Some questions have answers which are obsolete, or outdated. One prominent example is How to choose the number of hidden layers and nodes in a feedforward neural network? This question has an answer with an incredible (for stats.SE) 605 upvotes. The question is among one of the most up-voted questions on our site.
However, the time that this question was asked (2010), was just before neural networks had their recent renaissance. It would not be until 2012 that the AlexNet paper (Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton, "ImageNet Classification with Deep Convolutional Neural Networks") used CNNs for the ImageNet task and vastly out-performed their competitors, sparking a resurgence of interest in neural networks generally, convolutional neural networks specifically, and the technique of multi-layered networks along with it.
In other words, the current state of the art has advanced well beyond the suggestions in most of the answers to this question. The question itself, I believe, is implicitly asking about simple feed-forward neural networks because CNNs, RNNs, residual networks and other exotic architectures had not yet experienced this explosion in attention. Related neural network techniques, such as word2vec, hadn't even been invented yet. (Mikolov et al published in 2013.)
In light of the deficiencies in the answers to this question, and perhaps the limitations of the question itself (due to its place in time), what, if anything, should be done to direct readers to more recent, and more relevant, answers to the question?
These are two options that have occurred to me. There are probably more.
- One option is to simply use the SE machinery as it is to write a new answer and hope that intrepid readers will make it all the way down to the bottom of the list to find recent information. Or that they'll sort by recent... and scroll past the Accepted answer pinned to the top.
- Ask a new question which is deliberately distinct. "How do I choose the right number of layers and neurons in light of the many recent advancements in neural network architectures since 2011?" This question, lacking the upvotes and visibility that comes from 10ish years of accumulated hyperlinks, might be harder to find but would report more recent information. This question, because it deliberately asks about what's changed in the intervening time, cannot be closed as a duplicate of the older question.