AI Beats Human Experts At Distinguishing American Whiskey From Scotch (newscientist.com) 72
An AI system has outperformed human experts in distinguishing between American whiskey and Scotch, achieving 100% accuracy by identifying subtle differences in the chemical composition of the spirits. New Scientist reports: Andreas Grasskamp at the Fraunhofer Institute for Process Engineering and Packaging IVV in Germany and his colleagues trained an AI molecular odor prediction algorithm called OWSum on descriptions of different whiskies. Then, in a study involving 16 samples -- nine types of Scotch whisky and seven types of American bourbon or whiskey -- they tasked OWSum with telling drinks from the two nations apart based on keyword descriptions of their flavors, such as flowery, fruity, woody or smoky. Using these alone, the AI could tell which country a drink came from with almost 94 per cent accuracy.
Because the complex aroma of these spirits is determined by the absence or presence of many chemical compounds, the researchers also fed the AI a reference dataset of 390 molecules commonly found in whiskies. When they gave the AI data from gas chromatography -- mass spectrometry showing which molecules were present in the sample spirits, it boosted OWSum's ability to differentiate American from Scotch drams to 100 percent. Compounds such as menthol and citronellol were a dead giveaway for American whiskey, while the presence of methyl decanoate and heptanoic acid pointed to Scotch.
The researchers also tested both OWSum and a neural network on their ability to predict the top five odor keywords based on the chemical contents of a whisky. On a score from 1 for perfect accuracy to 0 for consistent inaccuracy, OWSum achieved 0.72. The neural network achieved 0.78 and human whisky expert test participants achieved only 0.57. The study has been published in the journal Nature Communications Chemistry.
Because the complex aroma of these spirits is determined by the absence or presence of many chemical compounds, the researchers also fed the AI a reference dataset of 390 molecules commonly found in whiskies. When they gave the AI data from gas chromatography -- mass spectrometry showing which molecules were present in the sample spirits, it boosted OWSum's ability to differentiate American from Scotch drams to 100 percent. Compounds such as menthol and citronellol were a dead giveaway for American whiskey, while the presence of methyl decanoate and heptanoic acid pointed to Scotch.
The researchers also tested both OWSum and a neural network on their ability to predict the top five odor keywords based on the chemical contents of a whisky. On a score from 1 for perfect accuracy to 0 for consistent inaccuracy, OWSum achieved 0.72. The neural network achieved 0.78 and human whisky expert test participants achieved only 0.57. The study has been published in the journal Nature Communications Chemistry.
Translated... (Score:5, Insightful)
Re: (Score:1)
>Chemical sensor beats general-purpose organs, something something AI
Now we wait for monies.
Re: Translated... (Score:2)
Yes. Fraunhofer is sure to derive billions in royalties at $1 per "tasting".
Re: If it was so simple (Score:2)
Re: (Score:2)
It was.
With a spectrograph. Which is, in case you didn't work it out... a necessary part of this process.
This is using a system to compare spectral data, a very easy computational task. If it requires an 'AI' then you're doing it wrong.
The part that requires AI is the marketing.
So... (Score:3)
What's the value proposition here? Is there American-made whiskey being labelled and sold as Scotch?
Re: (Score:1)
The value proposition is to get publicity. Oh, you mean value for you? None. At least if you are like me and like to try it yourself to decide whether you like it or not. You know, like normal people do.
Re: (Score:2)
"product performance analysis", which is one of the business applications of these guys. they've been doing that for ages and apparently just discovered a slightly better method using a neural network.
then again i would assume that labelling and marketing of such products can involve quite some shenanigans indeed. if end consumers would actually appreciate any difference (considering that even human "experts" are not much better than a random guess (0.57)) or even really would want to know is another matter
Re: (Score:1)
100 - 57 -> 43.
I think the difference between "guessing" right and being wrong: is 14 points.
So 0.57 is *much* better than a random guess.
Re:So... (Score:5, Informative)
The value proposition is killing off the "sommelier profession", which is as much a "profession" as "professional astrology".
An article from the age before everything was called "AI": https://www.theguardian.com/li... [theguardian.com]
Also, automated chemical analysis isn't "AI".
