What could a machine possibly learn about tea? And why would any AI want to learn about tea? We can barely find any humans to study tea!
My company uses machine learning and narrow-band artificial intelligence to understand what people taste in complex food and beverage products – the AI we’re building probably won’t take over the word any time soon, but if it does, it’ll just make everything taste better.
So why AI?
AI is a bad term for a wide range of research and capabilities; our narrow-band AI can be better thought of as automated machine intelligence to identify and control for subjectivity in human sensory data. Our goal is to strip away the influences of age, sex, race, socio-economic status, past tasting experience, first language, and smoking habits from what people claim to taste in a product, and arrive at the underlying chemical composition of flavor active compounds – without any lab equipment.
Once we do that, the real fun and intrigue begins. For example, we use that data not only to build real-time quality control monitoring for flaws, taints, and contaminations in beer – but also to determine what percentage of the population will be able to taste any quality control problem and the overlap with the producer’s target customer demographics. In bourbon, we’ve developed flavor-profile optimization strategies in the production process, and have helped producers predict the optimal barrel and bottle aging, removing a lot of the guess work from the process.
What projects do I have in mind for tea?
Nearly all of my research on tea to date has been though the Institute; but I was recently asked by Teamaster Teaparker during the Yixing Teapot Exhibition to put together a quick analysis of three Foshou Shan high mountain oolongs brewed comparatively in a Yixing and gaiwan.
It’s clear (and no surprise to anyone) that the 1982 (33 year old) tea was the taster’s choice – but what’s more interesting is the effect of brewing the tea in a real Yixing pot.
That analysis got me thinking about the types of valuable tools I could build for the tea industry, and what human tasters must struggle with while purchasing tea.
Age and Claim Verification of Pu’er Tea
Verifying the age of pu’er is a difficult undertaking. Cakes can be faked, wrappers can be forged, even the manufacturing house has an incentive to make the tea taste “older” than it is (hence all of the sheng cakes with “just a little bit” of shou in them from the 90’s). Through the Tea Institute, I have access to some very carefully sourced pu’er and the resources to have them radiocarbon dated.
The masterpiece era of pu’er ran from approximately 1950 – 1965, during which HongYin, LuYin, and HuangYin were produced; and about the same time as the atmospheric nuclear tests throughout the cold war, giving us the “bomb curve”. Usually radiocarbon dating is only accurate +/- 80 years (far too large a range for most teas!), but the atmospheric nuclear tests and the resulting release of nearly double the amount of carbon-14 into the atmosphere gives us seasonal and, in some cases, nearly monthly resolution in the age of the sample.
What does this have to do with sensory and machine learning?
Last time I went to China to purchase some pu’er, I found lugging my radiocarbon lab a bit inconvenient, and, being an inventive chap, I realized (with some re-inspiration from this last event) that I could model the flavor signatures from known old samples of pu’er tea to get a fairly accurate age range from sensory data alone!
I’ve been experimenting with multivariate adaptive regression splines on data preprocessed with a unsupervised metric learning implementation of local fisher discriminant analysis. There are a few other data tuning tricks I will need to use, but for now this is yielding viable results.
Does it work?
There’s not enough data yet to know. I’ve done some small scale experimentation on sensory reviews from highly experienced Tea Institute members, but that’s not a truly fair test.
My eventual goal is to obtain the same level of time-resolution (seasonality or month specificity) from the sensory reviews of less-experienced reviewers as is available from the radiocarbon dating. Then it could be used while making purchasing decisions in the field!
Cool… but where’s the AI?
Ah! I’m happy you asked. It’s important to understand that you can’t just take a few human sensory reviews, people recording what they taste, and start building statistical models (although that’s what everyone else in the field is doing… so it’s no surprise most products fail at launch) – you need to control for their personal attributes and experience. Our models first determine the age, sex, socio-economic status, and past tasting experience of every reviewer, and then classify them into a tasting population and preference archetype using displayed phenotypic classification by learning their sensitivities to gauge how they physiologically respond to flavor.
Not sure… I’d like to find some more time to build an age verification tool for the pu’er tea, and then maybe I’d work on a brew analysis tool. It would take your review of a brew of tea, and determine if it was over or under-brewed and if it was balanced or imbalanced – it might even be able to give brewing suggestions in the future.