Distribution Matching for Machine Teaching

Abstract

You may need to know the following questions before reading our manuscript.Q1: What is machine teaching? Machine teaching is an inverse problem of machine learning that aims at steering the student learner towards its target hypothesis, in which the teacher has already known the student's learning parameters. In simple terms, machine teaching can help to find the best training data for student learners automatically.Q2: What is our research problem? Previous machine teaching studies on machine teaching focused on balancing the teaching risk and cost to find those best teaching examples deriving the student model. This optimization solver is in general ineffective when the student learner does not disclose any cue (i.e. a black-box) of the learning parameters. To supervise such a teaching scenario, this paper presents a distribution matching-based machine teaching strategy.Q3: Why it is significant? Our study will help machine learning to find the best teaching examples on a blackbox setting, then the training examples will be closed-form. In this paper, we present a distribution matching perspective to resolve this issue from optimization.Q4: What is our novelty? Technically, our strategy can be expressed as a cost-controlled optimization process that finds the optimal teaching examples without further exploring in the parameter distribution of the student learner.Q5: What is the related study? The most relevant machine learning study to machine teaching is active learning. Its key assumption is that a student learner who frequently interacts with a teacher (annotator) would do better or no worse than other passive student learners who randomly solicit the training examples. Generally, active learning forwardly updates the current training model into its target, while the target is always agnostic.