Bayesian neural networks for classification: how useful is the evidence framework

摘要: This paper presents an empirical assessment of the Bayesian evidence framework for neural networks using four synthetic and real-world classification problems. We focus on three issues; model selection, automatic relevance determination (ARD) use committees. Model selection criterion is only tenable if number training examples exceeds network weights by a factor five or ten. With this available examples, however, cross-validation viable alternative. The ARD feature scheme useful in with many hidden units data sets containing irrelevant variables. also as hard method. Results applying to showed that committees achieved accuracies similar best alternative methods. Importantly, was achievable minimum human intervention.