Robust Training in High Dimensions via Block Coordinate Geometric Median Descent
Acharya, Anish, Hashemi, Abolfazl, Jain, Prateek, Sanghavi, Sujay, Dhillon, Inderjit S., Topcu, Ufuk
Geometric median (\textsc{Gm}) is a classical method in statistics for achieving a robust estimation of the uncorrupted data; under gross corruption, it achieves the optimal breakdown point of 0.5. However, its computational complexity makes it infeasible for robustifying stochastic gradient descent (SGD) for high-dimensional optimization problems. In this paper, we show that by applying \textsc{Gm} to only a judiciously chosen block of coordinates at a time and using a memory mechanism, one can retain the breakdown point of 0.5 for smooth non-convex problems, with non-asymptotic convergence rates comparable to the SGD with \textsc{Gm}.
Jun-16-2021
- Country:
- North America > United States
- Virginia (0.04)
- Massachusetts (0.04)
- Texas > Travis County
- Austin (0.04)
- Illinois > Cook County
- Chicago (0.04)
- Europe
- Switzerland > Neuchâtel
- Neuchâtel (0.04)
- Spain > Catalonia
- Barcelona Province > Barcelona (0.04)
- Switzerland > Neuchâtel
- Asia
- Japan > Honshū
- Tōhoku (0.04)
- Afghanistan > Parwan Province
- Charikar (0.04)
- Japan > Honshū
- North America > United States
- Genre:
- Research Report (0.82)
- Industry:
- Information Technology > Security & Privacy (0.68)
- Technology: