Approximate coding computing
summary
Describes concrete schemes that trade fidelity for speed by producing an approximate matrix product in stages then exact recovery later, using MatDot-like polynomial encodings, complex-valued evaluations, random permutation of subblocks and grouping. Simulations with 40 workers multiplying 100×8000 matrices (K=8, standard threshold 15) show discrete error jumps: first useful approximation after 6 or 8 workers, improved at later arrivals. Proposes successive approximation coding with a three-group design and alternatives using OrthoMatDot bases; works in simulation but has numerical stability and wall-clock evaluation caveats.
AI-written summary · about
chapters (12)
- 0:00Coded computing research
- 1:02Approximate coded computing
- 3:30Straggler approaches
- 6:03MatDot codes
- 8:29Epsilon MatDot codes
- 13:55Early recovery via polynomials
- 16:47Layered coding design
- 21:58Orthogonal polynomials
- 24:49Coding schemes comparison
- 25:32Open questions
- 26:37Conclusion
- 26:59Q&A
related talks
28mCommunication Avoiding Algorithms for Linear Algebra and Beyond
25mSimple, Efficient and Neural Algorithms for Sparse Coding
57mCoding for Interactive Communication
44mStochastic Integration via Error-Correcting Codes
01h 31mPast, Present and Future of Randomized Numerical Linear Algebra II
28mRandomized Constrained Matrix Decompositions
49mAnalysis of Algorithms on Dense Matrices using Approximate Message Passing
44mTutorial on List Decodable Coding Schemes and Interactive Communication in Networks (Part 2 of 2)
