No one uses generalization bounds to do model selection; practitioners use cross-validation.

Read 'no one' as a claim about common practice in 2014, not a literal count. Generalization bounds means theoretical bounds such as VC, Rademacher or PAC-Bayes bounds and structural risk minimization. AIC/BIC-style information criteria do not count as generalization bounds. TRUE: as of 2014, practitioners chose models mainly with cross-validation or held-out validation, and model selection driven by generalization bounds was essentially absent from practice. MOSTLY TRUE: cross-validation dominated and bound-based selection appeared only in rare or niche research uses. MOSTLY FALSE: bound-based model selection was a minority but clearly real and used in practice, so the absolute claim overstates. FALSE: bound-based model selection was common in practice, or cross-validation was not the main method.

Set before the research began.

1. E1 CS5339 Lecture Notes #10: Model Selection National University of Singapore (Jonathan Scarlett lecture notes) · 31 March 2021

   “One of the most widespread practical methods for choosing a model is cross-validation:”

2. E2 CS5339 Lecture Notes #10: Model Selection National University of Singapore (Jonathan Scarlett lecture notes) · 31 March 2021

   “We can use this idea to perform model selection –”

3. E3 A Study of Cross-Validation and Bootstrap for Accuracy Estimation and Model Selection IJCAI (Ron Kohavi, Stanford University) · 1995

   “We review accuracy estimation methods and compare the two most common methods: cross validation and bootstrap.”

4. E4 On Over-fitting in Model Selection and Subsequent Selection Bias in Performance Evaluation Journal of Machine Learning Research (Cawley and Talbot, University of East Anglia) · July 2010

   “the findings are quite general and apply to any model selection practice involving the optimisation of a model selection criterion evaluated over a finite sample of data, including maximisation of the Bayesian evidence and optimisation of performance bounds.”

Independent sources from before and after 2014 describe cross-validation as among the most common or widespread practical methods for accuracy estimation and model selection (E6, 1995; E3, 2021). A 2010 JMLR paper lists 'optimisation of performance bounds' among model selection practices (E7), so bound-based selection did exist, at least in research. The evidence supports cross-validation as the dominant method, with bound-based selection as a real but narrower practice. That fits 'mostly true' better than the literal claim that no one did it.