Machine Teaching for Bayesian Learners in the Exponential Family

By Xiaojin Zhu. 2013

90 citations recorded by the source corpus.

What if there is a teacher who knows the learning goal and wants to design good training data for a machine learner? We propose an optimal teaching framework aimed at learners who employ Bayesian models. Our framework is expressed as an optimization problem over teaching examples that balance the future loss of the learner and the effort of the teacher. This optimization problem is in general hard. In the case where the learner employs conjugate exponential family models, we present an approximate algorithm for finding the optimal teaching set. Our algorithm optimizes the aggregate sufficient statistics, then unpacks them into actual teaching examples. We give several examples to illustrate our framework.

Subfields: Artificial Intelligence, Probabilistic & Bayesian Methods

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Probabilistic & Bayesian Methods

Citations in the corpus

Attack of the Tails: Yes, You Really Can Backdoor Federated Learning (2020, methodology): “It is typically used to speed up training [59, 60, 61] by choosing the optimal training sequence.”
Dataset Distillation (2018, methodology): “However, methods (Zhu, 2013; 2015) inspired by this concept need the existence of target models, which our method does not require.”
Using Machine Teaching to Identify Optimal Training-Set Attacks on Machine Learners (2015, methodology): “Our training-set attack formulation is closely related to machine teaching (Zhu 2013; Patil et al. 2014).”
Machine Teaching: An Inverse Problem to Machine Learning and an Approach Toward Optimal Education (2015, background): “This is in general still an open problem, but we now understand the solution for certain machine teaching tasks (Zhu 2013; Patil et al. 2014).”
Towards Poisoning of Deep Learning Algorithms with Back-gradient Optimization (2017, methodology): “ 20 0 5 10 15 20 25 Initial sample (y=5) 0 10 20 0 10 20 25 Poisoning sample (y=6) 0 10 20 20 25 Attack changes Figure 5: Poisoning samples targeting the CNN. based on the concept of machine teaching [29, 40]. The fact that these techniques require full re-training of the learning algorithm at each ”
Teaching Categories to Human Learners with Visual Explanations (2018, background): “To date, a variety of different approaches have been explored for modeling the teaching of students from assuming perfect learners [12, 40, 21], heuristic-based approaches [2], Bayesian models [7, 11], recurrent neural networks [23], and reinforcement learning based methods [25, 1, 37].”
Becoming the expert - interactive multi-class machine teaching (2015, methodology): “The family of methods referred to as Machine Teaching offers a general solution to the problem of teaching humans [43, 18, 42, 39, 33].”
An Overview of Machine Teaching (2018, background): “The learner can take a variety of forms, including version space learners for the theoretical study of Teaching Dimension (to be discussed later) [18], Bayesian learners [47], deep neural networks, or cognitive models which model how humans learn [33].”
The Security of Latent Dirichlet Allocation (2015, methodology): “Our framework is similar to machine teaching (Zhu 2013, Zhu 2015, Mei & Zhu 2015, Patil, Zhu, Kopec & Love 2014) where the teacher plays the role of the attacker in our framework.”
Towards Understanding Knowledge Distillation (2019, background): “Transferring knowledge via machine teaching techniques is extremely effective: perfect transfer is often possible from a small finite teaching set (Zhu, 2013; Liu & Zhu, 2016).”
Optimization Models for Machine Learning: A Survey (2019, methodology): “In [217], a method is proposed for the Bayesian learners, while [174] focuses on Generalized Context Model learners.”
The Teaching Dimension of Linear Learners (2015, background): “This has been observed before for Bayesian learners, too (Zhu, 2013).”
Toward a general, scaleable framework for Bayesian teaching with applications to topic models (2016, background): “…proposed and investigated probabilistic models of optimal example selection [4–7]; algorithmic teaching researchers have proposed deterministic methods to select examples that rule out confusable concepts [8, 9]; and machine learning researchers have become interested in machine teaching [10, 11].”
Optimal Teaching for Limited-Capacity Human Learners (2014, methodology): “Such a result would establish a conceptual link between idealization manipulations from psychology and optimal teaching procedures from machine learning [7, 6, 8, 2, 9, 10, 11].”
Machine Teaching for Inverse Reinforcement Learning: Algorithms and Applications (2018, methodology): “Under the assumption of error-free demonstrations, SCOT is a (1 − 1 /e ) -approximation to the Machine Teaching Problem for IRL (Section 3.3) for the following learning algorithms: (Lopes, Melo, and Montesano 2009; Babes et al. 2011) Proof.”