Information

Time: 4:15 PM - 6:00 PM, Feb-08-2020

Location: Sutton North, 2nd Floor, Hilton Midtown New York

Slides: [pdf]

Abstract

Machine teaching is the inverse problem of machine learning. It aims at constructing an optimal data set according to a given target concept so that the target concept can be learned on this data set.

Based on the different types of machine teaching space, we will introduce several applications under various teaching settings:

1. Machine teaches human, i.e., supervising the crowdsourcing workers to learn and label in the form of teaching (e.g., teaching the crowdsourcing workers a concept such as labeling an image or categorizing a document);
2. Machine teaches machine, i.e., an adversary can intentionally modify the training data and enforce the training framework to end up with an ill-trained model (e.g., an adversarial attack and defense);
3. Human teaches machine, i.e., in AI system building, machine teaching can enable a machine learning system to be trained faster and more accurately via teaching by a human domain expert by simply providing labeled data and selected features.

For each teaching setting, we will provide a comprehensive review of existing techniques, and discuss the related applications.

Tutorial Outline

1. Introduction (10 minutes)

• Concepts of machine teaching.
• Applications and examples of using machine teaching.

2. Part I: Machine teaches machine (30 minutes)

• Training set poisoning attack on classification and regression.
• Training set poisoning attack on auto-regressive models.

3. Part II: Machine teaches human (45 minutes)

• The concept of crowd teaching.
• Incremental/iterative crowd teaching (for sequential learners)
• Global crowd teaching (for batch learners)

4. Part III: Human teaches machine (10 minutes)

• Concept teaching overview.
• Platform for interactive concept learning.

5. Future Outlook and Open Problems (10 minutes)

Tutors