VectorStream is a from-scratch implementation of a few Artificial Neural Network primitives:
- Automatic differentiation
- Gradient descent
- Linear units
- Convolutional units
- Max pooling
- Dropout
- Arbitrary activations
- Softmax classification
E.g. below is the output during training of a network for handwriting recognition on the MNIST data set. In this case, the network NewNet(LU(10,10), Re(LU(10,16), Re(LU(16,784))))
reaches about 96% accuracy.