The management of the Mars confectionery company turned to Google with a request to create a sweet dish based on Maltesers using AI. An employee of Google Cloud AI and a fan of baking at leisure, Sarah Robinson, has created a neural network that was trained on recipes for sweet pastries — in sum, more than several hundred were used. After the training, the neural network came up with a new cooking method that implies an unusual ingredient.

Google explained that since Maltesers are widely popular in the British isles, the new dessert had to contain ingredients typical for British sweet pastries. Given this condition, when creating and training the neural network, Sarah Robinson made use of a data set based on the methods of cooking British dishes. The data set included 4 pastry types common in the UK: cakes, cookies, buns, and pastries. When building the AI model, Robinson used the Keras API for TensorFlow. She also applied AI Platform Hyperparameter Tuning to optimize hyperparameters. After selecting a perfect combination of hyperparameters, Sarah unrolled the AI model with the help of AI Platform Prediction. The model has been able to compose a list of ingredients along with their volume and form a hybrid dessert. This is how Maltesers AI Cakes were created.