Back-Fitting
An iterative method of fitting additive models, by fitting each term to the residuals given the rest. It is a version of the Gauss-Seidel methods of numerical linear algebra.
 
 
     
 
  Back-Propagation
is the method used to calculate the gradient vector of a fitting criterion for a feed-forward neural network with respect to the parameters (weights). Also used for a steepest-descent algorithm with the gradient vector computed in this way.
 
 
     
 
  Bartlett Test of Sphericity
Statistical test for the overall significance of all correlations with a correlation matrix
 
 
     
 
  Bayes Formular
An elementary formula of probability. If are disjoint events, and then
 
 
     
 
  Bayes Rule
is a rule which attains the Bayes risk, and so is the 'gold-standard', the vest possible for that problem.
 
 
     
 
  Bias
has two meanings. (a) The bias of an estimator is the difference between its mean and the true value. (b) For a neural network, parameters which are constants (rather than multiplying signals) are often called biases.
 
 
     
 
  BIC
has two similar meanings. Akaike (1977, 1978) introduced 'information criterion B'. Schwarz (1978) introduced something which has become known as a 'Bayesian information criterion'. Although most references mean Schwarz's BIC, to avoid confusion this is also known as SBC('Schwarz Bayes Criterion'). Both penalize the deviance by log n times the number p of free parameters for n examples, but Akaike's has O(p) terms not depending on n.
 
 
     
 
  Bootstrap
(Efron, 1979) An idea for statistical inference, using training sets created by re-sampling with replacement from the original training set, so examples may occur more than once.
 
 
     
 
  Box¡¯s M
Statistical test for the equality of the covariance matrices of the independent variables across the groups of the dependent variables. If the statistical significance is greater than the critical level(eg., 0.01), then the equality of the covariance matrices is supported. If the test shows statistical significance, then the group are deemed different and the assumption is violated
 
 
     
 
  Branch-and-Bound
bound A technique in combinatorial optimization to rule out solutions without evaluating them.