How many neurons are optimal for distinguishing sounds?

Question

Neural networks

Hello.
I'm making a neural network for sound recognition.
I select the optimal number of neurons, and now I settled on this structure:
72 -> 2000 -> 1000 -> 500 -> 200 -> 100 -> 2
72 -> these are chroma_stft features, I get them using librosa .
2 days off, that's 2 sounds that I need to distinguish.
I teach in such a way that I cut the sound into equal pieces of 300 ml seconds, and I feed these pieces into the neural network.
Let's say I get 3 pieces from a sound in 1 second, and I feed these 3 pieces to the neural network three times, and just if the neural network gave out more in favor of some type of sound, then this is it.

I also want to know which activation function to use, again, judging by the experiments, Tanh works best , and sigmoidin this case, it performs very poorly.

Answer 1

You have a neural network. You have data. You can start the network, get a response, calculate the error. Fine.
If you really understand what you are doing, then you should run experiments, each time changing the network configuration, changing activation functions, running a couple of hundred experiments with different data and comparing the results - i.e. comparing the errors obtained in each experiment. And then tell us what happened as the best solution. Well, or publish your results at least in the form of a scientific work, at least as an article on Habré.
And what answer in terms of optimality do you expect to hear on the forum? What would someone have guessed the answer to? Well, someone from a great fantasy will tell you some numbers. Do you blindly trust them?