The 'perfect combination of food and drink' prediction by AI is highly accurate and in line with the view of a professional
Although chefs and culinary researchers from all over the world are studying food and cooking methods daily, researchers at Korea University in Seoul have taught only a few combinations of ingredients and are perfect for their ingredients. We developed an AI that discovers different food and drink combinations.
KitcheNette: Predicting and Recommending Food Ingredient Pairingsusing Siamese Neural Networks
(PDF file) https://arxiv.org/pdf/1905.07261.pdf
AI predicts new food and drink pairings | VentureBeat
A research group at Korea University published a paper on a preprint server , arXiv.com , to post unreviewed papers, 'We developed an AI that predicts combinations of ingredients using machine learning.'
The AI, which the research group named ' KitcheNette ', used a training method called ' Siamese Networks '. This is one of the machine learning methods in which text and images are learned while two neural networks work in a complementary manner, and effective training can be performed even with a small amount of data that can be used as a model. It is a good method .
The research group used Siamese Networks to analyze about 6.35 million combinations of food ingredients extracted from recipes of over 1 million dishes. As a result, it is said that they succeeded in predicting the relevance of ingredients such as ' Gin & Aquabit ', ' Lime & Nopal ', and 'Wasabi & Nori' from the 'Known food pattern' for 5%. Wasabi and Nori are ingredients that are familiar to Japanese people, but “Gin and Aquabit” and “Lime and Nopal” are ingredients and drinks classified as Nordic dishes and Mexican dishes, respectively.
In addition, 'KitcheNette' was evaluated for a combination often used in cocktails such as 'squeezed juice of champagne and orange' and 'squeezed juice of orange and sparkling wine'. The combination that you do not usually do, such as 'is evaluated low.
In addition, I instructed 'KitcheNette' to find ingredients suitable for alcohol such as red wine and white wine, but recommended combinations such as 'beef stock (red bean soup for red wine)' and 'mussels for white wine'. These combinations are highly accurate matches to the combination of food and drink mentioned in the book “
In addition, 'KitcheNette' has found even unknown combinations that are not found in existing recipes. The image below is a comparison of the ingredients recommended by 'KitcheNette' to match 'red wine' and the combinations recommended by Donnenberg's book. On the right side is a list of ingredients that match the red wines selected by professionals, from the top in alphabetical order such as 'beef', 'cheese' and ' hunting meat '.
On the other hand, the ingredients recommended by 'KitcheNette' are listed on the left side, and 'beef_stock (beef stock)', 'beef_cheeks', 'lamb_shank', etc. are arranged in descending order of rank. It's out.
Although some ingredients recommended by 'KitcheNette' are somewhat novel, we have also discovered combinations of new ingredients not found in existing recipes. Food combinations marked with '*' are unknown combinations, from 'pan_juice' to 'saltpeter (
In addition, the ingredients that match the sake recommended by 'KitcheNette' are this. Japanese foods and seasonings are selected in Japan, regardless of whether they can be used as snacks, such as seasoning 'Mirin' or ingredients such as 'Katakuri'.
In order to make the function more practical in the future, the research group seems to be planning to develop a learning model using information obtained by chemically analyzing food.