines can help make better, more relevant perfumes, but rather that this new level of relevance is impossible for humans.
but will be responding to the entire post.
I'm not afraid of AI. Firstly, whatever we consider ambitious AI becomes mundane and obvious some time later and we stop calling it AI. Then, we could profit from adding sophistication and freedom of expression in perfumes. Which is much easier when using more direct senses, like hearing or vision.
What influenced me profoundly, all thanks to contrabande, was how human drives are intertwined with human intelligence. A conscious machine needs a gender and some sort of sexuality, I thought, and blended a perfume with this in mind. I don't think the perfume itself conveyed anything of that though.
 

contrebande

Well-known member
Dec 6, 2019
Hi Xii and to anyone seeing this through the crap on this thread, there would be a lot to say, of course. But not here. And God knows we tried our best with the normies (or I did, anyways). It's just not going to happen and it's not worth it. AI/ML in general but also the philosophy of Parfumerie dangereuse both require a safe space away from the people that are systematically offended by those things and have it censored away. Again, thank God for the Wayback Machine. Regardless, the Basenotes DIY forum obviously cannot provide that space. As you may know, we will be re-launching Contrebande both as an ingredient retailer and a lab for Parfumerie dangereuse, perfumery data, DIY machine olfaction, DIY perfume engineering, DIY VOC sensing hardware, DIY home fragrance diffusion hardware, and much more. Much in the way of the IAO but for tech/science/engineering/maker-oriented people, where the content and the practical ateliers are free because they backed by the perfume ingredient retail business. Part of that relaunch will include a private, invite-only Element/Matrix server (a kind of open source, distributed, uncensorable Discord/Gitter of sorts). The only place for sure I know it will be announced when it happens is our Instagram account and we can use the Intagram DMs in the meantime until the Element/Matrix server is up. I know creating an Instagram account can be a pain, but certainly no more than having one here. And I think it's worth it, given the situation. That's where it's at for now, anyways.
 
Status
Not open for further replies.

Latest News

Top