Project 1: FNN FOR MNIST

Due Jan 22, 2020 by 11:59pm
Points 100
Submitting a website url or a file upload

In this section we learn the workflow mentioned in above sections through a “Hello World!” project using the MNIST data set to go through the key steps in building a deep neural network.

The starter code can be found at

https://www.kaggle.com/scaomath/simple-mnist-numpy-from-scratch Links to an external site.

For this starter project, we need to do the following:

Split the 42,000 training sample further into a training set and a
cross-validation set.
Add an extra hidden layer in the model, so that our model now has 2 hidden
layers.
Train the network using mini-batch SGD instead of the full gradient descent
method.
Validate the performance of the model (e.g., change the combination of the
sizes of the hidden layers, the strength of the regularization), to keep it from
over-fitting using the two datasets from the first step.