Hello,
in order to patch a target drive for my synth sound breeding genetic algorithm "Ewolverin", I need to measure the tonal distance of different sounds to a target sample. Is there a way to do this with an artificial neural network in Pd? If yes, who knows how and could help me?

Regards
Elden

If you are talking about evolution, then it's likely you need something like a recursive neural network (RNN), which [neuralnet] can't create (at least not yet). Unless this evolution you mention can be encapsulated in one snapshot. If it's spectral snapshots with a 256 bin resolution, and you need four of these snapshots to classify your data, then a network with 1024 inputs could worh with [neuralnet].
The MNIST dataset example (example 4) is a classification example. You should follow that to build your network structure (concerning activation functions, loss function, etc.).
With the current versino of the object you can also build autoencoders, which can be sort of good for compressing data. That being said, you could perhaps use an autoencoder and map the values in the latent space. The next version will hopefully include variational autoencoders (working on this), so that could also be an option.
As I don't know the structure of your data though, I don't think I can be of more help right now.

Ok, thanks a lot! Yes, the audio signal will be recorded and encapsulated as a sample of the same length like the target sample. No recurrent network needed. I just don't get how I could distribute the inputs of the network across the fft waterfall spectrum, you know? Every single input has to detect an amplitude value on a certain position in time and frequency and I don't know how [neuralnet] does that. Or is that already implemented somehow?

[neuralnet] does no FFT. You'll have to do all the FFT and IFFT after you get output from [neuralnet]. This object only enables you to create neural networks. What sort of data you'll use and how you'll prepare it is totally up to you.
From what I understand from your post, you should probably store your FFT spectra, and classify them according to the amplitude you mention. Then feed them together with the classes you have created to [neuralnet]. Does this make any sense?

Yes, it does. Now that makes me think about leaving the FFT-realm and turning to [TimbreID] to directly extract features from a sample over time and have them compared using [neuralnet]. Then it's no longer an FFT-issue, but simply a comparison of several X-Y-curves. I'll see what I can achieve. Thanks for your help!

@alexandros By the way: do you have a scheme of the internal operations of a single neuron inside [neuralnet] for me to check out?

It's just an Multi-Layer Perceptron (MLP) - a dense feed-forward neural network. Each neuron sends its output to all neurons in the next layer. Each neuron in the next layer multiplies its of its inputs by a weight. The weighted inputs are summed and added to a bias value, then passed through the activation function of the layer (e.g. a sigmoid function to ensure everything remains between 0 and 1 and don't go to +/- infinity) and passed to all the neurons in the layer after that.
It's all in the C code. If you want to learn more about this though (in case you don't know), I recommend the "Neural Networks from Scratch in Python" book, which is where I initially translated Python code to C to create this object.

Ok, thanks. Great work! That should do the job I need it for. I'll get back to you if I run into some hurdles. Thanks a lot!

@alexandros Hi, I get an SHA 256 error when I try to install neuralnet via deken. Is there some Windows binary available somewhere else ? Thanks

@jyg it hasn't been digitally signed, so I guess that's why you get this error. Can't the object be installed with this error? Otherwise you can get the source and compile yourself https://github.com/alexdrymonitis/neuralnet