Re:So... (Score:4, Insightful)
In our days, everything that utilizes a neural network is considered AI, even if it is only specialized for a single task.
Re: (Score:3)
In our days, everything that utilizes a neural network is considered AI, even if it is only specialized for a single task.
Actually, in IT consulting industry these days, anything that resembles "computer making decisions independently " is sold as AI. Neural networks, statistical analyses, machine learning models and even decisions trees - anything goes.
Re: So... (Score:2)
Nobody in user land knows the actual difference, and now you meet user expectations of being on top of the latest shiny object.
Win-win!
Re: (Score:3)
The value proposition is killing off the "sommelier profession", which is as much a "profession" as "professional astrology".
as usal the clickbait headline is misleading: their "professionals" were already made obsolete by their prior algorithm (OWSum), and this new method is only marginally better than that. admittedly the accuracy still leaves a bit to be desired, but relative to both methods the "professionals" barely even score. there surely was bad energy in the test room.
Also, automated chemical analysis isn't "AI".
convolutional neural networks are indeed fundamental part of the ai toolset.
Re: So... (Score:2)
Sommelier is a real job; it's there to help select the best products for the people who appreciate the finest things in life.
I guess it's modern parlance, you could say they're similar to influencers reviewing products. Or evangelists using slightly older buzzwords.
Re: (Score:2)
The value proposition is killing off the "sommelier profession", which is as much a "profession" as "professional astrology".
Actually it's not. Sommeliers aren't just snobs who say "American" or "Scotch". They understand how flavour profiles pair with food in a way to elevate both. That isn't "professional astrology", that is "professional understanding of what humans like".
Re: (Score:2)
Indeed, it's not.
Unlike the "professional" astrology, which has a foot in physical reality - the natal charts do begin with a sky chart - the "sommelier" is more of a bad novelist, all make-belief and cliches.
Re: (Score:2)
The value proposition is killing off the "sommelier profession", which is as much a "profession" as "professional astrology".
Nepo-babies need to make money somehow. Why would you take this from them? It only affects the other nepo-babies and their families. The existence (or not) of a sommelier does not affect my life in any way.
Re: (Score:2)
There is hardly any other useful angle to this story.
Wake me up when it can rank them (Score:2)
Re: Wake me up when it can rank them (Score:2)
Re: (Score:1)
I grew out of drinking alcohol a long time ago, but back then the same alcohol drink used to taste differently on different days of the week, times of the day, on a cold or warm day, before or after eating something, etc.
So, the first insurmountable task here is to define "tastes better", because there is really no such thing.
Re: Wake me up when it can rank them (Score:2)
Possibly. If you start noticing CAPTCHAs with shots of liquors coming out of your monitor, and you're asked to rate them to prove you're not a bot, then they're already working on it. Or you've been drinking too much, which is the more likely explanation.
Wake me up when it can get liver disease (Score:2)
And another meaningless stunt (Score:2)
Seriously, this is getting as stupid as all the claims of crapto being useful (for anything except crime, that is).
Re: (Score:2)
Maybe he just understands how it functions, mathematically speaking, he understands that when crypto miners say 'work' what they mean is 'we threw away a lot of hashes', and that any blockchain based system generates 'value' through waste...?
Maybe the idea that a whole bunch of supposedly economically savvy people actually aren't and they're just posturing... is an argument you make yourself in other contexts. Think about it a little.
Re: And another meaningless stunt (Score:1)
Re: (Score:2)
Type (Score:2)
Were they Islay or Speyside scotches? You can generally just smell an Islay and tell it's an Islay. A mild Speyside can be indistinguishable from a dryer bourbon.
Re:Type (Score:5, Insightful)
I certainly hope most people don't need an AI to help them figure out whether a dram is Laphroaig or Lagavulin versus Jim Beam!
Re: Type (Score:2)
The J&B expert has spoken!
Read the Label? (Score:1)
According to the gas chromatograph (Score:2)
Re: (Score:2)
"'Mom', 'Love', and 'Screen Door' are all trademarks of MomCorp."
I didn't use 15MW of energy to learn that (Score:2)
And at least I enjoyed tasting either of them.
Re: (Score:2)
What's a MW of energy?
I for one welcome our (Score:1)
...Bender overlord.
Re: (Score:2)
Nah, it's Bender.
