An convolutional neural network with drop out built from scratch. Its purpose is to classify images into one of six hard-coded categories: airplane, butterfly, flower, piano, starfish, and watch. The network supports arbitrarily many convolutional/pooling/fully connected layers by exposing a simple builder pattern. Clients may also set hyperparameters in this fashion.
Ensure you have Java SE 8 installed on your machine.
Run make at the top level to compile the code and then make test to train and test the network. The network will output its train, tune, and test accuracy at every epoch.
- Convolution layers do not support biases.
- There is currently a bug in the convolution layer's backpropagation logic, causing test accuracy on this dataset to cap at 65%.