Classification 4: Many Species Classification

This trial was focused on expanding the network to include more species of birds than the previous networks. A similar network to the basic network described above was built with the exception that the output and hidden axons both contained eight neurons. This change was made to accommodate the eight species selected for this trial. The birds selected were the Black Cap Chickadee, Bewick's Wren, Northern Flicker, House Finch, Purple Finch, Acadian Flycatcher, Dark-eyed Junco and Violet-green Swallow. These birds were selected for the availability of sample recordings. The testing of this network showed less favourable results than previous trials. After several networks were trained and tested, perfect classification was never observed. While the Northern Flicker proved the most difficult to classify, the other species were misclassified in an apparently random fashion from network to network.

In order to improve results, the number of neurons in the hidden layer was increased to fifteen. The test results of this new network showed that all species except the Northern Flicker and Black Cap Chickadee were consistently classified correctly. The Northern Flicker continued to be misclassified in most tests, but perfect classification was seen in 10% of runs. Increasing the number of neurons in the hidden layer to twenty allowed the network to train itself much faster than the previous network as seen in, but the results obtained in the testing phase were unchanged.

Next the number of hidden layers in the network was increased to two. The first hidden axon was given twenty neurons and the second was given fifteen. Testing showed that perfect classification was achieved in approximately 40% of the networks trained, with the same two species as noted above being responsible for all misclassifications. This required more epochs to train, and due to the greater number of synapses, took longer to run each epoch.