Automatic early stopping using cross validation: quantifying the criteria

被引:644
作者
Prechelt, L [1 ]
机构
[1] Univ Karlsruhe, Fak Informat, Karlsruhe, Germany
关键词
early stopping; overfitting; cross validation; generalization; empirical study; supervised learning;
D O I
10.1016/S0893-6080(98)00010-0
中图分类号
TP18 [人工智能理论];
学科分类号
081104 ; 0812 ; 0835 ; 1405 ;
摘要
Cross validation can be used to detect when overfitting starts during supervised training of a neural network; training is then stopped before convergence to avoid the overfitting ('early stopping'). The exact criterion used for cross validation based early stopping, however, is chosen in an ad-hoc fashion by most researchers or training is stopped interactively. To aid a more well-founded selection of the stopping criterion, 14 different automatic stopping criteria from three classes were evaluated empirically for their efficiency and effectiveness in 12 different classification and approximation tasks using multi-layer perceptrons with RPROP training. The experiments show that, on average, slower stopping criteria allow for small improvements in generalization (in the order of 4%), but cost about a factor of 4 longer in training time. (C) 1998 Elsevier Science Ltd. All rights reserved.
引用
收藏
页码:761 / 767
页数:7
相关论文
共 22 条
[1]  
[Anonymous], 1992, NIPS 91 P 4 INT C NE
[2]   Temporal Evolution of Generalization during Learning in Linear Networks [J].
Baldi, Pierre ;
Chauvin, Yves .
NEURAL COMPUTATION, 1991, 3 (04) :589-603
[3]  
COWAN JD, 1994, ADV NEURAL INFORMATI, V6
[4]  
Fahlman S., 1990, ADV NEURAL INFORMATI, V2, P524
[5]  
Fahlman S., 1988, CMUCS88162 SCH COMP
[6]  
Fiesler E., 1994, INT C ART NEUR NETW
[7]   IMPROVING MODEL SELECTION BY NONCONVERGENT METHODS [J].
FINNOFF, W ;
HERGERT, F ;
ZIMMERMANN, HG .
NEURAL NETWORKS, 1993, 6 (06) :771-783
[8]   NEURAL NETWORKS AND THE BIAS VARIANCE DILEMMA [J].
GEMAN, S ;
BIENENSTOCK, E ;
DOURSAT, R .
NEURAL COMPUTATION, 1992, 4 (01) :1-58
[9]  
HANSON SJ, 1993, ADV NEURAL INFORMATI, V5
[10]  
Hassibi B., 1992, ADV NEURAL INFORM PR, V5