Challenges at the intersection of ML.
summary
Spam-fighting work at Facebook and Rockset shows that real-time ML needs both low query latency and low data latency, so classifiers run synchronously and must be fast enough to block bursts. Vector search (ANN, HNSW) powers nearest-neighbor clustering for images but high dimensions (≈1500) force approximation and a tradeoff of 99% accuracy for 10x speed. Key unsolved problems are incremental indexing (live inserts deteriorate indexes), hybrid queries with metadata filtering, and selectivity-aware execution (cost-based optimization vs naive rebuilds or post-filter overfetching).
AI-written summary · about
chapters (5)
- 0:00Introduction and motivation
- 4:49Focus on vector search
- 13:16Indexing vector data
- 16:47Hard problems in vector search
- 26:59Summary and closing
related talks
01h 09mVector Search: Ask Me Anything!
06mMachines and the magic of fast learning
36mKeynote: AI, ML, Cloud for ML
41mChallenges of Machine Learning at Scale
25mMachine Learning is Going Real-Time
24mVector Search at Uber
44mVector Search Beyond the Hype: Real-World Implementation and Performance Insights
38mAccelerating Research with Competitions