Efficient Decentralized Federated Singular Vector Decomposition
summary
Excalibur is a decentralized federated SVD system that replaces external servers with a multiplicative matrix sharing scheme and a communication-optimized workflow based on one-sided bidiagonalization. It implements rotation-based dense-matrix sampling, QR preprocessing, and overlapping all-reduces in C++ with Intel MKL inside Docker, achieves nearly lossless accuracy, reduces communication rounds by ~66%, and reports 3.1×–6.0× faster than SOTA server-aided methods on billion-scale matrices and >23000× throughput over HE-based systems.
AI-written summary · about
chapters (6)
- 0:00Federated SVD introduction
- 3:50Threat model and goals
- 8:03Matrix protection with sharing
- 12:37Decentralized workflow
- 15:27Performance results
- 18:00Conclusion and future
related talks
09mOptimizing Sparse matrix-vector multiplication on the EPAC architecture
01h 27mPast, Present and Future of Randomized Numerical Linear Algebra I
35mRandomized Algorithms for Computing Full Matrix Factorizations
18mFLASH: Towards a High-performance Hardware Acceleration Architecture for Cross-silo...
22mJuliet Hougland, Data Scientist, Cloudera
43mThe Linear Algebra of Deep Learning
58mTensor Decomposition, Sparse Representations and Applications
32mAdvanced Techniques for Low-Rank Matrix Approximation