Interpretable Machine Learning Through Teaching

Researchers propose a two-stage student-teacher framework for machine teaching of concepts.

A student network first learns from random labeled examples, then a teacher network learns to pick the smallest useful set of examples.

Because the teacher stays grounded in the student’s concept space, the communication is more interpretable than arbitrary private codes.

The approach is shown on concepts like size, color, shape, and border, such as using different red objects to teach “red.”

The goal is more human-understandable AI explanations through example-based communication that preserves the intended concept.