You may work with a partner on this assignment. Be sure to add both names to the top of every file. Only one of the partners needs to turn in the code.

Before you begin do an update63 to get a copy of the starting point file, backprop.py, which you'll need to complete this homework. It should appear in the directory cs63/homework/05.

Take some time to read and understand the starting point code. It uses three classes: Unit, Connection, and Network. You will need to complete several methods of the Network class that are involved in:

• forward propagating activation
• teaching the network a pattern by backward propagating error
• computing error during training

Follow the comments provided in the file when implementing these methods. Feel free to create additional methods within the other classes in order to complete the Network class.

Your completed code should be able to successfully learn the two examples provided at the end of the file. The XOR problem may take several thousand iterations to find a solution (and in some cases, it may even fail after 10,000 iterations). If you'd like to improve learning speed, read about how to include momentum in the back-propagation algorithm.

• The activations of units in the input layer should simply be the values in the input pattern. The sigmoid function should NOT be applied to input unit activations.
• Be sure that the bias unit is contributing to the net input of the units in the hidden and output layers. If you loop over all incoming connections to a unit, you will automatically get the connection from the bias unit.
• Calculate all of the weight increments throughout the entire network before modifying any of the weights. If you modify weights as you go, you will not be able to calculate the correct values for back-propagation of error.

Thanks to Jim Marshall at Sarah Lawrence College for the idea for this assignment.

Once you are satisfied with your program, hand it in by typing handin63 at the unix prompt.