作者： David E. Rumelhart , Andreas S. Weigend
DOI:
关键词: Generalization 、 Minimum description length 、 Artificial intelligence 、 Bayesian probability 、 Prior probability 、 Network complexity 、 Time series 、 Term (time) 、 Mathematics 、 Mathematical optimization 、 Machine learning 、 Multivariate adaptive regression splines
摘要: Abstract : Inspired by the information theoretic idea of minimum description length, we add a term to usual back-propagation cost function that penalizes network complexity. From Bayesian perspective, complexity can be usefully interpreted as an assumption about prior distribution weights. This method, called weight-elimination, is contrasted ridge regression and cross-validation. We apply weight-elimination time series prediction. On sunspot series, outperforms traditional statistical approaches shows same predictive power multivariate adaptive splines.