Computational learning theory

In computer science , computational learning theory (or just learning theory ) is a subfield of Artificial Intelligence Devoted to studying the design and analysis of machine learningalgorithms. [1]

Overview

Theoretical results in machine learning mainly deals with a type of inductive learning called supervised learning. In supervised learning, an algorithm is made that are labeled in some useful way. For example, the mushrooms can be described and the labels are edible. The algorithm takes these previously labeled samples and uses them to induce a classifier. This classifier is a function that assigns labels to samples that have never been previously seen by the algorithm. The goal of the supervised learning algorithm is to optimize some measure of performance such as minimizing the number of mistakes made on new samples.

In addition to performance bounds, computational learning theory studies the time complexity and feasibility of learning citation needed ] . In computational learning theory, a computation is considered feasible if it can be done in polynomial time citation needed ] . There are two types of time complexity results:

  • Positive results – Showing a certain class of functions is learnable in polynomial time.
  • Negative results – Showing that certain classes can not be learned in polynomial time.

Negative results often rely on a commonly believed, but yet unproven assumptions citation needed ] , such as:

  • Computational complexity – P ≠ NP (the P versus NP problem) ;
  • Cryptographic – One-way functions exist.

There are several different approaches to computational learning theory. These differences are based are making Assumptions about the inference principles used to generalize from limited data citation needed ] . This includes different definitions of probability (see frequency probability , Bayesian probability ) and different assumptions for the generation of samples citation needed ] . The different approaches include citation needed ] :

  • Exact learning , proposed by Dana Angluin ;
  • Probably approximately correct learning (PAC learning), proposed by Leslie Valiant ;
  • VC theory , proposed by Vladimir Vapnik and Alexey Chervonenkis ;
  • Bayesian inference ;
  • Algorithmic learning theory , from the work of E. Mark Gold ;
  • Online machine learning , from the work of Nick Littlestone.

Computational learning theory has several practical algorithms according to whom? ] . For example, PAC theory inspired boosting , VC theory led to support vector machines , and Bayesian inference led to belief networks (by Judea Pearl ).

See also

  • Grammar induction
  • Information theory
  • Stability (learning theory)
  • Error Tolerance (PAC learning)

References

  1. Jump up^ http://www.learningtheory.org/

Surveys

  • Angluin, D. 1992. Computational Learning Theory: Survey and selected bibliography. In Proceedings of the Twenty-Fourth Annual ACM Symposium on Theory of Computing (May 1992), pages 351-369. http://portal.acm.org/citation.cfm?id=129712.129746
  • D. Haussler. Probably approximately correct learning. In AAAI-90 Proceedings of the Eight National Conference on Artificial Intelligence, Boston, MA, pages 1101-1108. American Association for Artificial Intelligence, 1990. http://citeseer.ist.psu.edu/haussler90probably.html

VC dimension

  • V. Vapnik and A. Chervonenkis. The convergence of relative frequencies of events to their probabilities. Theory of Probability and its Applications, 16 (2): 264-280, 1971.

Feature selection

  • A. Dhagat and L. Hellerstein, “PAC learning with irrelevant attributes”, in Proceedings of the IEEE Symp. on the Foundation of Computer Science, 1994. http://citeseer.ist.psu.edu/dhagat94pac.html

Inductive inference

  • Gold, E. Mark (1967). “Language identification in the limit” (PDF) . Information and Control . 10 (5): 447-474. doi : 10.1016 / S0019-9958 (67) 91165-5 .

Optimal O notation learning

  • Oded Goldreich , Dana Ron . On universal learning algorithms . http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.47.2224

Negative results

  • Mr. Kearns and Leslie Valiant . 1989. Cryptographic limitations on learning boolean formula and finite automata. In Proceedings of the 21st Annual ACM Symposium on Theory of Computing, pp. 433-444, New York. ACM. http://citeseer.ist.psu.edu/kearns89cryptographic.html

Boosting (machine learning)

  • Robert E. Schapire. The strength of weak learnability. Machine Learning, 5 (2): 197-227, 1990 http://citeseer.ist.psu.edu/schapire90strength.html

Occam Learning

  • Blumer, A .; Ehrenfeucht, A .; Haussler, D .; Warmuth, MK “Occam’s razor” Inf.Proc.Lett. 24, 377-380, 1987.
  • A. Blumer, A. Ehrenfeucht, D. Haussler, and MK Warmuth. Learnability and the Vapnik-Chervonenkis dimension. Journal of the ACM, 36 (4): 929-865, 1989.

Probably approximately correct learning

  • L. Valiant. A Theory of the Learnable. Communications of the ACM, 27 (11): 1134-1142, 1984.

Error tolerance

  • Michael Kearns and Ming Li. Learning in the presence of malicious errors. SIAM Journal on Computing, 22 (4): 807-837, August 1993. http://citeseer.ist.psu.edu/kearns93learning.html
  • Kearns, M. (1993). Efficient noise-tolerant learning from statistical queries. In Proceedings of the Twenty-Fifth Annual ACM Symposium on Theory of Computing, pages 392-401. http://citeseer.ist.psu.edu/kearns93efficient.html

Equivalence

  • D.Haussler, M.Kearns, N.Littlestone and M. Warmuth , Equivalence of models for polynomial learnability, Proc. 1st ACM Workshop on Computational Learning Theory, (1988) 42-55.
  • Pitt, L .; Warmuth, MK (1990). “Prediction-Preserving Reducibility” (PDF) . Journal of Computer System and Science . 41 (3): 430-467. doi : 10.1016 / 0022-0000 (90) 90028-J .