The two-way knowledge interaction interface between humans and neural networks

Neural networks need better communication with humans: While AI systems often outperform humans at specific tasks, they currently cannot easily explain their reasoning in ways humans intuitively understand, creating a barrier to human oversight and correction when errors occur.

A new "language" for human-AI dialogue: The paper proposes using structured visual concept maps as a common language between humans and AI systems, allowing neural networks to explain their decisions through relationship diagrams that people can actually comprehend.

Humans can actively improve AI systems: Rather than being passive users, humans can directly modify and guide AI decision-making by sharing their knowledge through this interface, demonstrating that human expertise and AI capabilities can work together rather than in isolation.

Two-way interaction improves both understanding and performance: The research shows that when humans and neural networks can communicate bidirectionally through this interface, the AI system becomes more interpretable to humans while simultaneously improving its actual performance on tasks.

Explainability alone isn't enough: Previous explainable AI research focused only on making neural networks more transparent, but this work demonstrates that true responsible AI requires an interactive system where humans can actively participate in correcting and guiding AI behavior.