Seminar - Michael Friedlander

Data Fitting and Optimization With Randomized Sampling

Abstract: 

For many structured data-fitting applications, stochastic gradient methods offer inexpensive iterations by sampling only subsets of the data. They make great progress initially, but eventually stall. Full-gradient methods, in contrast, often achieve steady convergence, but may be prohibitively expensive for large problems. Applications in machine learning and robust seismic inversion motivate us to develop a sampling method that exhibits the benefits of both stochastic and full-gradient methods.