Abstract

Inverse problems are paramount in Science and Engineering. In this paper, weconsider the setup of Statistical Inverse Problem (SIP) and demonstrate howStochastic Gradient Descent (SGD) algorithms can be used in the linear SIPsetting. We provide consistency and finite sample bounds for the excess risk.We also propose a modification for the SGD algorithm where we leverage machinelearning methods to smooth the stochastic gradients and improve empiricalperformance. We exemplify the algorithm in a setting of great interestnowadays: the Functional Linear Regression model. In this case we consider asynthetic data example and examples with a real data classification problem.

 

Quick Read (beta)