|Customers: DoDo Pizza
Contractors: Moscow Institute of Physics and Technology (MIPT), Skoltech (Skolkovo institute of science and technologies, Skoltech)Product: Artificial intelligence (AI, Artificial intelligence, AI)
Project date: 2019/04 - 2019/09
2019: Creation of pizza using artificial intelligence
On October 8, 2019 DoDo Pizza reported to TAdviser that it using artificial intelligence developed the recipe of pizza which considers molecular compatibility of ingredients. Pizza received the name Open Source ("Opensors") pizza. The recipe of pizza and the source code of artificial intelligence are uploaded publicly and available to use in any purposes, including commercial.
"Опенсорс" pizza was developed by artificial intelligence which was created by Dodo jointly with neuronets specialists from MIPT and Skoltech. To learn to find unevident communications between ingredients and to understand how ingredients are combined among themselves, the artificial intelligence analyzed more than three hundred thousand recipes and also results of scientific research about molecular compatibility of ingredients.
The recipe of the pizza which is thought up by artificial intelligence includes ten ingredients — as traditional (tomato sauce, a chicken, cherry tomatoes, broccoli, mozzarella), and very unusual (a melon, pear, a tuna, mint, muesli).
"Опенсорс" pizza was created as an experiment to check whether it is possible to describe structurally such subjective phenomenon as taste and whether it is possible to teach to find artificial intelligence unusual combinations of ingredients which will seem tasty to most of people.
|Generation of recipes of culinary recipes — not the most widespread problem, generation of pizzas — especially. Collecting of a data set was the most difficult as it usually happens during the work with machine learning — there are no such bases anywhere, it was necessary to process automatically the large culinary websites, and on the websites users because of what there are all troubles like errors in the text write recipes,|
speaks one of developers Egor Baryshnikov
The initial data set included about 300,000 recipes and about 1,000 ingredients. That the neuronet could work with recipes, they needed to be described in a mathematical type. That to make it, developers submitted the recipe as a set of vectors of ingredients. A vector of ingredients were created using Skip-Gram Negative Sampling (SGNS) — an algorithm word2vec which is based on occurrence of words in a context.
For creation of recipes two recurrent neuronets were used. Developers assumed that in the general space of recipes there is a subspace which is responsible for recipes of pizzas, and that the neuronet learned to think out recipes of pizzas, it is necessary to find this subspace. It was found after developers 'banished' recipes of pizzas through two recurrent neuronets.
After the problem of creation of the recipe of pizza was solved, the criterion of molecular compatibility on the basis of results of a research of scientists from Cambridge and several universities of the USA was added to model. The research established that ingredients are best of all combined with the greatest number of the general molecular pairs. Therefore during creation of the recipe the neuronet gave preference to ingredients with similar molecular structure.
Having learned to create the recipe of pizzas, using adjustment of coefficients the neuronet can think out as classical recipes (for example, Margarita or Pepperoni), and recipes with pear, a tuna and mint, one of which became a basis for "Opensors" of pizza.
"Опенсорс" pizza will be available in a limited number (only 500 pieces) to guests of a pizzeria of Dodo on October 10, 2019 in Moscow on New suburban, 46. Dodo laid out the recipe of pizza and the source code of artificial intelligence in open access, it can be used in any purposes, including commercial.