This approach focused primarily on multilayer perceptrons the
categorize the bird calls. This type of network consists of two major
components, axons and synapses. The axons contain a number of neurons,
and apply a discriminating function to any signal. These axons are
connected to each other by synapses. Each synapse connects a single
neuron in one axon to a neuron in another axon. When the first axon
transmits a signal through the specific neuron, the synapse scales
the signal by its weigh, which may be positive (excitory) or negative
(inhibitory), before passing it to the next neuron. By connecting
several axons together in this manner, a non-linear transfer function
can be applied to a set of input data.
The key to a neural network comes in its training. When
a set of data is input to the network, the resulting values can be retrieved
and compared to the desired value. Then the difference in the values can be
propagated backwards through the network by applying the inverse of the
discriminatory functions and synaptic weights. The synaptic weights are
then changed to bring them slightly closer to the weights required to produce
the desired results. The training of the network is achieved in this way
with a large amount of data and many iterations through the data (epochs).
With a sufficient number of epochs, the synaptic weights can approach values
that will correctly categorize a variety of different input signals.
A variety of data sets were used as described above. It
was found that due to the time variance of the signals, the longer interval
lengths tended to provide much better results. As a result, the interval
parameter was removed from consideration, and the entire vocalization was
analyzed as one interval. The simplicity of this approach encapsulated all
of the characteristics of the vocalizations in a compact form thus reducing
the processing time dramatically.
It was decided that the dominant frequency of each call
was an important characteristic of the vocalization.
There were four different styles of network built. The
first was built to distinguish between the House and
Purple Finches. The second was trained and tested on four species in
the Empidomax genus of Flycatcher. The
third investigated the networks response to song
mimickery. The final network was built to accommodate a
larger number of bird species.
Classification 1: House and Purple Finches
Classification 2: Empidonax Flycatchers
Classification 3: Mimickery
Classification 4: Many Species Classification