“Technology is becoming faster, bigger or smaller depending on the situation, more efficient, lighter, heavier, whatever it needs to become. Humanity and innovation allow us to adapt and progress through whatever challenges arise.” Marc Briancon
We all recognize food as being more than simply the thing we eat. Food is everywhere and it goes beyond simply the cooking and preparation of ingredients. It’s in our culture and shapes history. We memorialize it to mark special occasions. (Remember that Thanksgiving dinner three years ago?). Food is studied for its art and its science. In the former, there’s gastronomy (the art of cooking and eating food), molecular gastronomy (the physicaland chemical transformations of ingredients during cooking), and oenology (the science of wine) in the first category. In the latter, two come to mind, food science (the physical, biological and chemical makeup of food), and food technology (the design, distribution, packaging, storage and use of food).
Chefs, sommeliers, food scientists, cooks, farmers, foodies, truckers, dietitians, nutritionists have the traditional trades supporting food. Software engineers are now joining the fray as consumers embrace a slew of food-related digital innovations. Consumers are finding recipes on their smartphones, scanning products, ordering groceries online, subscribing to delivery services, and sharing experiences through social media. To support those innovations, smarts are implemented, the most advanced of which is Machine Learning (ML).
What is machine learning? Machine learning is a set of software techniques (at times referred as algorithms) that automate the creation of models and the use of these models in every day life. These models learn from data and make predictions about data. This is why, at times, machine learning is referred to as big data. Machine learning used to be referred to as Artificial Intelligence (AI). There is not one machine learning technique, rather there are numerous techniques each better suited to specific applications. You might not realize it, but you are experiencing machine learning every day in your digital life. Netflix or Amazon suggests a movie or product recommendation? Machine learning. VISA calls you because of a suspicious activity? Machine learning. Google’s car drives by itself? You guessed it: Machine learning! The smarts behind Kitchology’s app that profiles consumers’ activities and matches food to activities? You already know the answer.
Let us explore different scenarios in how this concept might/should/could change the food world this coming year going from simple to more complicated.
Advertising selection: You have all experienced cookies (not the chocolate kind) and their impact on our internet. You check Amazon for sugar free water flavoring or for a lawnmower and for the next 30 days, you will see that lawnmower just about everywhere you surf. If you search for recipes on Yummly or all recipes, you’ll see an ad related to your surfing history appear while you’re glancing at recipes. Things become more complicated on smartphones because of the reduced screen space. The good news is that smartphones capture a lot of information (time, location, etc..) and machine learning can incorporate that knowledge when deciding which ad to show (chicken or tuna ad?).
Prepopulating shopping lists: Imagine you write salads a lot on your shopping list, including salads, carrots, and greens. In this example machine learning should prepopulate your shopping list with salad dressing, taking into account the size of your family (number of ounces of dressing per week per person) and of course your dietary restrictions (say, no sesame). You don’t want to add this item every time to your shopping list, only when it makes sense. The challenge is that it’s hard to determine the shopping activities of every single family.
Which product will sell? Machine learning also helps suppliers. Product decisions are essential for brands. According to statistics published by Kuczmarski & Associates, companies are experiencing 12 product failures for every 1 success. Those odds can have a great impact on their bottom line. You are not drinking New Coke, snacking on WOW! chips, eating Cosmopolitan yogurt or dining on Colgate chicken. I did not think so. Machine learning can assess which attributes of food are relevant to acceptance by consumers, from design to packaging to branding. Machine learning is used more and more to decide when to pull the plug after a product launch. The key to designing it is capturing the right knowledge about how consumers plan food purchases, buy food and, more and more, how they use food. This is especially true for folks who deal with food allergies, intolerances or special diets. These individuals insist on certain key attributes of food and no amount of marketing can convince them to buy the “wrong” thing.
Predicting ingredient-pairing/creating recipes/selecting recipes. This is a very rich area of research and development. You need first to organize food in buckets and map the relationships between these sets. This is typically done in a database (or if you are really into geekiness “Ontology Library”). Different companies base their design on different datasets. There is not really one right model. Designs differ mostly based on which choice of primary data is used for their model. The choices are: 1) recipes, 2) taste or 3) cooks’ insight.
IBM Chef Watson is the most famous example of the first type. IBM implements a set of algorithms that crunches through many datasets, tens of thousands of recipes, regional and cultural knowledge and statistical, molecular and food pairing theories. Based on that analysis, Chef Watson generates new recipes, often with unusual ingredients. IBM refers to machine learning applied to recipe analysis as cognitive cooking.
McCormick developed FlavorPrint, a prototypical use of the second type of machine learning applied to food. McCormick’s research about food tastes, textures, aromas, and preparation techniques was encoded into a set of algorithms that helped select recipes. Because they both rely mostly on foods and compositions, McCormick’s and IBM’s designs are categorized as food-internal stimuli based machine learning (a lot of geeky words, sorry folks!).
A third approach is to use a cook centric dataset and use the collective wisdom of chefs as the primary (not exclusive) dataset for machine learning. It is closely related to what people who make food designs use as a technique. This is the approach Kitchology adopted. The principle of food pairing is taught in culinary schools and they integrate with a proprietary database that helps users modify recipes. It has the advantage of allowing consumers to help in the design over time. We use this machine learning technique to modify recipes so that consumers can eat the dishes they want with simple changes.
I’m guessing you would like chicken tonight? It’s one thing to figure out if you like chicken, it is another to figure out if you would like it tonight. This is true because in addition to the food internal stimuli (taste and liking), one makes food decisions based on other factors. These factors include social environments (eating alone or with friends), location, ideational factors and anticipated consequences (I will starve tomorrow, so I can eat more tonight). In a perfect world, all these factors need to be considered. Given this level of complexity we are probably 5 to 10 years away from being able to make those decisions. Food decisions are more complicated than just dealing food internal stimuli. This is true, in part, because the data is not available for the learning. No data, no learning. No learning, no chicken. When the day the learning happen, we will have reached the Arthur C. Clarke’s definition of magic.
These are but a few examples of machine learning. As I hope you can tell, machine learning will provide ever more magical digital applications. We are only at the cusp of this revolution. So to my fellow machine learning friends, let’s get cooking.