Feedback appreciated

Over the past few years, I invested quite a bit of time in trying to develop artificial intelligence software that would understand principles of food & wine pairing. We've come a long way since and the result is now available on:
http://www.winewinewine.com/en-gb/winestein

Anyone interested in this subject is invited to try and provide feedback.

Thanks in advance for this!

Best, and good to be back here after some time,

Cor

Cor, Hello and Welcome to the Forum....

I do have a collection of minor questions but for the moment will do with two major issues:

(a) I am not particularly computer sophisticated but I do know a bit about artificial intelligence and fail to see any connection between your wine pairing suggestions and AI.

(b) Unless I am seriously mis-reading your DNA analyses, it does seem that a good many dry red wines (e.g. Alter Ego of Palmer) show a high level of botrytis. Perhaps something amiss with the program? And of course, as a sub-question, is this truly a DNA analysis or just the use of a rather cute catch-phrase?

Best
Rogov

Cor, Hello and Welcome to the Forum....

Thanks! I can't say why exactly, but I used to be a regular here some years ago and somehow disappeared. Guess I've kept myself too busy with other wine stuff

(a) I am not particularly computer sophisticated but I do know a bit about artificial intelligence and fail to see any connection between your wine pairing suggestions and AI.

What I've been trying to do is dissect both wines and dishes (their ingredients, preparations, sauce and other addings) in taste, textural, aromatic and a few other components. This is what was named 'DNA', which is indeed a bit of a catch-phrase. Subsequently, all these data on wines and dishes were combined with scores on how well or how bad certain wines would match with a set of dishes used to train the software. This was done working from basic gastronomic principles, together with a chef who is quite knowledgeable on wine. I think this leaves enough free space to have people pick a wine that also matches their personal preferences: as you will know, quite a few dishes have a rather broad span of matching wines. However, if you'd try and play with e.g. oysters, you will find that WineStein offers a much more narrow span of matching wines.

In total, we used about 50,000 data on food and wine combinations, good and bad. From there, the system is now able to make preedictions regarding pairing wines for basically any dish that comes somehow close to a dish it is familiar with (and meanwhile, that's a lot).

(b) Unless I am seriously mis-reading your DNA analyses, it does seem that a good many dry red wines (e.g. Alter Ego of Palmer) show a high level of botrytis. Perhaps something amiss with the program?

This should of course not be the case, and as far as I know never has been. Let me check if something is wrong in the representation of structural elements of wines (we just released an update of the site).

If you'd have any more questions, please drop them and I'll try and answer them.

Best,

Cor

Cor, Hi...

Thanks for the response. And, of course, good luck with developing the program and site.

Best
Rogov

[David M. Bueker]

Welcome back Cor! Good to see you around.

Unfortunately I am getting errors trying to open your web site, but I will try another browser later on & give you some feedback.

Cor, what's behind the curtain? ...expert system, neural nets, something else?

Cor, what's behind the curtain? ...expert system, neural nets, something else?

Hey Jeff,

The whole system is neural based, so we don't use any type of rule sets like in expert systems. So basically, software is fed with both analytic properties of wines and dishes, and with a set of training data regarding food & wine pairings. From there, the system is now able to make predictions. Additional advantage is, that this approach makes the system self-learning. Practically, this means it will only improve with time. Furthermore it's fully 'dynamic': once you have properly trained the thing what to do, it will be able to make predictions on every imaginable new combination of ingredients you enter into the system.

It's kind of scary to realize that the whole things actually has no clue what - in terms of what we perceive as expertise - it's actually calculating The experts from Nijmegen Univeristy, with whom we developed this, tell me that the software makes a 'fit' based on the data it is fed with, a statistical optimization. Like when it sorts out that a lot of salad dishes containing vinegar (=high acidity) dressings show no correlation with sweet, fortied red wines, it will give any new dish containing these properties a low ranking for that wine type. But... there are LOADS of such correlations to consider.
In total, it performs these optimizations on every single property of a wine type, and on every single property of a dish of choice, also incorporating strong or weak correlations within the specific DNA-profiles of dishes and wines. This totals something dazzling like over 40 million permutations before it'll spit out what it thinks

In terms of an image, WineStein is maybe best described as a 'gastronomic brain'. With sheer number crunching, it's able to discover all sorts of explicit and implicit aspects that play a role in pairing a dish with a wine type. Interestingly it will incidentally come up with rather unexpected matches, combos that you somehow never considered. Yet, when you try to think of how it might have reached that conclusion, it always makes some sort of gastronomic sense. Basically, that is a result of a number cruncher being able to go through ALL the alternatives, and on and again, this means it will come up with something that never crossed your mind.

If you have any more questions, don't hesitate

Best,

Cor

Welcome back Cor! Good to see you around.

Unfortunately I am getting errors trying to open your web site, but I will try another browser later on & give you some feedback.

