Google AI ranks baking recipes and explains its predictions


One of the goals of AI researchers is to figure out how to make machine learning models more interpretable so that researchers can understand why they are making their predictions. Google says this is an improvement over considering predictions from a deep neural network without understanding what contributed to the model’s output. Researchers have shown how to build an explainable machine learning model capable of analyzing baking recipes.

The machine learning model can develop its own new recipes, and no data science expertise was needed to build the model. Sara Robinson works on AI for Google Cloud. During the pandemic, she enjoys cooking and has directed her AI skills towards this hobby. She started by collecting a recipe dataset and built a TensorFlow model to take in a list of ingredients and spit out predictions like “97% bread, 2% cake, 1% cookie.”

The model was able to categorize the recipes by type accurately, and she used it to come up with a new recipe. His model determined that the recipe was 50% cookies and 50% cake. He was nicknamed a cake. Robinson said the AI’s recipe was delicious and tasted like what she would imagine would happen if she told an AI to make a hybrid of cookies and cakes.

Robinson teamed up with another researcher to create a Baking Model 2.0 with a larger dataset, new tools, and an explainable model to provide insight into what makes cakes, cookies, and bread. The model came up with a new recipe called “breakie,” a bread cookie hybrid. The dataset used by the researchers included a long list of 16 basic ingredients and 600 recipes.

As the last part of the preprocessing, the researchers used a data augmentation trick. Data augmentation is a method of creating new training examples from the data you already have. The AI ​​was designed to be insensitive to a recipe’s serving size, so researchers would randomly double and triple ingredient amounts.

The machine learning model could predict the type of recipe and provided a dialog for researchers to name the model, how long they wanted the model to train, and indicate which input features to use in training. The result was a model that was able to correctly predict the category of a given recipe and specify importance scores for the ingredients that contributed the most to its prediction.


Comments are closed.