TLDR: This research introduces AI-ODC, a novel defect classification framework specifically designed for AI-based software systems. It adapts the traditional Orthogonal Defect Classification (ODC) by adding AI-specific attributes (Data, Learning, Thinking), expanding the severity scale to include ‘Catastrophic’ for high-risk AI applications, and using AI/AIP quality characteristics for impact assessment. A case study on Machine Learning bugs demonstrated that ‘Learning’ phase defects are most common and often high-severity, while ‘Thinking’ phase defects significantly impact trustworthiness and accuracy. AI-ODC helps identify high-risk defects, guiding more effective quality assurance in AI development.
Artificial Intelligence (AI) systems have become an integral part of our daily lives, from managing emails to automating complex industrial tasks. While these systems are primarily software-based, ensuring their quality is paramount. A critical aspect of software quality assurance is defect analysis, which involves identifying, classifying, and resolving issues. However, traditional defect analysis models often fall short when applied to AI systems due to their unique attributes, such as learning capabilities, data dependencies, and decision-making processes.
A new research paper introduces a groundbreaking framework called AI-ODC (AI-based Orthogonal Defect Classification) designed specifically for AI-based software systems. This framework, inspired by the well-established Orthogonal Defect Classification (ODC) paradigm, adapts its principles to recognize and categorize the special characteristics and defects inherent in AI. The study demonstrates the feasibility and effectiveness of this modified approach in classifying AI system defects.
Adapting Defect Classification for AI
The core idea behind AI-ODC is to adjust the existing ODC model to accommodate the distinct nature of AI. The researchers introduced several key modifications:
- New AI Attribute: A crucial addition is a new classification dimension that categorizes defects based on their origin within the AI system. These categories include ‘Data’ (issues within training/testing datasets), ‘Learning’ (errors during model training), ‘Thinking’ (mistakes in inference or decision-making), and ‘Not Related’ (general software defects unrelated to AI logic).
- Expanded Severity Scale: Recognizing the potentially catastrophic impact of AI system failures, especially in critical applications like healthcare or autonomous vehicles, the severity attribute was expanded. It now includes a ‘Catastrophic’ level, making it a five-tier scale (Catastrophic, Critical, High, Medium, Low). This new level accounts for defects that could lead to irreversible damage or significant risk to human safety.
- AI-Specific Impact Areas: Instead of traditional software impact areas, AI-ODC substitutes them with characteristics pertinent to AI systems, drawing from established AI/AIP (AI Platforms) quality models. This allows for a more accurate assessment of how defects affect specific AI quality attributes like trustworthiness, accuracy, robustness, and explainability.
Real-World Application and Key Findings
To validate the AI-ODC framework, a case study was conducted using a publicly available dataset of Machine Learning bugs from the Keras framework on GitHub. The defects were classified by expert annotators according to the new AI-ODC attributes and severity levels. The results provided valuable insights:
- Defects occurring during the ‘Learning’ phase were found to be the most prevalent. These issues often relate to suboptimal batch sizes, incorrect loss functions, or optimization problems during model training.
- Defects identified in the ‘Thinking’ phase, which involve inference, logic, or decision-making, had a disproportionate effect on critical quality characteristics such as trustworthiness and accuracy. This highlights the importance of robust validation for an AI model’s output.
- The two-way classification, correlating AI attributes with severity, revealed high-risk defect clusters. For example, ‘Learning’ phase defects were frequently linked to high-severity classifications, indicating that issues during training can have significant downstream consequences.
Also Read:
- AI-Driven Automation Transforms Software Testing Efficiency
- Navigating the Landscape of Automated Code Review: A Comprehensive Analysis
Implications for AI Quality Assurance
The findings from this research have profound implications for how AI software is developed and maintained. By providing a clear connection between the origin of a defect and its potential severity and impact on AI quality characteristics, AI-ODC empowers development teams to prioritize and address the most critical issues. For instance, a ‘Thinking’ phase defect compromising trustworthiness in a healthcare AI system would demand immediate attention, while a ‘Learning’ phase defect reducing efficiency in a large-scale industrial AI might be prioritized differently based on computational costs.
The structured AI-ODC framework can be integrated into continuous integration and delivery (CI/CD) pipelines, enabling real-time assessments of defect risks and proactive quality management. Future work aims to expand the framework’s validation to more diverse AI systems, automate defect classification using Large Language Models (LLMs), and integrate it with automated issue-tracking systems for longitudinal trend analysis.
In conclusion, AI-ODC offers a much-needed, tailored approach to defect classification for AI systems, moving beyond traditional software quality models. By recognizing the unique ‘Data’, ‘Learning’, and ‘Thinking’ aspects of AI, and adapting severity and impact assessments accordingly, this framework is poised to significantly enhance the dependability and overall quality of AI solutions across various critical domains. For more details, you can read the full research paper here.
