Results 3

Using NN with 10 hidden nodes.

The false positive problem (FP)

Four classes of examples, class 5 is new

Accuracy on the four classes: percent error: 73.6%.
Note: y axis scales vary to better show waveform shapes,
mean amplitudes in each row may be very different.

The NN is confused by classes 2 and 3 (rows 2 and 3) because their
waveforms are similar, and because the NN has seen about the same
number of each (spike counts are similar).

The NN was not trained or tested on the 105 class 5 (row 5) waveforms
at this point.

Test the NN on the 105 new class 5 waveforms.

The FP rate with threshold p >= 0.90 for class 5 waveforms was 1.0.
The NN fails on this task because it classifies the new spike waveforms
as members one of the four known classes based on similarity even if the
most similar class is not very similar at all. In fact, a test input of
all zeros might be classified with a p >= 0.90.