Error-Tolerant Techniques for Classifiers Beyond Neural Networks for Dependable Machine Learning

Dependability is a key requirement of machine learning (ML) systems in safety-critical applications such as vehicles, finance, and medicine domains. To handle errors in the underlying ML hardware and provide a dependable outcome, error-tolerant techniques have been extensively investigated for several high-performance ML algorithms/schemes such as neural networks. However, the dependability of other simpler algorithms that can also provide good learning performance has received significantly less attention. In this chapter, the state-of-the-art error-tolerant techniques for K Nearest Neighbors, Random Forest, and Support Vector Machine ML classifiers are reviewed. By investigating the impact of errors on the final classification results and allowing “no-impact” errors, these solutions have been designed to be very efficient in terms of protection overhead (e.g., power dissipation). The overall goal is to provide readers with a set of techniques for hardening the classifiers when they are used in safety-critical applications.