Bored with finding cures for cancer and beating Jeopardy champions at their own game, the IBM computer system known as Watson has taken up a hobby: cooking.
For the past three years, the system’s keepers have fed it a steady diet of cookbooks and food theory. They’re trying to train a machine — which can’t even taste (!) — to understand what makes a good recipe.
Recently, Watson got so pro that, along with chefs from the Institute of Culinary Education in New York, it published a cookbook, an eccentric 231-page tome crowded with what its creators call “recipes for innovation.”
There’s a weird beef burrito accented with chocolate and edamame. A risotto studded with candied ginger, of all possible things. A pumpkin-ricotta cheesecake with savory mushroom meringues.
“Watson amplifies human creativity,” said Steve Abrams, an IBM engineer who worked on the Chef Watson team. “It’s a collaboration that allows Watson and the chef to discover more than either of them could independently.”
My personal adventures with Watson begin, as so many kitchen adventures do, with some overlooked, frost-bitten produce I needed to use. I’d planned to throw the frozen corn into a soup, but in the age of “cognitive cooking,” that’s for amateurs.
Home cooks, alas, don’t have access to quite the same version of Chef Watson that the Institute of Culinary Education did. But IBM, in partnership with Bon Appétit magazine, has released a slightly less robust Web app that basically uses the same technology.
You input your ingredients and preferences: a dessert with corn and sugar, I said.
And Watson generates pairing suggestions: pumpkin puree, medjool dates . . . horseradish.
From there, you can add and subtract ingredients, cuisines and dishes from Watson’s list until the system generates a satisfactory recipe template.
Watson is enormously complicated, so it’s hard to explain exactly what the site is doing when it makes its recommendations. But in the most basic terms, Watson ingests a huge amount of unstructured data — recipes, books, academic studies, tweets — and analyzes it for patterns the human eye wouldn’t detect. (If you’ve seen the recent blockbuster “Ex Machina,” you have a general, if sci-fi, model for this type of machine-learning.)
To create the Web app for home cooks, IBM researchers input nearly 10,000 recipes from Bon Appétit. To make the professional version of Chef Watson, researchers went even further: 30,000-plus recipes, scraped from the Internet; spreadsheets on the molecular makeup of different flavor and odor compounds in food; and academic research into the smells and tastes that people find most pleasurable, an obscure field known as “hedonic psychophysics.”
Much of that information is too technical, too literally microscopic, to register on most chefs’ radar. (Did you know, for instance, that white wine and tomatoes share the chemical compound hexenal?) As Watson crawls the recipe log, it calculates which foods appear in recipes together, and the statistical frequency of each match. It plots chemical affinities on a complex computational knowledge graph.
Corn, Watson decided, pairs pretty well with berries. Using the Watson app, I eventually ended up with a recipe for Corn Wedding Tarts, adapted from a more conventional version with rhubarb and phyllo pastry.
At the Institute of Culinary Education, chefs working with IBM probed Watson for similarly novel pairings. Asked for Spanish pastry ingredients, Watson suggested pepper, saffron, coconut milk, lemon extract and honey. Later, chef James Briscione turned to Watson for inspiration for a Creole-spiced dumpling. The system spit out okra, tomatoes, lamb and shrimp, among other things.
“If you look at the list of ingredients, it’s just gumbo,” Briscione said. “But I would never have thought to condense gumbo and put it in a dumpling like that. . . . It’s such a collaboration between you and the man in the machine.”
Briscione likens the Watson experience to solving a puzzle: You know the ingredients will work together, but not how or in what form. It’s like a more cerebral take on the popular Food Network show “Chopped” — which Briscione, incidentally, has won twice.
My Corn Wedding Tarts are . . . interesting. Not interesting in quite the same way as the chocolate-beef burrito from Watson’s cookbook, with its confoundingly cohesive garnishes of soybean and Edam cheese. Also not, thankfully, interesting in the same way as the mushroom-flavored whipped cream that topped an otherwise excellent cheesecake. (We fed that to our dog, who thought it was great.)
Instead, the corn tarts are just benignly peculiar: crooked rectangles of pie crust, topped with a sticky mound of corn and ginger in a strawberry-jam glaze. Strawberries and corn, it turns out, contain high concentrations of an organic compound called furaneol, which is why this combination, against all odds, approaches okay.
Still, I find myself wondering why Watson didn’t suggest a corn panna cotta with strawberry sauce, or maybe a sweet corn ice cream. In the introduction to their cookbook, the tech-heads at IBM expound at length on the concept of “cognitive computing” — the idea that systems like Watson have become so advanced, they can actually rival human creativity.
Alas, all the engineering advances in the world haven’t taught Watson that mixing corn and jam creates a sticky, mottled goop — or that said goop doesn’t make for appropriate wedding food.
“There will always be a collaborative relationship between the cook and the information,” Briscione said. “You could have the best computer in the world, but without someone who knows what they’re doing in the kitchen, it’s not going to work.”
That said, Watson is getting better all the time, both as a chef and as a machine. Watson incorporates user feedback into its algorithms, which means it’s always improving.
The cookbook is not the end of the road for Chef Watson: IBM’s Abrams says the company is researching applications for use by the food and beverage industries. (The Post reached out to several chefs for comment, but none cared to speculate on their future under the machines.) Future home versions of Watson could also know your health and dietary needs, creating recipes that are vegan or low-sugar or gluten-free. The Institute of Culinary Education, still partnered with IBM, uses the program to instruct its trainees.
All of this is a sideshow for IBM, of course; the company created Chef Watson only as “a metaphor,” Abrams says, for the types of creative “thinking” Watson can do. Elsewhere in IBM’s labyrinthine operation, Watson is advising veterans on their finances and researching potential therapies to treat cancer. A Canadian company recently launched a Watson-powered program that it says recommends medical treatments to veterinarians, “just like a colleague in the room.”
Abrams bristles at the suggestion that Watson could someday replace actual human chefs — or actual human ingenuity. “That’s not what the program was intended for,” he said, without explicitly denying the possibility.
I crunch my way through a slice of Corn Wedding Tart, secure that such a day is still pretty far off. As long as Watson is pairing frozen corn and jam, we humans still have a shot.