This is still a bit chaotic right now, because I'm just trying to get my stuff out there. For the moment, I have introduced
-
a Neural Network class which allows doing any fully-connected neural network with an arbitrary number of layers with each layer having its own number of nodes/activation function, which can be either the sigmoid, ReLU or LeakyReLU
-
two examples of application of this network in practice on randomly generated data for binary classifiers, one to classify almost linearly separable data and one for non-linearly separable data but using the same ideas
-
a PDF format explanation of how to use this network and how it works behind the scenes.
The files linear-logistic-regression.py and non-linear-logistic-regression.py are ready to be compiled in python3 ; you only need to have NeuralNetworkMBGD.py in the same directory since the data sets used in those examples are generated by the random library within the file.
Because the data is randomly generated, there is an extra "data accuracy" parameter which says how accurate a perfect model would classify the data points in their categories.
For any questions/comments, please contact me at patrick1dasilva@protonmail.com.