The Center for Data Innovation spoke with Jason Cohen, chief executive officer of Analytical Flavor Systems. Cohen discussed how demographics can influence people’s perception of taste as well as how to evaluate the trustworthiness of food reviewers.
Joshua New: I’ve had the opportunity to interview sensory scientists that use data science to better understand people’s’ taste preferences. Is that what Gastrograph does, or is there more to it?
Jason Cohen: Gastrograph AI is an artificial intelligence platform for modeling and predicting human sensory perception and consumer preference—so our goals overlap with traditional sensory science, but differ in some key ways. Our thesis is that most products are only generally acceptable, but not optimized for a specific target consumer cohort, which means that most food and beverage companies are developing products that lots of people like, but no one loves. We help those companies develop better products by targeting specific segments of the population with a better flavor profile that meets their flavor, aroma, and texture preferences.
The other thing that differentiates our models from what other sensory scientists are doing is its predictive power—we can use tiny amounts of data, often times less than 10 reviews of a product, to predict how the entire world will taste, and the distribution of preferences for each population. We can do this because our models are all hierarchical in nature—they build on our past data and past predictions to become increasingly accurate with less data over time.
New: Before Analytical Flavor Systems, you launched and ran the Tea Institute at Penn State. Can you describe the work you did there? How did this lead you to your current role?
Cohen: The Tea Institute at Penn State is an interdisciplinary research institute dedicated to the study and preservation of tea, tea ceremony, and tea culture. In addition to overseeing the research of about 30 students in over five fields of study, I did my own research originally in sensory science before moving to machine learning and artificial intelligence. My goal at the institute was to develop models of human sensory perception that could be used to predict what different tasters would perceive in a variety of products. That original research created the technology base for Analytical Flavor Systems, and we keep improving on it today.
New: Gastrograph factors in demographic data such as age and socioeconomic status to predict consumer preferences. How much do demographics actually influence people’s tastes?
Cohen: Demographic factors are the classic nature versus nurture debate. Both have a large impact in what you perceive and what you prefer! We take into account the major demographic factors like age and socioeconomic status to help companies develop better products for everyone. What we find is that the mix of genetics and past experience cause cohorts to emerge amongst similar individuals, which make their perception and preferences predictable.
In 1979, Pierre Bourdieu wrote Distinction, a book about how individuals acquire their preferences; most tellingly, he wrote: “the arts of the bourgeois will remain the arts of the bourgeois because they have the formative experiences in which to appreciate it.” It’s an interesting quote on a variety of fronts! For example, if an individual has never tasted a lychee or pine nut, then they couldn’t use those descriptive terms when describing a product. That could be a problem, because it could look like wealthier individuals are better tasters, when in reality they only have a larger tasting vocabulary because of the wider range of their past experiences.
We knew when we were developing our models and system of sensory data collection that it couldn’t only work for those individuals, and we spent a lot of time developing a system that can collect and infer high resolution data even from individuals with very small vocabularies. Going back to lychee and pine nut as an example of that: even if an individual had never tasted a lychee, they could still describe it as fruity and tart and sweet, or other lower-resolution flavors, and they could still describe the pine nut as nutty and metallic.
New: It seems Analytical Flavor Systems’ major focus thus far has been beer. Is there a reason for this? How easy would it be to adapt the models you’ve developed for other kinds of foods and drinks that might not have such an “enthusiast” culture?
Cohen: My research and all of our original work was in tea, but it was very difficult to train American university student as tea tasters, because of their lack of familiarity with the product. I started collecting data on coffee, and while the data quality was better, it was still difficult to attract enough tasters. Then I started holding beer tasting, and no longer had a data problem—people would show up for free beer!
For Analytical Flavor Systems, beer is now a small portion of the products we work on. We also work on: chocolate, tea, noodles, dairy, soft drinks, baked goods, and spirits! It’s very easy for us to enter a new product category, it just takes a few days of data collection and some model training time.
New: You published some interesting research about how to determine the trustworthiness of food or beverage reviewers. How does this work? How do you go about quantifying something that seems so heavily subjective?
Cohen: Everyone likes to believe that they are unique and their opinions and responses are their own, but we can show that sensory perception is a response to an underlying stimuli. For example, we can both agree on what color an item is, and we can both agree on what flavor a product is. When both of us taste a product, the chemical composition doesn’t change, but our responses might—that’s the perceptual filter. What our AI learns is the distribution of the perception of specific demographic tasting populations and consumer cohorts within those populations, so that it can map the responses of stimuli between these groups.
To measure the accuracy and trustworthiness of an individual, we look at their in-group agreement based on how other tasters within the same demographic change their perception over time. That allows us to calibrate the experience curves across all users to see how an individual’s ability to identify subtlety and nuance in a product is predictive of their skill level.