Abstract

We consider the machine teaching problem in a classroom-like setting whereinthe teacher has to deliver the same examples to a diverse group of students.Their diversity stems from differences in their initial internal states as wellas their learning rates. We prove that a teacher with full knowledge about thelearning dynamics of the students can teach a target concept to the entireclassroom using O(min{d,N} log(1/eps)) examples, where d is the ambientdimension of the problem, N is the number of learners, and eps is the accuracyparameter. We show the robustness of our teaching strategy when the teacher haslimited knowledge of the learners' internal dynamics as provided by a noisyoracle. Further, we study the trade-off between the learners' workload and theteacher's cost in teaching the target concept. Our experiments validate ourtheoretical results and suggest that appropriately partitioning the classroominto homogenous groups provides a balance between these two objectives.

 

Introduction (beta)

None

 

Conclusion (beta)

None