Angluin, D. (1992). Computational learning theory: survey and selected bibliography. Proceedings of the twenty-fourth annual ACM symposium on Theory of computing (pp. 351--369). New York: ACM Press.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....specified and values in the case of PAC learners [Val84] We, however, have a firm total budget, specified before the learning begins. Budgeted learning falls under the framework of bounded rationality (e.g. RS95] and is an instance of active learning and cost sensitive learning (e.g. [Ang92, CAL94, Tur00, GGR02]) Feature costs in [Tur00, GGR02] refer to costs occuring at classification time, while we are concerned with cost during the learning phase. In typical poolbased active learning, the learner knows the feature values but not the label of the instances in the pool. Our problem is closer to unknown ....

D. Angluin. Computational learning theory: survey and selected bibliography. In Proc. 24th Annu. ACM Sympos. Theory Comput., pages 351--369. ACM Press, New York, NY, 1992.

....of ineciency and insigni cant gains (speci cally see [LMRar] However, we observe that in our case the greedy method often has poor performance, and that looking deeper can pay signi cant dividends. Budgeted learning is also related to cost sensitive learning and active classi cation (e.g. [Ang92, Tur00, GGR02]) although feature costs in [Tur00, GGR02] refer to costs at classi cation time, while we are concerned with cost during the learning phase. Nonethe Xn H H Hj Figure 1: Na ve Bayes Structure less, like several active learning results[LMRar, ....

D. Angluin. Computational learning theory: survey and selected bibliography. In Proc. 24th Annu. ACM Sympos. Theory Comput., pages 351-369. ACM Press, New York, NY, 1992.

....sub H iv 1 Introduction 1.1 Overview A central topic in query learning is to determine which classes of Boolean formulas are efficiently learnable. A number of classes have been shown to be learnable in polynomial time using membership and equivalence queries (for an incomplete survey, see [4]) For classes that are difficult to learn, we would like to know whether the difficulty is purely informational (it takes more than a polynomial number of queries to learn the class) or whether it is computational (a polynomial number of queries may suffice, but it is difficult to generate and ....

Angluin, D. (1992), Computational learning theory: Survey and selected bibliography, in "Proceedings of the 24th Annual ACM Symposium on Theory of Computing," pp. 351--369.

....f is symmetric if the value of f depends only on the number of inputs that are 1. For a symmetric function f and integer i, we let f(i) denote the value of f when exactly i of its inputs are 1. We study learning in the distribution free or Probably Approximately Correct (PAC) learning model [74, 2]. In the PAC learning model, we assume that the learning algorithm has available an oracle EXAMPLES(c) that when queried, produces a labeled example h v; c( v)i according to a fixed but unknown probability distribution D. If C and H are concept classes, we say that algorithm A learns C by H if for ....

Dana Angluin. Computational learning theory: Survey and selected bibliography. In Proceedings of the Twenty-Fourth Annual ACM Symposium on Theory of Computing, pages 351--369, May 1992.

....algorithms and programs which can learn concepts from positive and negative examples is an important goal of artificial intelligence, and it has motivated much research in the field. In this thesis, we approach the problem of concept learning from the perspective of computational learning theory [1] in that we are concerned with the ability to learn concepts efficiently. A machine or algorithm for concept learning is said to be efficient if the quantities of resources it uses (e.g. time, space, examples, etc. are bounded by polynomials 11 in the various learning parameters. Many models of ....

....examples of that concept. In the above example, for instance, the known concept class may be circular regions in the plane, and the unknown target concept is circular region of radius 1 centered at the origin. A great many algorithms have been devised to PAC learn various concept classes [1]; however, nearly all of these algorithms are brittle in the sense that they cannot tolerate noisy data. In Part I of this thesis, a model of PAC learning is studied which is both general in the sense that it encompasses nearly all known PAC learning algorithms and robust in the sense that many ....

[Article contains additional citation context not shown here]

Dana Angluin. Computational learning theory: Survey and selected bibliography. In Proceedings of the 24 tn Annual ACM Symposium on the Theory of Computing, 1992.

....classes. 68T05 Learning and adaptive systems. 1. Introduction Inferring an unknown function from examples of its input output behavior is a fundamental problem in scientific investigation. Recent work has approached this problem from the standpoint of computational complexity [59, 8, 27]. One focus has been to determine those types of Boolean formulas that may be efficiently inferred by algorithms that interact with their environment, or with a teacher, in specified ways. In this paper, we consider a standard learning model called exact learning with membership and equivalence ....

D. Angluin, Computational learning theory: survey and selected bibliography. In Proc. 24th Annu. ACM Sympos. Theory Comput., pages 351--369, ACM Press, New York, 1992.

....algorithms and programs which can learn concepts from positive and negative examples is an important goal of artificial intelligence, and it has motivated much research in the field. In this thesis, we approach the problem of concept learning from the perspective of computational learning theory [1] in that we are concerned with the ability to learn concepts efficiently. A machine or algorithm for concept learning is said to be efficient if the quantities of resources it uses (e.g. time, space, examples, etc. are bounded by polynomials 11 in the various learning parameters. Many models of ....

....examples of that concept. In the above example, for instance, the known concept class may be circular regions in the plane, and the unknown target concept is circular region of radius 1 centered at the origin. A great many algorithms have been devised to PAC learn various concept classes [1]; however, nearly all of these algorithms are brittle in the sense that they cannot tolerate noisy data. In Part I of this thesis, a model of PAC learning is studied which is both general in the sense that it encompasses nearly all known PAC learning algorithms and robust in the sense that many ....

[Article contains additional citation context not shown here]

Dana Angluin. Computational learning theory: Survey and selected bibliography. In Proceedings of the Twenty-Fourth Annual ACM Symposium on Theory of Computing, pages 351--369, May 1992.

....5. 3 Problems in the synthesis of state diagrams The state machine synthesis has been treated as an inductive inference problem by Koskimies and M akinen [54] Inductive inference studies algorithms for inferring rules from their application instances, usually considering functions and languages [3, 4]. Applying the BK algorithm to the state diagram synthesis, however, differs from a typical inductive inference, causing inaccuracies to the synthesized state machine. For this reason, learning results of the BK algorithm do not necessarily hold for state diagrams. The main source of problems is ....

Angluin D., Computational learning theory: survey and selected bibliography, In Proc. 24th Ann. ACM Symp. on the Theory of Computing, 1992, pp. 351--369.

....Web site management systems. AQUA can be used as an XML graphical query environment. While other graphical query languages such as XML GL[2] require users to specify queries, AQUA infers queries from instance based example operations. Many results in the area of computational learning theory[1] should be applicable to development of improved algorithms for AQUA. Introduction of negative examples and various techniques for active learning is one of the improvements we are planning. ....

D. Angluin. Computational Learning Theory: Survey and Selected Bibliography. Proc. 24th Annual ACM Symposium on Theory of Computing, pp. 351-369, 1992.

....Proposition 6 There are some concept classes C (together with A, P , err( c( such that nding the optimal active classi er is trivial if we are given 2 C, but otherwise is not known to be possible. Proof: This claim reduces to the fact that not everything is known to be PAC learnable [Ang92] because, if all costs c(x i ) are zero, the classi er can ask for all attributes and then it will classify optimally if and only if it can identify the concept. The preceding claims (Proposition 4, Theorem 5 and Proposition 6) show that, while the learn then optimize approach is certainly ....

D. Angluin. Computational learning theory: survey and selected bibliography. In STOC-92, pages 351-369, 1992.

....the reader to [84, 52] for details. Another very important variant is that in which the learning algorithm can ask questions concerning, for example, the classification of a chosen example; see, for example [2, 26, 49] Such query learning is an active area of research, and the paper by Angluin [3] provides a good survey. In a similar vein, one may ask how much easier learning becomes when there is a helpful teacher providing cleverly chosen examples as training sample. This is very different from the PAC model in that the training examples are no longer randomly chosen. Such models of ....

D. Angluin. Computational learning theory: survey and selected bibliography. In Proc. 24th Annu. ACM Sympos. Theory Comput., pages 351--369. ACM Press, New York, NY, 1992.

....for each concept is called a representation class. By learning a representation class we mean to find a representation for the concept of interest with limited access to it. Among the representation classes which have received much attention are DFAs, DNFs and CIRCUITS. See the article by Angluin [Ang92] for a survey on learning theory. A major area of research in computational learning theory is the classification of different representation classes with respect to the difficulty of learning them in any reasonable learning model. In this direction many interesting results are known. In ....

D. Angluin. Computational learning theory: survey and selected bibliography. In Proc. of the 24th ACM Symposium on Theory of Computing, pages 351--369, 1992.

....examples (and sometimes in place of random labelled examples) was mentioned by Valiant (1984a) and has been studied extensively in recent years. We refer the reader to the papers by Angluin (1988) Angluin, Frazier and Pitt (1990) Baum (1990,1991) Maass and Turan (1990,1992) and the survey by Angluin (1992) and the references therein. 8.3 The graph dimension As far as applications to arti cial neural networks are concerned, the most signi cant and important extension of the basic pac model is to the learning of general function spaces. The basic model concerns f0; 1g valued functions only; that ....

Angluin (1992): D. Angluin, Computational learning theory: survey and selected bibliography, in Proceedings of the Twenty-Fourth Annual ACM Symposium on the Theory of Computing.

