Industry Presentations
summary
Private aggregation proposes PREAMBLE: ~100x less communication than Gaussian, shows shuffle-model needs roughly d messages per user or utility collapses, and gives a secure-aggregation construction with random projections and DPF blocks that matches central DP. Judgmental forecasting builds an LLM pipeline matching expert Brier scores, finds search improves forecasts by ~10% when agentic and ensembling plus extremization yields 5–10% gains, and warns naive retrieval can hurt. EV routing uses an auxiliary graph with discretized battery levels and spanners (1% edges → 1% error); queue monitoring proves ping-on-arrival policies are asymptotically optimal. Federated Byzantine Agreement explains quorum slices and lists open problems: leader election, erasure-code and threshold-signature generalizations.
AI-written summary · about
chapters (6)
- 0:00Industrial partnership overview
- 2:15Private aggregation in federated learning
- 20:32Judgmental forecasting with language models
- 37:27Algorithms for online navigation platforms
- 52:42Federated Byzantine Agreement overview
- 1:07:00Open research questions & Q&A
related talks
34mEfficient Robust Private Aggregation for DP
29mBuilding Probabilistic BFT Protocols Through Subsampled Voting
15mProduction Multi-Party Computation via the Distributed Aggregation Protocol
49mIntroduction and Research Fellows Lightning Talks
16mPreliminary Study on Message Aggregation Optimizations for Energy Savings in PGAS(…)
14mDistributed Private Aggregation in Graph Neural Networks
44mFederated Heavy Hitters Discovery with Differential Privacy
01h 12mTutorial: Federated Optimization, Part I