A Survey on Verification and Validation, Testing and Evaluations of Neurosymbolic Artificial Intelligence

Neurosymbolic AI offers a solution to the "black box" problem: By combining symbolic AI (which is interpretable and rule-based) with sub-symbolic AI (like deep learning), neurosymbolic systems can make AI predictions more explainable, addressing a major challenge in testing and validating traditional AI systems.

Current testing and validation methods are only partially adequate: The survey finds that existing techniques for testing and evaluating AI are "partly sufficient" for neurosymbolic systems, suggesting that new or adapted approaches are needed to properly verify these hybrid AI architectures.

The symbolic component can improve verification processes: The research demonstrates that the symbolic part of neurosymbolic AI can be leveraged specifically for testing and validation purposes, providing a pathway to make AI systems more trustworthy and auditable.

Different neurosymbolic architectures require different validation approaches: The survey reveals that how symbolic and sub-symbolic components are combined significantly impacts which testing and validation methods will be effective, meaning there is no one-size-fits-all approach to responsible AI deployment.

Neurosymbolic AI is an emerging solution for responsible AI development: The paper suggests neurosymbolic approaches have "great potential" to address the verification and validation challenges that have made it difficult to deploy sub-symbolic AI systems responsibly in high-stakes applications.