David,

Sorry to hear this and this should of course no occur. I dropped you a PM so we can sort out what caused the problem. Apologies for this.

Best,

Cor

[David M. Bueker]

I have re-tried the site with IE 8 and Firefox and it seems to work now (IE 7 does not work, but no big deal, IE 7 does not work with much).

I experimented a little bit with a roasted lamb dish on the site, and I was very surprised that however I changed the herb components/saucing I kept being steered back to Aussie Grenache. Normally, with the exception of something like a relatively plain, grilled piece of steak or lamb I make wine decisions based on saucing/spice/herb elements rather than the meat choice. If I am doing something wrong let me know.

(b) Unless I am seriously mis-reading your DNA analyses, it does seem that a good many dry red wines (e.g. Alter Ego of Palmer) show a high level of botrytis. Perhaps something amiss with the program? And

Daniel,

I checked, but didn't see any botrytis in the wine DNA of Alter Ego, nor in any other dry red. I take it you did you read the graph the correct way around? Could you perhaps cross check with e.g. a Sauternes to see if you get a similar representation on botrytis? And if you'd really be getting botrytis in a dry red then, please PM me what browser you're using?

Cor, Hi.....


As I say, I may be reading this incorrectly, but see the DNA profile at http://www.winewinewine.com/en-gb/findw ... eau-Palmer

Best
Rogov

Thanks, Cor. I appreciate some of the "mysteries" of neural nets. Is it secret or can you tell us how many interior nodes there are? Have you tried looking at the activation levels to see whether you could map any of them to a human sense stimulus? (In some simple object-recognition nets, for example, it is possible to look at the nodes and say, "ah, that one is highly activated whenever the animal has wings, so it's a sort of bird indicator".)

And, one more q: Nets need to be trained so whose tastes did you use when saying "this combo works, this combo doesn't" ?

Cor, Hi.....


As I say, I may be reading this incorrectly, but see the DNA profile at http://www.winewinewine.com/en-gb/findw ... eau-Palmer

Best
Rogov

Daniel,

I saw it....and see zero botrytis there. The darker color in the graph indicates the level.Guess that is not entirely clear, so will check this with the design boys.

Best,

Cor

Thanks, Cor. I appreciate some of the "mysteries" of neural nets. Is it secret or can you tell us how many interior nodes there are? Have you tried looking at the activation levels to see whether you could map any of them to a human sense stimulus? (In some simple object-recognition nets, for example, it is possible to look at the nodes and say, "ah, that one is highly activated whenever the animal has wings, so it's a sort of bird indicator".)

Jeff,
It looks like you're a bit more of an expert on the technicalities of this than I could claim to be. All I can say is I did the gastronomic input according to what the neural experts specified. I understand at a basic level what they do, but I don't really feel qualified to give you a proper answer here, sorry

And, one more q: Nets need to be trained so whose tastes did you use when saying "this combo works, this combo doesn't" ?

Good point, and I agree that - in principle - there's a subjective element here. We trained according to fairly general gastronomic principles regarding good and bad matches, leaving a usually fairly broad band of possible combo's so that people can add in an element of personal taste. We're currently working on letting personal preference profiles of registered users help to further narrow down what would be good 'personal' matches. The training was done by myself and a chef who is good with wine combo's, and we regulalry use feedback sessiosn with other experts in this field to fine tune. E.g. we recently found Holland's first three star Michelin chef (now retired) willing to spend a few thoughts on this as well.

Best,

Cor

I have re-tried the site with IE 8 and Firefox and it seems to work now (IE 7 does not work, but no big deal, IE 7 does not work with much).

I experimented a little bit with a roasted lamb dish on the site, and I was very surprised that however I changed the herb components/saucing I kept being steered back to Aussie Grenache. Normally, with the exception of something like a relatively plain, grilled piece of steak or lamb I make wine decisions based on saucing/spice/herb elements rather than the meat choice. If I am doing something wrong let me know.

David,
I did a few lamb roasts as well. With a mild curry cream sauce it suggests Medoc or Italian Bordeaux blend, with truffle gravy it suggests NW grenache, Rioja, NZ pinot noir and with garlic and a rosemary gravy it suggests Chateauneuf.
I didn't check all the exact scores, but since many meat dishes with not too specific sauces allow a very broad band of red wines to match well, there could well be a lot of scores that are actually very close.

Best,

Cor

[David M. Bueker]

I have to admit that I did not try curry.

Cor,
I know enough about neural nets to be dangerous. What model do you use? How many nodes and layers, etc.?

I also know that innovative combinations of flavors might spoof a trained network. I guess I can play around to see how much flexibility I am allowed.

Then there is the question about a wine's age. We'll see what this can do...

Thanks for taking this project on. It seems intriguing.

Well, already problems.

I want to try a recipe that I invented for a maple-chile glazed salmon that is baked. I cannot find where I define the additional flavors in the glaze.

Also, why can't I find the saute preparation?