Due: April 26
Turn in hardcopy in class
Use http://playground.tensorflow.org/ on the two-spirals problem (see Data on left side)
Perturb the following parameters, considering extreme values as well as moderate ones (with respect to the available parameter settings):
- Ratio of training to test data
- Data noise
- Learning rate
- Regularization rate and type (L1 or L2)
- Number of hidden neurons
- Number of hidden layers
- Alternative input features
a. Test/training loss (is there overfitting?)
b. Convergence rates (does the network stabilize? how fast?)
c. Output map (what pattern does the network learn? in what ways does it differ from the spirals?) and maps within hidden neurons