DEVOURED
Epicure: Navigating the Emergent Geometry of Food Ingredient Embeddings
"Epicure" is a new family of three skip-gram ingredient embeddings, trained on 4.14 million multilingual recipes, designed to map food ingredient relationships based on co-occurrence and chemical properties.
What: Researchers Jakub Radzikowski and Josef Chen introduced "Epicure," a set of three skip-gram ingredient embeddings. These models were trained from scratch on a corpus of 4.14 million recipes spanning seven languages and normalized 1,790 canonical ingredient entries using an LLM-augmented pipeline. The three variants (Cooc, Chem, Core) differ in their random-walk schema on ingredient co-occurrence and typed FlavorDB ingredient-compound graphs.
Why it matters: This research exemplifies the application of advanced AI techniques like embeddings to highly specific, real-world domains like culinary science, demonstrating how complex relationships (flavor, chemistry) can be modeled to uncover novel insights or power recommendation systems.
Decoder
- Skip-gram: A neural network architecture used in word embedding models (like Word2Vec) that predicts context words given a target word, allowing it to learn vector representations (embeddings) of words based on their co-occurrence in text.* Embeddings: Numerical representations (vectors) of text, images, or other data that capture semantic relationships, allowing machines to process and understand them.* NPMI graph (Normalized Pointwise Mutual Information): A graph where nodes are ingredients and edges represent the statistical association or co-occurrence strength between them, often used in natural language processing to understand word relationships.
Original article
Epicure is a family of three sibling skip-gram ingredient embeddings retrained from scratch on a multilingual recipe corpus.
Explore all 61 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi