Here I will implement my own Machine Learning library from scratch that will use feed-forward fully-connected neural networks. I will train and test it on MNIST hand written digits data set, and, once the library is functional, I will use it as a part of my next Java API projects.
Progress status: the operational code for a library is written. The accuracy prediction of handwritten digits varies between 83-87% which is not great but also isn't bad for a classic stochastic gradient descent optimizer. Use of more advanced optimizers like Adam should further improve the prediction accuracy, but I won't be implementing it here for now.
What's next: I need to turn this code into an actual lirbary that could be used by other applications.