Isn't the name "perceptron" going to be a little confusing in this context?

Yes, I found the name misleading also.

That was my first thought as well

Why exactly?

...Hence why I've called my "feedforward artificial neural network model that maps sets of input data onto a set of appropriate outputs" builder Perceptron...

This would be like calling a linear model builder "Regression"

No need to take the name so literally though - it's rare names are taken literally at all!

Not in particular. Just a satisfiying name, and relevant enough.

I respect the choice not to use a framework. I keep flirting with the idea of posting some of my trained char-rnn models online, in hopes people find them amusing and possibly share theirs in turn. Trained RNN models aren't small, but they're not huge either. 20-30mb generally. In other words it wouldn't be prohibitive to accumulate a collection of trained, GNU models, if the outputs would be useful and interesting to people.

Out of curiosity how are you serializing the output models? Are the files memory-mapped bytecode, or does the training encode a nice, clean data structure?

when you say output models are you referring to the optimized ANNs, or the organisation of output neurons?

Very nice idea.

What is the objective function (loss function) that `back_propagate` is optimizing exactly? I'm new to the world of NN, so really excited to see NN code that works from first principles. Reminds me of https://github.com/mnielsen/neural-networks-and-deep-learnin...

Perhaps a little more detail about the math behind the code would be in order? I'd be happy to help out.

Another cool direction would be to convert the python code to cython and see how much faster it will run...

Cheers for feed back. The cost/error function is simply the 1/2(a-y)^2 ...(I think its called the square cost function). I wrote all the code myself (no libs) so people like yourself can see all steps required to train an ANN. I was thinking of getting things working for GPU, which should even make cython look slow. All depends on how much attention I get. Right now it's still got quite a few bugs, and in general is under done.

For didactic purposes there's also:

- http://www.ra.cs.uni-tuebingen.de/software/JavaNNS/

- http://www.cs.waikato.ac.nz/ml/weka/

These are most likely more matured than Perceptron at the moment. More work to do!

Yes. But mature projects also have larger codebases and many knobs to play with. That can make it hard to serve as introductory stuff. Focused projects have their purpose too... like what Minix is to Linux. So it's fine.

Other software to play with neural networks can be R and Scilab, not to mention others like: Tensorflow, Torch, Theano, Pandas, Scikit-learn, Caffee and many others. All of them free, great communities and videos can be found in YouTube with nice intros.

Another way to learn can be just grab Python and Numpy and roll your own multilayer perceptron with backpropagation, then play with other concepts.

Jupyter notebooks are also a great way to explore what's out there.

This looks great, but I think that it would be even better if there was some way of bootstrapping it to existing libraries, say for example keras. I don't know how easy this would be, but I believe that it would definitely be very useful. Even without this capability, it looks like a great learning tool!

great to hear! The reason I don't use libraries is because I want the code, in itself, to be educational. So people can read the code and see what algorithms are being used for every step. This project is less about trying to compete with amazing performance, and more of an educational play ground.

Thanks so much for this! I'm just getting started with ML and deep learning so anything with visuals and "knobs"/"dials" with real-time feedback to visually show relationships of changes and their effects is a huge eye opener!

Really happy to hear it's living up to its purpose.