Result-Based Re-computation for Error-Tolerant Classification by a Support Vector Machine

Abstract

Support Vector Machines (SVMs) are widely used in Machine Learning (ML) to perform classification. For a given element, an SVM computes a value using a kernel function and several support vectors to determine its class. Unfortunately, the computational units can be affected by externally induced phenomena such as soft errors; therefore, an erroneous result can be obtained. This is an issue when an SVM is used in safety-critical applications in which a change of the classification result is not acceptable. To ensure that errors do not change the classification result, traditional protection schemes can be used. For example, if the SVM computation is done on a processor, calculations can be executed twice, and the results get compared. If they are different, the error is detected, and the calculation can be done for a third time, voting can then be utilized to obtain the most likely result. However, this approach incurs in a large cost for computing resources and may not be acceptable when an SVM is used in resource constrained platforms such as Internet of Things (IoT) devices. In this paper, Result-based Re-computation (RBR) is proposed; RBR is an efficient technique to protect SVMs from errors in the kernel function, which is the most complex part in the SVM implementation. RBR is based on the observation that the SVM result is a sum of kernel terms to detect the terms that can modify the classification result and only these terms must be re-computed. The evaluation results using several publicly available datasets show that compared to a traditional protection scheme, the proposed RBR reduces up to 95.58% of the re-computation needed to protect an SVM against errors. Impact Statement-Error tolerance is critical in safety-critical Artificial Intelligence (AI) applications due to their dramatic consequences and safety implications. However, conventional protection solutions always incur in a large computational cost that could become prohibitive in resource-constrained platforms and application domains such as Internet of Things devices. The Algorithm-based Error Tolerance (ABET) approach for Support Vector Machines (SVMs) proposed in this paper has a significant advantage in terms of low computational demands. This advantage makes it very attractive in practice. The proposed algorithm reduces computational time with up to 95.58%, when compared to a classic protection scheme. This saving will promote the use of ABET approaches in industry, especially in embedded low-power systems.