I'm building a neural network to attempt to distinguish human speech. I'm looking into analyzing a spectrum of frequencies, and found that I'm likely going to be using a large number of nodes (between 55k to 6k)- I came to this by using a value told to me by my AI book, that the hidden layer should have 2n/n nodes in it (AI: a modern approach, 3rd edition p 732 footnote).
I have not, however, found any sort of information regarding the number of hidden layers that should exist for a given system, or the number of nodes in the input. Does anyone know of any resources that would describe this, or know themselves?
[link] [26 comments]