The IBM Watson supercomputer has already demonstrated
success in medicine and game shows. It is
planned to be used in technical support services instead of live operators. However, all these tasks are connected rather with finding the right answer to user requests based on known information. IBM believes that true artificial intelligence should be able to find creative solutions, create and invent new things, and not just analyze the old.
For the development of Watson's creative abilities, its creators
chose culinary art. This is a very convenient testing ground: cooking is a very “human”, intuitive process, poorly amenable to algorithmization and standardization. And any person from the street is able to evaluate the result. Spanish-style almond-chocolate cookies, Ecuadorian strawberry dessert, grilled tomatoes on saffron croutons - these and other dishes created by Watson have already been prepared and eaten with pleasure during the experiments. A couple of weeks ago
, a preprint of an article describing the algorithms and mathematical models that Watson uses to create original recipes
was published .
Watson's Prescription Cookies
Any creative solution must simultaneously satisfy two criteria - be new and be of high quality. Novelty is relatively easy to achieve simply by combining ingredients and processing techniques. But with the quality of the situation is much more complicated. Teaching a computer to understand what the taste, flavor, texture and appearance of the dish will be is extremely difficult.
')
The original data for Watson served several million recipes collected on the Internet. He was passed through proven natural language processing algorithms that were used to win the quiz and to teach Watson medicine. From Wikipedia, information was obtained about typical ingredients and processing techniques characteristic of cuisines of different nations of the world. Finally, Watson received a thorough knowledge of the chemistry and physiology of human perception of taste and smell.
New recipes were generated on the basis of existing ones using a genetic algorithm, the values ​​of novelty, amenity and compatibility were used as a function of fitness.
The mathematical model for estimating the novelty of a recipe is based on Bayes theorem; the so-called "
Bayesian surprise " approach was used, originally designed to simulate the behavior of the viewer when watching a video. In a nutshell, the essence of the method is that the difference between the a priori and posterior probability is met to meet some combination of products in the recipe space when adding a new one to it. So, combinations of nuts with chocolate or mustard with sausages are completely commonplace and causes almost no change in the probabilities of the various combinations. But sausages in chocolate will affect these probabilities much more significantly.
Chemistry was used to assess the pleasantness. Knowing the chemical composition of the products and the order of their mixing and processing, the computer calculated which substances would determine the taste and smell of the dish. Interestingly, the smell was much more important than the taste of the dish. Our taste perception is very much connected with smell and aroma. Man distinguishes only a few basic tastes - sour, sweet, salty, bitter. In different cultures, several more basic tastes are distinguished, for example, tart or
umami . But the variety of smells is much more and they are not limited to simple basic combinations.
Finally, the assessment of the compatibility of products also relied on a serious scientific base, in particular, on a joint study of American and British scientists "
Network of flavors and principles of combination products ", which analyzed about 50,000 recipes and constructed maps of the compatibility of products typical of kitchens of different regions .
As a result, an application was created in which you can specify a set of products, a national style and a variety of dishes, after which Watson gave out a set of recipes that can be ordered by degree of novelty, amenity and compatibility. In addition to individual dishes, Watson is able to create entire menus, achieving diversity and the right combination of dishes through the use of
thematic modeling . This is a way to build a model of a collection of text documents, which breaks the collection into themes and determines which theme each document belongs to. Watson applies this model to recipes — separate ingredients act as keywords, and recipes themselves as documents.
According to Lav Varshni, one of the authors of the methodology for modeling creative abilities, the company is already discussing the issue of using Watson with several major manufacturers of products and perfumes.