boston_housing.md

• Open notebook 3.7 from the book repo in Jupyter (use CNTK icon if on lab desktop)
• Go through the notebook:
  • Training for 100 or 500 epochs as the notebook suggests is slow, reduce that to 30 and 200 respectively
  • Try using sklearn.model_selection.KFold instead of the manual validation data selection in the first big loop.
  • Does validation accuracy degrade if you use one less hidden layer?
  • Try using and mae loss function instead of the mse
  • Train a model with a single layer and a single neuron (no activation) - this is equivalent to a linear regression model. Take a look at the weights using model.get_weights(). Compare them to the coefficients found by sklearn.linear_model.LinearRegression for the same data.
  • When computing the smoothed version of the validation curve, try using a sliding window average (could be done via a convolution) instead of the exponential decay average used by the book author.
• When done, train the same network on house price data from the Kaggle House Prices competition