.... 1978) However, using Angluin s query learning, deterministic finite automata can be identified in polynomial time (Angluin 1987) Learning of context free grammars is computationally hard even with query learning (Angluin and Kharitonov 1991) For further details of the theoretical background see Angluin (1992). These pessimistic results led, Lucas (1993) suggests, to a general stagnation in the growth of new algorithms . However, their relevance to the practicability of grammatical inference is debatable. Gold s and Angluin s learning models demand exact identification of the language, whereas in ....

Angluin, D., 1992, Computational learning theory: survey and selected bibliography.

.... system examine in order to be confident that it will output a satisfactory wrapper Computational learning theory , a subfield of the machine learning and theoretical computer science communities, provides a rigorous foundation for investigating the expected performance of learning systems; see [Angluin 92] for a survey. We have applied these techniques to our wrapper induction application. Our results (see Section 4.6) are statistical in nature. We have developed a model which predicts how many examples our learning algorithm must observe to ensure that, with high probability, the algorithm s ....

Angluin, D. Computational learning theory: survey and selected bibliography. In Proc. 24th ACM Symp. Theory Comp., pages 351--69, 1992.

....may not benefit from repeatedly asking the same question. To our knowledge, this is the first algorithm for any concept class capable of coping robustly with such errors. CLASSIC LEARNING 11 1.5. Comparison to Previous Results Automating propositional concept discovery has been well studied (Angluin, 1992). In comparison, efficient first order learnability has been less well studied. Even so, some results are known. Cohen (1993b) gives a PAC learning algorithm for function free, two clause, closed, linearly recursive, ij determinant logic programs; he also shows (Cohen, 1993a) that when the ....

....classification and says I don t know. We sketch an argument that the weak union of Classic concepts is learnable. The idea of learning a (weak) union of Classic sentences using this method was suggested by Rob Schapire (private communication) and is reminiscent of other approaches (Page, 1993; Angluin et al. 1992). Let T be the set of concepts constituting the target. We maintain a set U of concepts as our hypothesis by replacing any one u 2 U with Prune(c Theta u) where c Theta u is a positive example and c is a (necessarily positive) counterexample to U . If no u 2 U satisfies this property, then ....

Angluin, D. (1992). Computational learning theory: survey and selected bibliography. Proceedings of the Twenty Fourth Annual ACM Symposium on Theory of Computing (pp. 351--369). Victoria, BC: ACM Press.

....clear, though, that this target function is very complex in terms of the input space and, in particular, may depend on relational information and even quanti ed predicates. Many years of research in learning theory, however, have shown that ecient learnability of complex function is not feasible [Ang92]. In the learning scenario described here, therefore, learning will not take e ect directly in terms of the raw input. Rather, we will learn the target de nitions in terms of an intermediate representation that will be generated from the input image. This will allow us learn a simpler functional ....

D. Angluin. Computational learning theory: survey and selected bibliography. In Proc. 24th Annu. ACM Sympos. Theory Comput., pages 351-369. ACM Press, New York, NY, 1992.

No context found.

Angluin, D. (1992). Computational learning theory: survey and selected bibliography. Proceedings of the twenty-fourth annual ACM symposium on Theory of computing (pp. 351--369). New York: ACM Press.

No context found.

D. Angluin. Computational Learning Theory: Survey and Selected Bibliography. In N. Alon, editor, Proceedings of the 24th Annual ACM Symposium on the Theory of Computing, pages 351--369, Victoria, B.C., Canada, May 1992. ACM Press.

No context found.

Dana Angluin. Computational learning theory: Survey and selected bibliography. In Proceedings of the 24th Annual ACM Symposium on the Theory of Computing, pages 351-369, Victoria, BC, Canada, May 1992.

No context found.

See, e.g., D. Angluin, "Computational learning theory: survey and selected bibliography ", in Proceedings of the 24th Annual ACM Symposium on Theory of Computing, Victoria, British Columbia, Canada, 4--6 May 1992.

No context found.

D. Angluin, "Computational learning theory: survey and selected bibliography," in Proceedings of the 24th Annual ACM Symposium on the Theory of Computing, ACM Press: NewYork, 1992, pp. 351--369.

No context found.

Dana Angluin. Computational learning theory: Survey and selected bibliography. In Proceedings of the 24th Annual ACM Symposium on the Theory of Computing, pages 351-369, Victoria, BC, Canada, May 1992.

No context found.

D. Angluin, `Computational learning theory: survey and selected bibliography', Proc. 24th Ann. ACM Symp. on the Theory of Computing, 1992, pp. 351--369.

No context found.

) D. Angluin. Computational Learning Theory: Survey and Selected Bibliography. In Proceedings of the 24th Annual ACM Symposium on the Theory of Computing, pages 351--369. ACM Press, 1992.

First 50 documents  Next 50

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC