This project achieves a 90% accuracy for given images. Expensive functions such as ‘forward’, ‘backprop’ and ‘update_weights’ are executed on GPU device. ‘mnist’ images has been used for training and testing purpose.

Architecture:

Input Layer

The input layer is a 28×28×1 grayscale image.

Fully Connected Hidden Layer

This is just a fully connected layer with ReLU activation. It consists of 1024 neurons with dropout rate of 0.4. This means that during training, any given neuron has a 0.4 probability of “dropping out”, which means that it is set to 0, regardless of the inputs.

Output Layer

This is the final classification layer. It is a fully connected layer consisting of 10 neurons, one for each class. It will compute a softmax.

Technologies: Modern C++(11/14), CUDA

Check out the Project Source Code for more detailed info on implementation.