Embeddings
summary
Twitter builds an embeddings pipeline that produces regularly retrained user and content vectors, stores them in the in‑house Strato feature store, and serves them for both batch and low‑latency online lookups. The stack includes SVD/ALS, skip‑gram, two‑tower co‑embedding networks for candidate generation, automated network splitting for batch and online inference, and large‑scale ANN indexes that hit 99.9% recall in one‑tenth the time. Teams compress billions of sparse IDs (example: 500M users into ~500 dims), benchmark and visualize neighbors, and version embeddings for downstream reuse.
AI-written summary · about
chapters (11)
- 0:00Introduction and team context
- 1:47What are embeddings
- 2:50Why use embeddings
- 6:43Embeddings as model features
- 7:03Feature compression with embeddings
- 9:53Nearest neighbor search
- 14:23Transfer learning benefits
- 16:10Embeddings pipeline goals
- 17:38Embeddings pipeline stages
- 22:29Future directions and challenges
- 23:11Q&A session
related talks
38mEmbeddings! Embeddings everywhere!
27mData with Purpose: The Power of Embeddings
28m"Embeddings: How Computers Learned to Read"
32mGenerative AI Text and Multimodal Embedding Models for Real World Use Cases
38mSearch: A New Era
38mHave you met Ada? Word Embeddings with Laravel
31mEmbeddings... Embeddings Everywhere
45mBehind the hype: managing billion-scale embeddings in Elasticsearch and OpenSearch