The two-tower embedding model is a method used in model training to connect embeddings from different modalities in the same vector space. This model is often used in personalized recommendation systems, where items and user histories are the two modalities. To ensure the embeddings from different modalities can be mapped to the same space, they need to have the same dimension. In personalized recommendations, the two-tower model combines user history and context with items to generate candidate recommendations. The model architecture consists of a query tower and an item tower, which are trained together using deep learning techniques. The Hopsworks platform can be used to manage feature data when building two-tower models. Other applications, such as image and text processing, can also utilize the twin-tower model architecture. Ongoing research is exploring the extension of two-tower models to more than two modalities.
https://www.hopsworks.ai/dictionary/two-tower-embedding-model