Catching up faster by switching sooner: a predictive approach to adaptive estimation with an application to the AIC–BIC dilemma

By Tim van Erven, Peter D. Grünwald, S. D. de Rooij. 2012

87 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Statistics and Probability, Scaling & Learning Theory

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Statistics and Probability

In Scaling & Learning Theory

Citations in the corpus

Learning and Evaluating General Linguistic Intelligence (2019, background): “A characteristic of the codelength metric is that models that perform signiﬁcantly worse at the beginning can have problems catching up to the performance of models that start well, often referred to as the catch-up phenomenon (Erven et al., 2012).”
The Description Length of Deep Learning models (2018, methodology): “The solution introduced by (Van Erven et al., 2012) against the catch-up phenomenon described in Section 3.4, is to switch between models, to always encode a data block with the best model at that point.”
A Bayesian information criterion for singular models (2013, background): “At a high level, they fall into two categories (Yang, 2005; van Erven et al., 2012; Wit et al., 2012).”
Using Stacking to Average Bayesian Predictive Distributions (with Discussion) (2017, methodology): “We suspect that in this M-closed case, stacking will behave like AIC, which, in the case of nested models, even asymptotically will select an overly large model with positive probability (for theoretical rate-of-convergence and consistency results for switching see Van der Pas and Grünwald (2018)).”
Model Selection Techniques: An Overview (2018, result): “That motivated an active line of recent advances in reconciling the two classes of model selection methods [24], [43], [52].”
Safe Testing (2019, methodology): “…the use of conditional e-variables as basic building blocks may also have advantages over general e-processes (for example, it is easier to switch (Erven et al., 2012) to a diﬀerent type of e-variable from one study to the next); sorting this out in detail remains a question for future work.”