DeepRecSys, лекция 4: Двухбашенные нейросети II

Lecture: Kirill Khrylchenko Seminar: Artem Matveev On the fourth week of the course, we continue our discussion of two-tower models and deep retrieval. While in the previous lecture we covered the base architecture and training of two-tower models, this time we take a deep dive into the tower architecture itself: 1. Trainable (ID-based) embeddings: their advantages and limitations 2. Content encoding and inductive bias 3. Unsupervised representation learning as a powerful alternative for learning embeddings 4. User encoders 5. Additionally, we briefly discuss deep retrieval approaches that go beyond the classical two-tower setup, including multi-interest models, mixture of logits, GPU retrieval, and generative retrieval. During the seminar, we focus on practical aspects of training neural recommender models: the main tools we use, what truly matters during training, and what the actual training code looks like in practice.