There’s a very easy tutorial to get your first neural network up and running at:

machinelearningmastery.com/tutorial-first-neural-network-python-keras

Do the tutorial to create a network for the Pima dataset and run it.

I’ll put a version of the well-known “Iris” dataset on the course website. This dataset, dating from 1936, relates four parameters of irises (petal length and width, sepal length and width) to their species (Iris setosa, I. versicolor, or I. virginica). I’ve manipulated this so that the dataset is shuffled and in a form amenable to a “softmax” output. A softmax output layer is one that contains \(n\) neurons, one for each possible category of output, and whose activations sum to 1: thus, each neuron’s activation can be viewed as the probability that the input is a member of the corresponding category. The format of the iris data is a CSV file:

sepal-length, sepal-width, petal-length, petal-width, I.setosa, I.versicolor, I.virginica

So this data item:

6.6, 3.0, 4.4, 1.4, 0, 1, 0

\noindent would represent an Iris versicolor with sepal length and width of 6.6 cm and 3.0 cm, and with petal length and width of 4.4 cm and 1.4 cm.

Your job is to create two neural networks to classify irises based on values of the four input parameters. You can use as many layers and neurons per layer you like for each; a simple single-hidden-layer network is sufficient, although you can experiment with other kinds to see how your accuracy changes. Although you could use any type of layer, I would suggest sticking to the “dense” layers used in the tutorial. The output layer, however, must be a softmax layer with 3 neurons, one for each type of iris. Vary the number of layers and/or the number of nodes per layer between the two neural networks you construct.

I’ll split the dataset into two files, iris_train.csv and iris_test.csv to use for training the net and testing it, respectively.

Turn in a file containing:

• A brief (1–2 page) description of the networks you created and a discussion of what their differences meant for the accuracy of their answers. Relate any insights or hypotheses you may have about why what you observed happened.
• Your Python programs for the two networks.

• Output from the two programs¹

  • Your output must include a line with the accuracy of your for test data.
  • It must also include the results of predicting the output given the inputs for the test data.

  For example, my output looks like this:

    (keras) [rmt@rturner-2 NN]$ python iris.py
    Using TensorFlow backend.
    Epoch 1/150
    132/132 [==============================] - 0s - loss: 0.6367 - acc: 0.6667     
    Epoch 2/150
    132/132 [==============================] - 0s - loss: 0.6365 - acc: 0.6667     
    ...
    Epoch 150/150
    132/132 [==============================] - 0s - loss: 0.3273 - acc: 0.7828     
    15/15 [==============================] - 0s
    
    acc: 71.11%
    
    Predictions (based on test dataset):
    prediction=[ 0.11855328  0.44186708  0.43957967]
    prediction=[ 0.98168379  0.01319122  0.00512492]
    prediction=[ 0.09259017  0.44522405  0.46218583]
    prediction=[ 0.10822026  0.44351041  0.44826928]
    prediction=[ 0.09259017  0.44522405  0.46218583]
    prediction=[ 0.09259017  0.44522405  0.46218583]
    prediction=[ 0.10330817  0.44415826  0.45253354]
    prediction=[ 0.09259017  0.44522405  0.46218583]
    prediction=[ 0.99175453  0.00614701  0.00209844]
    prediction=[ 0.09259017  0.44522405  0.46218583]
    prediction=[ 0.09259017  0.44522405  0.46218583]
    prediction=[ 0.09259017  0.44522405  0.46218583]
    prediction=[ 0.09259017  0.44522405  0.46218583]
    prediction=[ 0.12725127  0.44022453  0.43252417]
    prediction=[ 0.09259017  0.44522405  0.46218583]
  

  In this case, the accuracy of the network was 71%, and (for example) the prediction for the last of the tests is that it is type 3, with a probability 0.46 (it actually is a type 3).