Neural Network Engine

The Neural Network Engine (NNE) is a C++ program that I created as an exercise in understanding the basics and low-level operation of machine learning. It contains a self-written matrix library and several kinds of "layer" classes that when connected together, form a network. The NNE is currently geared towards solving the MNIST handwritten digit classification problem but is highly scalable and will easily adapt to other, more complicated problems.

NNE and the MNIST Dataset

The NNE is currently being used to classify items in the MNIST data set. This problem involves the use of alternating convolutional and pooling layers to break down 28x28 grayscale images of handwritten digits into a series of features and MLP layers to classify the digits based on those features. Shown to the left is the successful classification of the digit '6'.

NNE and the MNIST Dataset Cont.

The alternating convolutional and pooling layers extract the principal components of the digits and compress the convolved images to be more manageable for the MLP classification layers. 

Code Organization and PolyMorphism

The C++ code is organized such that networks can be created and executed in a very modular and intuitive fashion. Layer classes are organized by types, currently including pooling, element-wise, MLP, and convolutional. All inherit from the virtual layer class which provides a common interface for layer interaction.