## On the Complexity of Teaching (1992)

Venue: | Journal of Computer and System Sciences |

Citations: | 102 - 2 self |

### BibTeX

@ARTICLE{Goldman92onthe,

author = {Sally A. Goldman and Michael J. Kearns},

title = {On the Complexity of Teaching},

journal = {Journal of Computer and System Sciences},

year = {1992},

volume = {50},

pages = {303--314}

}

### Years of Citing Articles

### OpenURL

### Abstract

While most theoretical work in machine learning has focused on the complexity of learning, recently there has been increasing interest in formally studying the complexity of teaching . In this paper we study the complexity of teaching by considering a variant of the on-line learning model in which a helpful teacher selects the instances. We measure the complexity of teaching a concept from a given concept class by a combinatorial measure we call the teaching dimension. Informally, the teaching dimension of a concept class is the minimum number of instances a teacher must reveal to uniquely identify any target concept chosen from the class. A preliminary version of this paper appeared in the Proceedings of the Fourth Annual Workshop on Computational Learning Theory, pages 303--314. August 1991. Most of this research was carried out while both authors were at MIT Laboratory for Computer Science with support provided by ARO Grant DAAL03-86-K-0171, DARPA Contract N00014-89-J-1988, NSF Gr...

### Citations

11431 |
Computers and Intractability, A Guide to the Theory of NPCompleteness
- Garey, Johnson
- 1979
(Show Context)
Citation Context ... teach and an integer k. The question is then: Is there a teaching sequence for csof length k or less? We now show that this problem is equivalent to the minimum cover problem. (See Garey and Johnson =-=[5]-=- for a formal description of the minimum cover problem.) Theorem 8 The optimal teaching sequence problem is equivalent to a minimum cover problem in which there are jCj \Gamma 1 objects to be covered ... |

1745 | A theory of the learnable
- Valiant
- 1984
(Show Context)
Citation Context ...] have shown that this combinatorial measure of a concept class characterizes the number of examples required for learning any concept in the class under the distribution-free or PAC model of Valiant =-=[19]-=-. In the next section we briefly discuss some related work. In Section 4 we compare the teaching dimension to both the Vapnik-Chervonenkis dimension and Natarajan's dimension measure. Furthermore, we ... |

974 |
On the uniform convergence of relative frequencies of events to their probabilities. Theory Probab
- Vapnik, Chervonenkis
- 1971
(Show Context)
Citation Context ...nstances a teacher must reveal to uniquely identify any target concept chosen from the class. We show that this new dimension measure is fundamentally different from the Vapnik-Chervonenkis dimension =-=[3, 11, 20] and the d-=-imension measure of Natarajan [13]. While we show that there is a concept class C for which the teaching dimension is jCj \Gamma 1, we prove that in such cases there is one "hard-to-teach" c... |

668 |
Queries and concept learning
- Angluin
- 1988
(Show Context)
Citation Context ...ence is only required to eliminate those hypotheses that are not "close" to the target. Such a model would use a PAC-style [19] success criterion for the learner versus the exact-identificat=-=ion-style [1]-=- criterion that we have used here. In other related work, Anthony, Brightwell, Cohen, and Shawe-Taylor [2] define the specification number of a concept c 2 C to be the cardinality of the smallest samp... |

634 |
Learnability and the Vapnik-Chervonenkis dimension
- Blumer, Ehrenfeucht, et al.
- 1989
(Show Context)
Citation Context ...nstances a teacher must reveal to uniquely identify any target concept chosen from the class. We show that this new dimension measure is fundamentally different from the Vapnik-Chervonenkis dimension =-=[3, 11, 20] and the d-=-imension measure of Natarajan [13]. While we show that there is a concept class C for which the teaching dimension is jCj \Gamma 1, we prove that in such cases there is one "hard-to-teach" c... |

389 | Learning Decision Lists
- Rivest
- 1987
(Show Context)
Citation Context ...er the concept class of monotone decision lists. Let V n = fv 1 ; v 2 ; . . . ; v n g be a set of n Boolean variables. Let the instance space X n = f0; 1g n . The class C n of monotone decision lists =-=[14]-=- is defined as follows. A concept c 2 C n is a list L = h(y 1 ; b 1 ); . . . (y ` ; b ` )i where each y i 2 V n and each b i 2 f0; 1g. For an instance x 2 X n , we define L(x), the output from L on in... |

155 |
A greedy heuristic for the set covering problem
- Chvátal
- 1979
(Show Context)
Citation Context ...ing sequence problem is NP-hard even if it is known that each instance in X is a positive example for at most three concepts from C. While it is NP-complete to compute a minimum set covering, Chvatal =-=[4]-=- proves that the greedy algorithm (which is a polynomial-time algorithm) computes a cover that is within a logarithmic factor of the minimum cover. While this problem appears to be well understood, th... |

