Sorry for cross-posting.

In an effort to understand a bit more about neural networks, I wrote a Pd external, by translating Python code to C, from the book "Neural Networks from Scratch in Python". After a couple of months of working on this, I ended up with [neuralnet].

This is an object written in pure C, without any dependencies, for creating densely connected neural networks for classification, regression, and binary logistic regression. You can choose among different activation and loss functions, optimizers, and other settable features. I've created some examples, trying to replicate some of the examples in the aforementioned book, and some examples of my own.

The code is on GitHub (https://github.com/alexdrymonitis/neuralnet), and Linux amd64 and armv7-32 (Raspberry Pi) binaries are uploaded to deken. I don't have a Mac or Windows machine, and I don't know how to compile for these architectures on a Linux machine (or if that is even possible). I would be grateful if anyone can compile for any of these architectures and upload to deken.

I would also love to get feedback both on how the object performs, and on the source code itself.

Version 0.2 is already up on deken and GitHub, as there was an issue with Windows, which is now solved. Please try the new version, especially if you're on Windows. Hopefully binaries for Windows will come soon, as there was already some help on GitHub by a couple of people. I hope macOS binaries will also be available soon, but again, I'll need help for that.

Windows binaries are uploaded to deken, thanks to Lucas Cordiviola.

Huge job

While your finger are still warm, what would be great would be to be able to export "images" of the neural net training, layer by layer, like in all those images blending app.

I think it is a bitmap of the output values of each neuron of a layer.

@Il-pleut can you elaborate a bit more, or provide some link that explains what you say in a bit more detail?
How would it export it? In a .png file? Or maybe Gem or Ofelia?

Deep dream generator from google does it.

Maybe you have seen images generated by "IA". They are everywhere know.

They are in fact not the output of a neural network, but are visualisations of the internal process of a layer of a neural network "analysing" or "sorting" an input image.

This ability to see what's going on inside of a neural network was implemented in google's tensor flow for debbuging purpose, if I remember it correctly.

To make it simple, in a neural network with an input vector shaped as a screen, with each input being a pixel, and the pixels organised in lines one above the other, a code to generate images from, let say, instant output value of each pixels of a layer, or other values coresponding to the differents internal parameters of each neuron of a same layer.

I think a .bitmap format is straitforward to generate.

Binaries for all OSes (Linux, Mac, Windows, Raspberry Pi) are now available on deken.

@alexandros

Hi, I am trying out this external, curious about it.

Would it be possible for you to describe a bit more on how to explore and work with the examples provided? I have only a very basic understanding of neural networks and have not read the book you base the examples on.

@cfry Well, one case is to enable some sort of automation of parameter control. For example, the two examples that control a synth, enable the control of many parameters with only a few inputs. These examples control three or five parameteres, but I have used the object to control twenty parameters with the mouse coordinates, so with two inputs only. Additionaly, it helps you explore sounds as the parameter control will most likely fall in places that will creates sounds from a synth that you haven't created yourself, as the different combinations of twenty or so parameters are way too many to be able to thoroughly explore them all.
Concerning classification, it depends on what you would like to do. You could train it for example to clasify vowels, and based on speech input, keep a count of every vowel and do something with it.
Neural networks are just a tool to facilitate cumbersome processes, which can also provide interesting results. Anyway, I'm no specialist, so it's best if you keep searching to find ways to use NNs in your work.