Features

• Fully connected layers, obviously.
• Convolution, both standard and partially strided/transposed/deconvolution. The latter is used to calculate the output value errors.
• A gui: the neural network is fully customizable, as are most (if not all) hyperparameters. Deep networks are supported here, as well.
• Pooling in various sizes.
• Momentum
• Nesterov Momentum
• LeCun initialization
• RMSProp
• Batch normalization