Withdrawing an example from the training set: An analytic estimation of its effect on a non-linear parameterised model

摘要: Abstract For a non-linear parameterised model, the effects of withdrawing an example from training set can be predicted. We focus on prediction error left-out example, and confidence interval for this example. derive rigorous expression first-order expansion, in parameter space, gradient quadratic cost function, specify its validity conditions. As consequence, we approximate expressions given thereof, had been withdrawn set. show that influence model summarised by single parameter. These results are applicable to leave-one-out cross-validation, with considerable decrease computation time respect conventional leave-one-out. The paper focuses theoretical aspects question; both academic illustrations large-scale industrial examples described [9].