You would think that a pudding made of bacon, porcini and cumin, among other things, would probably be disgusting.
But you’d be wrong. Science says it’s delicious.
At least, that’s what Watson — IBM’s Jeopardy-winning supercomputer — has proved at the South by Southwest convention in Austin this week. The company’s cognitive computing group used the system to crunch through about 35,000 recipes sourced from Wikipedia and the Institute of Culinary Education to create an app that thinks up the most unusually delicious dishes based on molecular science.
The group spent two years orchestrating the project, which analyzed about 1,000 chemical flavor compounds to make educated predictions about which flavors pair together best and how surprising they might be. The result is something people have never tasted before. For example: custard with porcini-infused bacon, topped with raisins, figs, honey, sugar, orange and cumin — a culinary invention I consumed with sheer delight.
The food truck where recipes generated by IBM’s Cognitive Cooking project came to life at SXSW. Photo courtesy of IBM.
The system works like this:
First, you choose an ingredient for the recipe to be based on. It could be something you have lying around in the fridge that needs to be cooked before it goes bad or your absolute favorite food. For my run-through session with IBM, I chose artichokes (because they rule).
The system tells you which cuisines use artichokes the most. It turns out that Maltese, Italian, French and Israeli foods are most likely to contain my green friend. But I opted for a less-common Californian cuisine to honor my home state.
After that, it presents you with the types of dishes in which you’d most commonly find an artichoke. Turns out they’re usually in quiches, casseroles, stews and paella. I chose soup because I’m a rebel.
Now the system knows I’m making a California-style artichoke soup, so it goes through its database to figure out what categories of ingredients make that particular dish. It determined that I needed up to four veggies, three spices, one oil, one kind of starch, two kinds of herbs, one beverage, one dairy product and one meat. During this step, you have the option to add or remove ingredients, so if you’re a vegetarian or have a food allergy, it’ll adjust the recipe to fit your needs.
Finally, the system runs through trillions of quadrillions of possible ingredient combos that fit the parameters you set and produces a long list of options.
You can choose the one you want based on three little descriptor scales that come up when you tap a recipe: surprise, pleasantness and pairing. A dish’s surprise rating is based on how common a recipe is; its pleasantness is based on taste; and its paring is derived from something called the flavor pairing hypothesis: the idea that the more flavor compounds ingredients share, the more likely they are to taste good together in Western cuisine.
The dish I chose in the end was — drumroll please — artichoke, red bell pepper, red onion, Hass avocado (sup, California), salt, black pepper, paprika, parsley, basil, beef sirloin, soy milk, beef stock, water and olive oil. Not that surprising (as the system is full of soups), but literally 100 percent pleasant, according to Watson.
If you’re a trained chef, you could probably create your own recipe by just looking at those ingredients. But for us common folk, the system orders the food in a flow chart based on actions, which is then roughly translated into English.
“The idea is that those flavor compounds that are found in very different ingredients in the recipe could explain why it tastes good,” Florian Pinel, the senior software engineer from IBM’s Watson Group told Yahoo Tech.
This creation was something of a pet project for Pinel, who happens to have a degree from the Institute of Culinary Education.
“It helps you be more creative,” he said. “In the kitchen you might think of triplets of ingredients that work well together. But here we have lists of five, six, seven, eight ingredients that work well together.”
Ultimately, he feels the system’s processing ability signifies a new phase of culinary discovery.
“A long time ago, food didn’t taste that good and we didn’t know why,” he said. “Then we moved to a generation where the food tasted good but we didn’t know why. And now the food tastes good and we know why.”
The cognitive cooking app is still in the early stages of development, but IBM plans to release it to the public someday.