XAI: Essential Components and Challenges
Keywords: XAI
Abstract: Explainable AI (XAI) is of crucial significance in future Artificial General Intelligence development. Only when we build explainable AI, can we really trust AI and debug them. We first review major challenges in current XAI research. Current explainable structured modelling has low expressive power and is difficult for scaling-up learning. Interpreting intermediates results of black-box best-performing neural network models relies on specific model architectures and has made only primary steps. Then we propose essential components in the communicative XAI framework: an explainable structured model which has explicit causal chains, explicit Theory of Mind (ToM) modelling and context-aware communication message generating methods.
Submission Number: 223
Loading