Hex, Bugs and More Physics | Emre S. Tasci

a blog about physics, computation, computational physics and materials…

"The two ideas of neural network modelling and Bayesian statistics might seem uneasy bed-fellows. Neural networks are non-linear parallel computational devices inspired by the structure of the brain. ‘Backpropagation networks’ are able to learn, by example, to solve prediction and classification problems. Such a neural network is typically viewed as a black box which finds by hook or by crook an incomprehensible solution to a poorly understood problem. In contrast, Bayesian methods are characterized by an insistence on coherent inference based on clearly defined axioms; in Bayesian circles, and ‘ad hockery’ is a capital offense. Thus Bayesian statistics and neural networks might seem to occupy opposite extremes of the data modelling spectrum."

Probable Networks and Plausible Predictions – A Review of Practical Bayesian Methods for Supervised Neural Networks *

David J.C. MacKay *

Leave a Reply