Neural network from scratch

The only dependency required is NumPy.

• Support for the following is provided
  • Forward pass and backprop
  • Batch and mini-batch gradient descent
  • Activation functions - sigmoid, relu, tanh, linear, softmax
  • Loss functions - MSE, binary cross entropy, cross entropy

• Creating and training a model is simple (somewhat similar to the Tensorflow API):
  1. Create a list of Layer objects
  2. Pass the created list and necessary parameters to instantiate the NN object
  3. Call train on the NN object

• Benchmarked using the MINST dataset
  • Simple network with a single hidden layer achieved 88% accuracy on the test set