Towards Generic Interface for Human-Neural Network Knowledge Exchange

Neural Networks (NN) outperform humans in multiple domains. Yet they suffer from a lack of transparency and interpretability, which hinders intuitive and effective human interactions with them. Especially when NN makes mistakes, humans can hardly locate the reason for the error, and correcting it is even harder. While recent advances in explainable AI have substantially improved the explainability of NNs, effective knowledge exchange between humans and NNs is still under-explored. To fill this gap, we propose Human-NN-Interface (HNI), a framework using a structural representation of visual concepts as a ”language” for humans and NN to communicate, interact, and exchange knowledge. Take image classification as an example, HNI visualizes the reasoning logic of a NN with class-specific Structural Concept Graphs (c-SCG), which are human-interpretable. On the other hand, humans can effectively provide feedback and guidance to the NN by modifying the c-SCG, and transferring the knowledge back to NN through HNI. We demonstrate the efficacy of HNI with image classification tasks and 3 different types of interactions: (1) Explaining the reasoning logic of NNs so humans can intuitively identify and locate errors of NN; (2) human users can correct the errors and improve NN’s performance by modifying the c-SCG and distilling the knowledge back to the original NN; (3) human users can intuitively guide NN and provide a new solution for zero-shot learning.