Random sampler M-estimator algorithm with sequential probability ratio test for robust function approximation via...

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Outlier detection and robust model fitting are critical in data analysis.
• Existing neural network training methods struggle with data corrupted by outliers.
• The Random Sample Consensus (RANSAC) framework offers a basis for outlier-robust training.

Purpose of the Study:

• To develop an enhanced neural network training algorithm for accurate functional model fitting in the presence of outliers.
• To improve upon the Random Sample Consensus (RANSAC) framework for neural network training.
• To enhance both the accuracy, robustness, and time performance of outlier-robust model fitting.

Main Methods:

• A multilayered feed-forward neural network is employed for functional model fitting.
• The proposed algorithm integrates an M-estimator cost function for improved model selection.
• Wald's sequential probability ratio test is utilized for enhanced algorithmic speed.

Main Results:

• The new algorithm demonstrates superior accuracy and robustness compared to existing methods.
• Significant improvements in time performance were observed due to the statistical pretest.
• Successful evaluation on both synthetic and real-world datasets with varying outlier contamination levels.

Conclusions:

• The proposed enhanced RANSAC-based neural network training algorithm effectively handles outlier-corrupted data.
• The integration of M-estimators and sequential probability ratio tests leads to notable improvements in performance.
• This method offers a more reliable and efficient solution for practical applications requiring robust model fitting.