Exploiting Transformer Models in Three-Version Image Classification Systems
Shamima Afrin, Fumio Machida · 2024
Machine learning (ML) models are extensively employed in a wide range of real-world applications, including safety-critical ones. The reliability of ML application systems is a critical concern, particularly in situations where incorrect system outputs lead to severe consequences. This paper aims to enhance the reliability of ML-based image classification systems by exploiting a transformer-based model with non-transformer models in two distinct architectural frameworks, specifically the Majority Voting (MV) and Recovery Block (RB). Instead of relying on a single ML prediction, the proposed architectures leverage multiple ML models, executed simultaneously for the same input, to improve system output reliability. Our exper-imental results on image classification tasks show that three-version systems employing transformer models exhibit reliability enhancements in both MV and RB architectures. Moreover, considering performance overhead imposed by transformer models, we evaluated the response times as a performance metric of ML systems. The evaluation results show that RB architecture proves to have shorter response times than MV architecture.