LAB 10 - Neural Networks - Mike Smith


Average Number of correct (>.5) guesses for each
integer after 30,000 sweeps over 5 tries.

24	23	20	22	22	23	24	24	23	24
22	20	11	21	21	21	10	18	19	17
19	21	5	7	7	22	13	17	21	8

Network #1 got the most right. It is a simple network, a fully connected, forward feeding tree with ten hidden nodes. It recognized 229 of the 250 integers. That is 91.6% of the data. It had the most trouble with the twos.

Network #2

Network #2 did second best. It attempted to employ feature detectors by all 6 rows and 4 of the columns:

1 1 1 1 1 1 (1)
2 2 2 2 2 2 (2)
3 3 3 3 3 3 (3)
4 4 4 4 4 4 (4)
5 5 5 5 5 5 (5)
6 6 6 6 6 6 (6)
(7) (8) (9) (10)

(Node #)

This network got 168 of the 250 integers correct. That is 67.2%. This network may have done better if I had used different features, for example the four corners may have provided better information to the network if I had set up the nodes that way.

Network #3

This network did the most poorly of the three. It got 149 of the 250 integers correctly. That is 59.6% of the data. I would like to believe that this means that this network is the most generalized, and would do the best with novel data. This is kind of like when you teach children these days, you don’t just want them memorizing facts, you want them to make connections that they can use in areas of study which they haven’t specifically spent time memorizing. Again, this network may have been more successful had I done something different with Network #2 because this network is built on the same foundation, with just a fully connected 10 node second hidden layer.