Squashing activation functions in benchmark tests: towards eXplainable Artificial Intelligence using continuous-valued logic

Squashing activation functions perform comparably to conventional activation functions while offering improved interpretability. The paper's benchmark tests across five different network types showed that these functions achieved similar accuracy, loss, and training time metrics as standard approaches, suggesting they can be adopted without sacrificing performance.

Combining neural networks with continuous logic reduces the "black box" problem that makes AI systems difficult to understand. By integrating multi-criteria decision-making tools with continuous logic-based neural models, the researchers demonstrated a pathway toward more transparent AI systems that stakeholders can better understand and trust.

This approach maintains performance while improving safety and transparency. The paper emphasizes that achieving explainable AI through Squashing functions doesn't require trading off model effectiveness—the continuous logic framework can solve classification tasks "successfully and with high performance" while being more interpretable than traditional deep neural networks.

Squashing functions represent a practical, implementable solution for responsible AI. Rather than proposing purely theoretical improvements, the authors conducted rigorous benchmark testing showing these functions work across multiple network architectures, making this a concrete tool that practitioners can consider adopting to build more responsible AI systems.