Neural Networks from Scratch

Learn backpropagtion the hard way

In this repository, I will show you how to build a neural network from scratch (yes, by using plain python code with no framework involved) that trains by mini-batches using gradient descent. Check nn.py for the code.

In the related notebook NeuralNetworkfromscratchwith_Numpy.ipynb we will test nn.py on a set of non-linear classification problems

  • We'll train the neural network for some number of epochs and some hyperparameters
  • Plot a live/interactive decision boundary
  • Plot the train and validation metrics such as the loss and the accuracies

Example: Noisy Moons (Check the notebook for other kinds of problems)

Decision boundary (you'll get to this graph animated during training)

Decision boundary
Decision boundary

Loss and accuracy monitoring on train and validation sets

Loss/Accuracy monitoring on train/val
Loss/Accuracy monitoring on train/val

Where to go from here?

nn.py is a toy neural network that is meant for educational purposes only. So there's room for a lot of improvement if you want to pimp it. Here are some guidelines:

  • Implement a different loss function such as the Binary Cross Entropy loss. For a classification problem, this loss works better than a Mean Square Error.
  • Make the code generic regarding the activation functions so that we can choose any function we want: ReLU, Sigmoid, Tanh, etc.
  • Try to code another optimizers: SGD is good but it has some limitations: sometimes it can be stuck in local minima. Look into Adam or RMSProp.
  • Play with the hyperparameters and check the validation metrics