RANSAC algorithm with sequential probability ratio test for robust training of feed-forward neural networks

Abstract

This paper addresses the problem of fitting a functional model to data corrupted with outliers using a multilayered feed-forward neural network (MFNN). Almost all previous efforts to solve this problem have focused on using a training algorithm that minimizes an M-estimator based error criterion. However the robustness gained from M-estimators is still low. Using a training algorithm based on the RANdom SAmple Consensus (RANSAC) framework improves significantly the robustness of the algorithm. However the algorithm typically requires prolonged period of time before a final solution is reached. In this paper, we propose a new strategy to improve the time performance of the RANSAC algorithm for training MFNNs. A statistical pre-test based on Wald's sequential probability ratio test (SPRT) is performed on each randomly generated sample to decide whether it deserves to be used for model estimation. The proposed algorithm is evaluated on synthetic data, contaminated with varying degrees of outliers, and have demonstrated faster performance compared to the original RANSAC algorithm with no significant sacrifice of the robustness.