A NECESSARY CONDITION FOR LEARNING FROM POSITIVE EXAMPLES

被引:3
作者
SHVAYTSER, H [1 ]
机构
[1] CORNELL UNIV,DEPT COMP SCI,ITHACA,NY 14853
关键词
Concept learning; learning from positive examples; nonlearnability; predictable concepts; sample complexity;
D O I
10.1023/A:1022663809420
中图分类号
TP18 [人工智能理论];
学科分类号
081104 ; 0812 ; 0835 ; 1405 ;
摘要
We present a simple combinatorial criterion for determining concept classes that cannot be learned in the sense of Valiant from a polynomial number of positive-only examples. The criterion is applied to several types of Boolean formulae in conjunctive and disjunctive normal form, to the majority function, to graphs with large connected components, and to a neural network with a single threshold unit. All are shown to be nonlearnable from positive-only examples. © 1990, Kluwer Academic Publishers. All rights reserved.
引用
收藏
页码:101 / 113
页数:13
相关论文
共 10 条
  • [1] OCCAM RAZOR
    BLUMER, A
    EHRENFEUCHT, A
    HAUSSLER, D
    WARMUTH, MK
    [J]. INFORMATION PROCESSING LETTERS, 1987, 24 (06) : 377 - 380
  • [2] LEARNABILITY AND THE VAPNIK-CHERVONENKIS DIMENSION
    BLUMER, A
    EHRENFEUCHT, A
    HAUSSLER, D
    WARMUTH, MK
    [J]. JOURNAL OF THE ACM, 1989, 36 (04) : 929 - 965
  • [3] Bollobas B., 1978, LONDON MATH SOC MONO
  • [4] HAUSSLER D, 1988, 1988 P WORKSH COMP L, P42
  • [5] KEARNS M, 1987, 19TH P ACM S THEOR C, P285
  • [6] Minsky M., 1969, PERCEPTRONS
  • [7] Natarajan B. K, 1987, P 19 ANN ACM S THEOR, P296
  • [8] SHVAYTSER H, IN PRESS IEEE T PATT
  • [9] A THEORY OF THE LEARNABLE
    VALIANT, LG
    [J]. COMMUNICATIONS OF THE ACM, 1984, 27 (11) : 1134 - 1142
  • [10] VALIANT LG, 1985, 9TH P IJCAI, P550