Attached you’ll find a simple feedforward network leveraging stochastic gradient descent to train MNIST (handwritten digit recognition).
As part of your task, we kindly ask you to optimize the speed of convergence of this network.
You can either reduce the number of epochs required for convergence by improving the training method, or use improved technical implementation for an increased through-put of samples.
Proof your improved convergence speed by logging the convergence performance over time of the provided code and your improved version. You can export your log in a CSV or share a saved state notebook. Document your approach in a structured email and document implemented changes further in code.
Improving the neural network’s architecture for stronger performance (e.g. CNN) Leveraging frameworks that provide support based on computational graphs (e.g. TensorFlow) or native GPU support (e.g. Theano)
This repository builds on "Neural Networks and Deep Learning", Determination Press, 2015
Feel free to use their free resources to get a better understanding of the network (http:https://neuralnetworksanddeeplearning.com/).