Neural network uncertainty assessment using Bayesian statistics with application to remote sensing: 3. Network Jacobians

被引:26
作者
Aires, F
Prigent, C
Rossow, WB
机构
[1] Columbia Univ, NASA, Goddard Inst Space Studies, Dept Appl Phys & Appl Math, New York, NY 10025 USA
[2] Ecole Polytech, CNRS, Meteorol Dynam Lab, IPSL, F-91128 Palaiseau, France
[3] Observ Paris, CNRS, LERMA, F-75014 Paris, France
关键词
remote sensing; uncertainty; neural networks;
D O I
10.1029/2003JD004175
中图分类号
P4 [大气科学(气象学)];
学科分类号
0706 ; 070601 ;
摘要
Used for regression fitting, neural network (NN) models can be used effectively to represent highly nonlinear, multivariate functions. In this situation, most emphasis has been on estimating the output errors, but almost no attention has been given to errors associated with the internal structure of the NN model. The complex relationships linking the inputs to the outputs inside the network are the essence of the model and assessing their physical meaning makes all the difference between a "black box'' model with small output errors and a physically meaningful model that will provide insight on the problem and will have better generalization properties. Such dependency structures can, for example, be described by the NN Jacobians: they indicate the sensitivity of one output with respect to the inputs of the model. Estimating these Jacobians is essential for many other applications as well. We use a new method of uncertainty estimate developed in the work of Aires [ 2004] to investigate the robustness of the quantities that characterize the NN structure. A regularization strategy based on principal component analysis is proposed to suppress the multicolinearities that are a major concern when analyzing the internal structure of such a model. The theory is applied to the remote sensing application already presented in the work of Aires [ 2004] and Aires et al. [ 2004].
引用
收藏
页码:D103051 / 14
页数:14
相关论文
共 28 条
[1]   Remote sensing from the infrared atmospheric sounding interferometer instrument -: 1.: Compression, denoising, and first-guess retrieval algorithms -: art. no. 4619 [J].
Aires, F ;
Rossow, WB ;
Scott, NA ;
Chédin, A .
JOURNAL OF GEOPHYSICAL RESEARCH-ATMOSPHERES, 2002, 107 (D22)
[2]   Remote sensing from the infrared atmospheric sounding interferometer instrument -: 2.: Simultaneous retrieval of temperature, water vapor, and ozone atmospheric profiles -: art. no. 4620 [J].
Aires, F ;
Rossow, WB ;
Scott, NA ;
Chédin, A .
JOURNAL OF GEOPHYSICAL RESEARCH-ATMOSPHERES, 2002, 107 (D22)
[3]   Inferring instantaneous, multivariate and nonlinear sensitivities for the analysis of feedback processes in a dynamical system: Lorenz model case-study [J].
Aires, F ;
Rossow, WB .
QUARTERLY JOURNAL OF THE ROYAL METEOROLOGICAL SOCIETY, 2003, 129 (587) :239-275
[4]  
Aires F, 2002, J APPL METEOROL, V41, P144, DOI 10.1175/1520-0450(2002)041<0144:ARNNAF>2.0.CO
[5]  
2
[6]   Neural network uncertainty assessment using Bayesian statistics with application to remote sensing: 2. Output errors [J].
Aires, F ;
Prigent, C ;
Rossow, WB .
JOURNAL OF GEOPHYSICAL RESEARCH-ATMOSPHERES, 2004, 109 (D10) :D103041-12
[7]   Neural network uncertainty assessment using Bayesian statistics with application to remote sensing: 1. Network weights [J].
Aires, F .
JOURNAL OF GEOPHYSICAL RESEARCH-ATMOSPHERES, 2004, 109 (D10) :D103031-11
[8]  
Aires F, 2002, J ATMOS SCI, V59, P111, DOI 10.1175/1520-0469(2002)059<0111:ROEBTI>2.0.CO
[9]  
2
[10]   The "weight smoothing" regularization of MLP for Jacobian stabilization [J].
Aires, F ;
Schmitt, M ;
Chedin, A ;
Scott, N .
IEEE TRANSACTIONS ON NEURAL NETWORKS, 1999, 10 (06) :1502-1510