88 |
The knowledge complexity of interactive proofs
- Goldwasser, Micali, et al.
- 1989
(Show Context)
Citation Context ...oses examples tailored to a particular learner. To avoid collusion between the teacher and learner, they consider the interaction between the teacher and learner as a modified prover-verifier session =-=[9]-=- in which the learner and teacher can collude, but no adversarial teacher can cause the learner to output an hypothesis inconsistent with the sample. While it appears that the teacher's knowledge of t... |

34 |
On learning boolean functions
- Natarajan
- 1987
(Show Context)
Citation Context ...nt and NSF Grant CCR-9110108. 1 Introduction While most theoretical work in machine learning has focused on the complexity of learning, recently there has been some work on the complexity of teaching =-=[7, 8, 13, 16, 17]-=-. In this paper we study the complexity of teaching by considering a variant of the on-line learning model in which a helpful teacher selects the instances (this is the teacher-directed learning model... |

31 | Learning binary relations and total orders
- Goldman, Schapire
- 1993
(Show Context)
Citation Context ...nt and NSF Grant CCR-9110108. 1 Introduction While most theoretical work in machine learning has focused on the complexity of learning, recently there has been some work on the complexity of teaching =-=[7, 8, 13, 16, 17]-=-. In this paper we study the complexity of teaching by considering a variant of the on-line learning model in which a helpful teacher selects the instances (this is the teacher-directed learning model... |

30 |
On the complexity of learning from counterexamples
- Maass, Turan
- 1989
(Show Context)
Citation Context ...elta \Delta \Delta ; n \Gamma 1g d We now consider the concept class of orthogonal rectangles in f0; 1; \Delta \Delta \Delta ; n\Gamma1g d . (This is the same as the class box d n of Maass and Tur`an =-=[12]-=-.) Theorem 14 4 For the concept class C d of orthogonal rectangles in f0; 1; \Delta \Delta \Delta ; n\Gamma1g d : td(C d ) = 2 + 2d: Proof: We build the following teaching sequence T . Select any two ... |

29 | Exact identification of circuits using fixed points of amplification functions
- Goldman, Kearns, et al.
- 1993
(Show Context)
Citation Context ...oncept in the class. Note, however, that the optimal teaching sequence is allowed to vary with the target concept, as opposed to the universal identification sequences of Goldman, Kearns and Schapire =-=[6]. Fin-=-ally, we define the Vapnik-Chervonenkis dimension [20]. Let X be the instance space, and C be a concept class over X. A finite set Y ` X is shattered by C if fc " Y j c 2 Cg = 2 Y . In other word... |

28 | Results on learnability and the Vapnik-Chervonenkis dimension
- Linial, Mansour, et al.
- 1988
(Show Context)
Citation Context ...nstances a teacher must reveal to uniquely identify any target concept chosen from the class. We show that this new dimension measure is fundamentally different from the Vapnik-Chervonenkis dimension =-=[3, 11, 20] and the d-=-imension measure of Natarajan [13]. While we show that there is a concept class C for which the teaching dimension is jCj \Gamma 1, we prove that in such cases there is one "hard-to-teach" c... |

25 |
On exact specification by examples
- Anthony, Brightwell, et al.
- 1992
(Show Context)
Citation Context ... use a PAC-style [19] success criterion for the learner versus the exact-identification-style [1] criterion that we have used here. In other related work, Anthony, Brightwell, Cohen, and Shawe-Taylor =-=[2]-=- define the specification number of a concept c 2 C to be the cardinality of the smallest sample for which only c is consistent with the sample. Thus the specification number is just the length of the... |

25 | A computational model of teaching
- Jackson, Tompkins
- 1992
(Show Context)
Citation Context ...e by examples if the teaching dimension is polynomially bounded. The primary focus of their work is to establish a relationship between learnability and teachability 1 . Recently, Jackson and Tomkins =-=[10]-=- have considered a variation of our teaching model in which they can study teacher/learner pairs in which the teacher chooses examples tailored to a particular learner. To avoid collusion between the ... |

21 |
Learning with a helpful teacher
- Salzberg, Delcher, et al.
- 1991
(Show Context)
Citation Context ...nt and NSF Grant CCR-9110108. 1 Introduction While most theoretical work in machine learning has focused on the complexity of learning, recently there has been some work on the complexity of teaching =-=[7, 8, 13, 16, 17]-=-. In this paper we study the complexity of teaching by considering a variant of the on-line learning model in which a helpful teacher selects the instances (this is the teacher-directed learning model... |

13 |
Testing geometric objects
- Romanik, Smith
- 1990
(Show Context)
Citation Context |

11 |
Approximate testing and learnability
- Romanik
- 1992
(Show Context)
Citation Context ...here, the learner and teacher sometimes share such information, but we do not provide a formal characterization of what kind of information can be shared in this manner. The work of Romanik and Smith =-=[16, 15]-=- on testing geometric objects shares similarities with our work. They propose a testing problem that involves specifying for a given target concept a set of test points that can be used to determine i... |

5 |
Learning Binary Relations, Total Orders, and Read-Once Formulas
- Goldman
- 1990
(Show Context)
Citation Context |