Common sense testing and sanity checks in machine learning

Machine learning (ML) systems that perform admirably on traditional training and validation sets can exhibit critical shortcomings during deployment when fundamental common sense and sanity check considerations are neglected. From this lens, this paper posits that while commonly adopted metrics assess pattern fitting within training data, they may often fail to detect inconsistencies obvious to domain experts or violations of fundamental principles. To mitigate these shortcomings, the paper advocates for an expanded evaluation pipeline centered on common sense testing and sanity checks that respect physical laws, causal reasoning, domain-specific rules, etc. The scope of this review includes a structured examination of pre-deployment and post-deployment evaluation methodologies. Furthermore, the paper introduces functional (system-level robustness and reasoning-based) common sense and sanity checks and proposes a methodological framework for implementing common sense testing. Finally, the paper concludes with recurring challenges and possible future research directions.