Hokay (Score:2)
Re: (Score:3)
Re: (Score:2)
The test was to tell them apart from the descriptions alone. Which isn't a test of the AI so much as it is a test of whether or not the person writing the descriptions is consistent. For all we know it was keying on British vs American spellings, like "whisky" vs. "whiskey" and "colour" vs. "color".
Slightly mor
Re: Hokay (Score:2)
Stupid (Score:5, Funny)
The whole thing is stupid. First off, the test is deeply unfair! The AI gets mass spec data which shows a chemical breakdown of what the ingredients are, while humans get whisk... ... Hmm .. ok scratch that .. how do I sign up for this study? Anyway, my point is, if you have mass spec results you don't even need AI you only need "If then else".
Re: (Score:3)
(sip) Hmm... I think I got it, but better safe than sorry - give me another of each!
(sip) No, there's obviously some subtle characteristic I'm missing. One more sip of each to verify!
(sip) These really are closer to each other than I expected. Just bring me the bottles so I don't hafta keep botherin' you!
(a couple hours pass) Where'sh zat mash shpectrometereretr?? And we needz 'nothr bottlez!!
Re: (Score:2)
"You spent HOW MUCH on imported booze!?"
"It... It was for research. I swear!"
"RESEARCH!? What research? You're a fucking computer scientist!"
"My AI research..."
"Don't you fucking DARE tell me you're going to get an AI drunk."
"No, babe. No, of course not. I'm researching... um... whether or not... um... the AI can tell Scotch from American whiskey?"
"Right... And just how is the AI supposed to taste the whiskey?"
"It doesn't have to taste it! It can tell from... from... from reading the description of
Re: (Score:2)
you don't even need AI you only need "If then else"
As much as the marketers would like us to believe it, I'm not convinced any current LLM is much more than an extremely complicated if/ then /else tree. I was waiting for the part in this article where the "AI" was only fed data from a smell sensor and actually did an analysis, but we didn't get there.
It's not clear from the article if the LLM is provided with data that associates a smell or flavor with a compound. Given that info, I would expect a junior sys admin could easily put together a script to p
More stupid AI shit (Score:1)
There's no such thing as artificial intelligence.
Really just stop right there.
One day when so-called "AI" can distinguish between drinks, call me up to the bar.
Until then I'm more thinking about why I'm paying for a "shot" which is 25ml in the stingy UK, 30ml in the stingy AU, 40ml in the short-pouring US, or 46ml/1.5oz in the honest part of the world.
WHO CARES which 80% proof (abv) it is if you won't serve it you know-nothings?
I'm sorry, Dave. I'm afraid I can't do that (Score:2)
Forget turning on their human masters, let me know when the first robot gets a DUI
AI rules (Score:2)
It's just not that hard! (Score:2)
They stock them in different aisles!
Why "AI", though? (Score:2)
So they did mass spec to determine the quantities of 390 different organic compounds, and based on that, they could distinguish between American and Scottish whiskies. But surely they could have accomplished the same thing using regular multivariate analysis-- you know, the kind I was taught 30 years ago in a first-year statistics course? Especially if, as the article states, there are certain compounds like menthol which reliably distinguish the two.
Worthless study (Score:2)
This question only answers "Is chemical analysis with gas chromatography more accurate that tasting?" We already knew this.
If the humans had the gas chromatography results and were trained in how to read them, AKA given exactly the same information the AI was, would there have been any difference in the results? I don't think so.
The other test was just word analysis on a very, very, very limited set of words, which shouldn't have taken an AI to find the connections, and probably could be done on a Commodore
ChatGPT, please aprar this csv file... (Score:2)
In other news.. (Score:2)
Al has the same advantage. Why is this news?
Practice more (Score:2)
Who the *(&^Y*^(*& wrote this headline? (Score:2)
How about this one: "Gas chromatagraphy better than the human nose at determining differences between two specific chemical compounds." I would be willing to bet a HUMAN given the output of the gas chromatagraph would be at 100% accuracy too.
"predict the top five odor keywords based on the chemical contents of a whisky."?!? What the holy hell is this nonsense? "Computers better than humans at moving a cursor precisely 100 pixels, robot dominance soon to follow."