Data with Purpose: The Power of Embeddings
summary
Explains how embeddings convert product text and images (BERT, CLIP, ResNet) into dense vectors to power nearest-neighbor search, vector arithmetic (king − man + woman → queen) and similarity at hundreds of millions of items scale, creating a billion-vector problem. Describes GPU/TPU generation, CPU inference, vector DBs like millbus and quadrant, indexing approaches (Facebook size algorithm, Google scan), and tradeoffs—latency vs accuracy, quantization—and concrete retail uses: cataloging, personalization, seller assortment suggestions, chatbot routing and counterfeit detection.
AI-written summary · about
chapters (15)
- 0:00Introduction to embeddings
- 2:11Why embeddings matter
- 5:39From one-hot to BERT
- 7:49Text embeddings in action
- 8:53Image embeddings explained
- 10:19What is an embedding
- 12:02Common embedding operations
- 14:51Handling scale with embeddings
- 17:40Vector databases and indexing
- 19:05Embeddings in retail
- 21:36Customer experience use cases
- 23:03Driving business growth
- 24:52Efficiency and automation
- 26:17Fraud detection with embeddings
- 26:38Summary and Q&A
related talks
31mEmbeddings... Embeddings Everywhere
28m"Embeddings: How Computers Learned to Read"
46m"Embeddings - the lesser known Hero of AI"
32mDistributed Vector Databases - What, Why, and How
50mVector Ops: run vector embedding-powered apps in prod
16mEmpower AI: Building Smarter Solutions with Embeddings
23mEmbeddings: Core Concepts for Developers
27mMachine Learning for Advanced AI Search