CS 5043: HW2

Assignment notes:

Deadline: Friday, February 15th @11:59pm.
Hand-in procedure: submit to the HW2 dropbox on Canvas.
This work is to be done on your own. While general discussion about Python and TensorFlow is okay, sharing solution-specific code is inappropriate. Likewise, you may not download code solutions to this problem from the network.
Submit PDF and Jupyter notebook files only; do not submit zip or MSWord documents.

Multilayer Network

This week we took our first steps into sequential, multi-layer networks in Keras. In this assignment, we will apply these skills to our brain-machine interface data set.

Here, we will predict the zeroth column of the theta matrix, which corresponds to shoulder angle in radians. As with HW 1, use the first four segments of data for training. Furthermore, use fold 18 for our validation data set. Note that the validation data set, like the training set, was extracted in the example brain-machine interface code that I distributed before HW 1 (this corresponds to variables ins_val and outs_val).

Part 1: Linear Network

In Keras, create a linear network with no hidden layers.
When training a network, one can include the validation set in the fit() call. After every training epoch, the validation set is used to evaluate the model and the performance is reported:
```
model.fit(x=ins_training, y=outs_training,
          validation_data = (ins_val, outs_val),
          epochs=epochs)
```
Train the model for as many epochs that are necessary to reach optimum performance (after this point, the performance may begin to worsen).
Query the network: extract the predictions and the mse for both the training and validation data sets. The former can be obtained by calling model.predict(); the latter by calling model.evaluate(). For this latter case, set the batch_size to be the length of the data set being evaluated.
In a single figure, plot the first 1000 samples for both the ground truth and the prediction for the training set. Make sure to label your axes and provide a legend.
Do the same for the validation set.
Using the MSE, compute and print the FVAF for both the training and validation data sets

Part 2: Multi-Layer Network

Create a multi-layer, nonlinear network. The output unit should still be linear, however.
Train the network, plot the results and report the FVAF, as you did for part 1.

Submission

Save your notebook (keep a copy for your records).
Generate a PDF from your notebook (File/Export/PDF). Make sure that your PDF file includes: your code, your training runs and the FVAF results.
Hand in the Jupyter notebook and the corresponding PDF.

Hints

The Keras Model Class documentation
np.var() can be useful for this assignment

andrewhfagg -- gmail.com

Last modified: Sat Feb 9 01:19:14 2019